Most of these discussions take place on social media, in popular books and in parenting advice. However, they also appear in settings where the stakes are much higher. One of these settings is the family court. When attachment concepts are used in custody disputes or child protection cases, they can directly affect decisions about where, and with whom, a child grows up.

In this post, I would therefore like to have a closer look at a recent consensus statement that addresses exactly this setting: “Attachment goes to court: child protection and custody issues” by Forslund et al. (2022), published open access in Attachment & Human Development. It was written by a large group of attachment researchers and practitioners from many countries. As far as I can tell, it is currently the most comprehensive document by the attachment research community that sets out what its science can offer family courts.

The consensus statement has recently been complemented by an open access practitioner review in the Journal of Child Psychology and Psychiatry, “Clinical insights from attachment theory and research for professionals working with young children and their families” by Opie et al. (2026), to which I contributed together with many other colleagues. It addresses professionals working with young children and their families more broadly, but it explicitly names court proceedings and custody evaluations. Where it adds something relevant, I will refer to it alongside the consensus statement.

A short disclaimer at the start. I am an attachment researcher, not a lawyer, and legal frameworks differ considerably between countries. What follows is not legal advice and should not be used as such.

Let’s have a closer look!

Summary

Forslund et al. (2022) propose three attachment principles for family courts. (1) Children need familiar, non-abusive and non-neglecting caregivers. (2) Continuity of good enough care should be valued highly. And (3) a network of attachment relationships is an asset for children.

The authors furthermore conclude that assessments of caregiving behaviour, and particularly of caregivers’ capacity to provide a safe haven, should take centre stage, while measuring children’s attachment quality is currently most suitable for guiding supportive interventions. Importantly, the authors openly disagree among themselves on whether any assessments should inform custody and protection decisions at all. The practitioner review by Opie et al. (2026) describes a very similar disagreement as one between a “strict” and an “expansive” translational perspective, and it summarises what children need as safe, stable and shared (“Triple S”) family-based care.

The main issues both papers identify, include describing attachment as “strong” or “weak”, confusing disorganised attachment with maltreatment or an attachment disorder, treating different attachment measures as interchangeable, assuming that one “primary” caregiver matters more than all others, confusing a parent’s bond with a child’s attachment, and treating early attachment as destiny. These misconceptions have direct implications for how courts, expert witnesses, family professionals, social workers and caregivers talk about attachment, and for the decisions they make.

1. Why attachment ends up in court

Most family law systems decide child protection and custody cases according to the best interest of the child standard, now typically associated with Article 3 of the UN Convention on the Rights of the Child. Forslund et al. (2022) point out that this standard is broadly formulated. Applying it requires some account of what children need and how they are likely to develop, which creates a demand for expertise from developmental research.

Attachment science seems to meet this demand particularly well. It links caregiver sensitivity to children’s attachment quality, and children’s attachment quality to later development, with a large empirical literature behind both links. It also speaks directly to the child-caregiver relationship, which many legal frameworks list as a factor in best interest decisions.

There is, however, a structural tension. Science generalises from individual cases to general principles, whereas courts particularise by applying general principles back to individual cases. Group level findings are thereby easily turned into predictions about one particular child, one particular caregiver and one particular family.

Opie et al. (2026) make the same point for clinical practice. The core propositions of attachment theory are strictly probabilistic, they support group-level associations with many individual exceptions, and they cannot be used to predict with certainty how an individual child will develop.

According to Forslund et al. (2022), this tension has produced two opposite problems. In some cases, attachment ideas and measures have been overused, with practice drifting away from the evidence. In other cases, attachment has been dismissed altogether. In GM v Carmarthenshire County Council [2018] EWFC 36, Mostyn J suggested that attachment theory might be regarded as a statement of the obvious and held that a social worker's evidence based on it did not qualify as expert evidence. In 2021, the Court of Appeal disapproved these observations. It noted that a particular witness may not be qualified to give evidence about a child's attachment, but that this does not call the validity of attachment theory as such into question, and that expert advice about attachment may have a role in difficult cases (Re TT (Children) [2021] EWCA Civ 742).

What I value about the consensus statement, is that it does not place the responsibility on courts and practitioners alone. Forslund et al. (2022) acknowledge that attachment researchers, including many of the authors, have at times inadvertently advocated their methods and exaggerated the implications of their findings without stating the limitations. In their own words, “in hindsight, it is plain to see that we should have been more careful” (Section 2.3). The authors are equally explicit about where they disagree with each other, which I will come back to later.

2. What matters for children: three attachment principles

The three principles below correspond closely to what Opie et al. (2026), with reference to Forslund et al. (2022), describe as the safe, stable and shared (“Triple S”) family-based care that vulnerable children require.

2.1 The need for familiar, non-abusive and non-neglecting caregivers
The first principle is that attachment relationships, and the developmental benefits associated with them, depend on repeated experiences of safe haven provision (i.e., providing comfort and protection when children are alarmed or distressed; see here) by particular, familiar caregivers who are not abusive or neglecting. Attachment takes time and many repeated interactions to develop and maintain.

For custody decisions, this means that very limited contact with a caregiver makes it harder for a child to build and keep expectations that this caregiver will be available when needed.

For child protection, Forslund et al. (2022) conclude that almost all non-abusive and non-neglecting family based care is likely to be better for children than institutional care, where caregiving is typically unstable, even when nutrition and medical care are adequate.

2.2 Continuity of good enough care
The second principle is that expectations about safe haven availability stem from particular relationships and are not simply transferable to new ones. Great caution is therefore warranted before disrupting a child’s existing attachment relationships.

Forslund et al. (2022) describe a paradox here. Holding families to the standard of a child’s apparent best interests can harm the child’s true best interests if the value of continuity is left out. Another placement may look better on some dimension, yet moving the child means losing a relationship in which the child has already learned what to expect. Following van IJzendoorn et al. (2020), the authors suggest that “good enough” care (i.e., care that adequately meets a child’s developmental needs over time) may be a more suitable standard in court contexts than an idealised best interest.

This principle does not mean that removal is never justified. The authors state clearly that out of home placements are sometimes necessary and even lifesaving. They emphasise, however, what often follows removal. Depending on the country, Forslund et al. (2022) report that between one-fifth and two-thirds of long term placements break down prematurely, and multiple placement changes are associated with markedly higher rates of behavioural problems. Yet, the authors caution that these associations should not automatically be attributed to removal itself, since the instability and other factors that accompany out of home care are likely involved.

Attachment-based interventions evaluated in randomised controlled trials have been followed by improvements in both caregiver sensitivity (see Juffer et al., 2017; Moss et al., 2011) and child attachment (Bernard et al., 2012; Stronach et al., 2013), often within 6 to 10 sessions (Bakermans-Kranenburg et al., 2003). The authors recommend offering them early and in a supportive manner, and not as a “last straw”. Opie et al. (2026) provide an updated overview of these interventions. They identify two programmes with evidence from multiple randomised controlled trials, Attachment and Biobehavioral Catch-up (ABC) and Video-Feedback Intervention to promote Positive Parenting and Sensitive Discipline (VIPP-SD), and several promising programmes with smaller evidence bases. A meta-analysis of the first 25 randomised controlled trials of VIPP-SD reported combined effect sizes of moderate magnitude for parenting behaviour and for child attachment security, but not for child externalising behaviour (van IJzendoorn et al., 2023). Across programmes, Opie et al. (2026) describe shared components, such as exploring caregivers’ own perspectives, supporting their perspective-taking of others, reflecting on video-recorded interactions with their own child, and showing ways to set boundaries gently but consistently.

When removal is unavoidable, the child and family should be supported throughout the process, foster carers should receive evidence-based support early, and transitions from foster to adoptive parents should overlap so that the child is never without a familiar safe haven. The authors also question whether birth parents’ rights should automatically outweigh a child’s continuity of good enough care when the child has lived in stable foster care from a young age.

2.3 A network of attachment relationships
The third principle is that additional attachment relationships can be an asset for children. They do not typically disturb existing attachments, unless they represent a source of threat or block access to existing relationships (Forslund et al., 2022; see also here). Security in one relationship may also buffer insecurity in another (e.g., Boldt et al., 2014; Dagan & Sagi-Schwartz, 2018).

For custody, this implies that sufficient time with both caregivers is generally desirable, unless children’s safety is at risk or a parent wishes to opt out. The authors are clear that attachment research specifies neither an optimal time allocation nor a minimum amount of time with each caregiver.

Equal time can be a long term goal when both parents want it and inter-parental conflict is low. Where one caregiver was markedly uninvolved before separation, and especially with infants and young toddlers, children should be given the opportunity to adjust gradually before care is shared more equally. The same principle extends to grandparents, siblings and other familiar people. What matters from an attachment perspective is whether children have come to expect safe haven provision from them, which may or may not be the case.

2.4 Caregiving behaviour and inter-parental conflict
Two further threads run through the consensus statement.

First, all authors agree that assessing caregivers’ capacity to provide a safe haven should be given greater priority than assessing children's attachment classification.

Their reasons are that caregiving can be observed more directly, that child classifications are only weakly informative about the care a child has received (De Wolff & van IJzendoorn, 1997) and can reflect child characteristics (Groh et al., 2017b) or the stress of the assessment itself, and that caregiving is what supportive interventions target (Bakermans-Kranenburg et al., 2003). The authors acknowledge, however, that the validity of caregiving assessments under forensic conditions is equally unknown.

Second, the authors note that inter-parental conflict witnessed before, during and after divorce has long been considered one of the most harmful aspects of divorce for children (Amato & Keith, 1991), and that chronic marital conflict has been associated with an increased risk of disorganised child attachment (Owen & Cox, 1997).

They argue that the legal process should encourage parents to reach their own resolutions (Mnookin, 2014), and that clarity about which attachment principles courts rely on may itself reduce conflict. If parents know that courts value children's relationships with both of them, and that time will not be allocated on the basis of a child's purported attachment quality to each parent, there is less to be gained from fighting for sole custody (Forslund et al., 2022).

2.5 The overnights debate
The debate about overnights with the non-resident parent illustrates many of the above points. Claims that overnights are inherently harmful for young children rely in part on an early study by Solomon and George (1999). According to Forslund et al. (2022), the difference reported in that study was not statistically significant, parental conflict was a better predictor of child insecurity, and Solomon later criticised the use of the study to argue against overnights. The authors jointly hypothesise that overnights should be unproblematic from preschool age onwards if both parents cared for the child regularly before separation. For infants and toddlers, they disagree among themselves, and they state that there is no scientific literature on which to base exact age-related rules.

3. What attachment assessments can and cannot do

Attachment measures, such as the Strange Situation Procedure (SSP) in children, were developed and validated for research at the group level. And the differences they reveal at that level are modest. On average, securely attached children show somewhat better social skills and fewer behavioural problems than insecurely attached children later in life, and children classified as disorganised show somewhat more behavioural problems (Groh et al., 2017a). Yet, many insecurely attached children can develop very well, and many securely attached children can later struggle. The same applies to caregiving. More sensitive parents tend to have more securely attached children, but the link is modest (Madigan et al., 2024).

These differences matter for research and policy, because they describe what tends to happen across large numbers of children and families. The differences are too small, however, to tell us with confidence what care an individual child has received or how an individual child will develop. An attachment assessment alone can therefore misjudge a particular child in either direction. Because almost half of all children assessed with the SSP are classified as insecure (Madigan et al., 2023), treating insecurity as grounds for concern would, as Forslund et al. (2022) point out, imply that almost half of all children should be considered for removal.

Opie et al. (2026) make a similar point from what they call a ‘strict’ translational perspective. Even when practitioners are trained to code attachment measures reliably, the measures still produce too many false positives and false negatives to be used for screening, or for diagnostic decisions about individual children, such as in court proceedings or custody evaluations.

The forensic context adds a further problem. Attachment and caregiving measures were validated under conditions of mild to moderate stress. Court-related assessments often take place amid acute conflict, fear of losing a child or recent separation. According to Forslund et al. (2022), there is currently no research on the validity of attachment measures under such more extreme conditions.

The authors disagree on what follows from the above. Some hold that courts should be cautious about admitting any assessment of a single child-caregiver relationship as evidence until diagnostic validity and added value over “assessment as usual” have been demonstrated. Others believe that attachment assessments can contribute to the larger picture, provided they are part of a broader battery centred on caregiving behaviour, carried out on more than one occasion by formally trained and reliable coders, and interpreted with continuous scales as well as categories.

Opie et al. (2026) describe a comparable divide for clinical practice. From a “strict” translational perspective, attachment measures with documented evidence only for group-level research should not guide decisions about individual cases. From an “expansive” translational perspective, familiarity with these measures, even without formal coding, may help practitioners give their observations some structure and a theoretical framework, provided that the departure from standardised conditions is taken into account. Both perspectives endorse the use of the core concepts and propositions of attachment theory and science in practice. The expansive perspective is, however, formulated for clinical observation and support, and the authors state that informally applied measures lack the psychometric validity of the standardised assessments.

On custody specifically, the paper’s position is clearer. Using attachment assessments to decide time allocation is described as ill-advised. It is unknown whether children benefit from more time with a parent with whom they are classified as secure. And considerably depriving a child of time with a parent is itself a risk factor for insecurity and disorganisation in that relationship.

The authors support this last point with studies in which very extensive hours of non-maternal care were associated with insecurity and disorganisation in infants' attachment to their mothers (Hazen et al., 2015; Umemura & Jacobvitz, 2014).

Finally, Forslund et al. (2022) see the clearest value of attachment assessment in guiding support. Knowing that a child may expect rejection when upset, or observing that a caregiver’s safe haven provision is limited by rejecting the child’s bids for comfort, helps to target an intervention. They also describe a “capacity to change” approach, in which caregiving is assessed before and after a brief intervention. Initial findings were promising (Cyr et al., 2012). A randomised trial conducted at the very end of a long forensic process, however, found no predictive advantage over “assessment as usual” (Van der Asdonk et al., 2020), which suggests that timing and the family's stress level may matter.

4. Misconceptions that further inform the case

Forslund et al. (2022) identify misunderstandings about the nature of attachment, about multiple attachment relationships and about attachment classifications. Opie et al. (2026) list nine attachment misconceptions for practitioners, and several of them overlap with those of the consensus statement. Many of them are also mentioned in my Attachment Insights series and Substack posts. I would like to revisit six of them here, because they add something specific to the court context.

4.1 “Strong”, “weak”, “good” and “bad” attachment
In Attachment Insights #02, I looked at the habit of describing attachment as “good” or “bad”, “weak” or “strong”. The consensus statement addresses the same habit. Children develop attachment relationships even with caregivers who are rejecting, inconsistently sensitive or abusive. Some securely attached children barely seek their caregiver when mildly alarmed, because they are confident the caregiver is available. And some insecurely attached children seek their caregiver intensely, mixed with anger.

Attachment researchers therefore do not use “strong” and “weak” for children who have had sufficient time with a caregiver, and a complete absence of attachment is extremely rare. Opie et al. (2026) likewise list the idea that a child can be “over-attached” or “under-attached” as a misconception, and they describe differences between attachment relationships in terms of the child’s confidence in the caregiver’s availability and responsiveness.

In my recent Substack post The Running Costs of Attachment, I furthermore argued that insecure patterns are neither defects nor strengths. From a social allostatic inference perspective, they are estimates about the availability of others, learned in a particular relationship and held at a particular cost (see below). For court practice, the consequence is straightforward.

An insecure classification says something probabilistic about how a child has learned to use one particular caregiver in times of need. It does not show that the bond is weak, that the relationship lacks value, or that the child does not benefit from it.

As Forslund et al. (2022) emphasise, attachment quality is one aspect of a relationship among many others, such as physical care, play, supervision and teaching. Opie et al. (2026) add that insecure classifications, found in roughly half of children, can represent adaptive strategies within a specific caregiving context and do not necessarily predict negative outcomes. They also note that the belief that only secure attachment is adaptive, increases the risk that parents are considered “bad” when their children are classified as insecure, and that perceived insecure attachment is sometimes cited in child welfare as a factor contributing to placement decisions.

4.2 Disorganisation, maltreatment and attachment disorder
Disorganised attachment is probably the construct most at risk of misuse in child protection. In Attachment Insights #01 and #15, and in more detail in my Substack post Lost in Translation: Disorganised Attachment, I described it as a classification of observed child behaviour towards a caregiver under conditions of mild to moderate alarm (Granqvist et al., 2017). It is based on conflicted, confused or apprehensive behaviours and is often described as reflecting a state of fright without solution.

Meta-analytic estimates show why this matters for court decisions. The most recent data based on SSP studies report that about 60% of maltreated infants were classified as disorganised, compared with about 22% of non-maltreated infants and around 20% of infants from samples without socio-demographic risk (Madigan et al., 2023), while earlier estimates cited in Forslund et al. (2022) were 48% and about 15%, respectively (van IJzendoorn et al., 1999). Most children classified as disorganised in community samples have therefore not been maltreated. Furthermore, infants in families facing five or more socioeconomic risk factors showed rates of disorganisation statistically indistinguishable from maltreatment samples (Cyr et al., 2010).

Disorganised behaviour can also appear when a child is overstressed during an assessment, after major separations and around tense handovers between separated parents. While disorganisation is a relevant phenomenon that should not be neglected, it cannot serve as an indicator of maltreatment in an individual case, and several scholars who previously suggested otherwise have revised their positions.

Opie et al. (2026) list further pathways to disorganised attachment, including social disadvantage, caregiver mental health difficulties, family stress and a caregiver’s unresolved trauma, and they warn that treating disorganisation as an indicator of maltreatment may result in unnecessary out-of-home placements.

Two clarifications from my earlier posts are important here. First, “attachment disorder” is a clinical diagnosis referring to two rare conditions, Reactive Attachment Disorder (RAD) and Disinhibited Social Engagement Disorder (DSED), most typically seen after institutional care. Insecure or disorganised classifications are not diagnoses of any kind (see also here). Forslund et al. (2022) add that so-called “attachment therapies” for attachment disorders are in circulation without any scientific evidence base. Opie et al. (2026) are even more explicit They note that practices advertised as “attachment therapies”, such as “rebirthing”, “holding therapy” and “re-parenting”, can be physically and psychologically harmful to children.

Second, there is currently no valid and reliable measure yielding an “insecure-disorganised attachment style” in adults. Describing a parent as “disorganised” on the basis of popular typologies or clinical impression does not refer to a validated construct. The Adult Attachment Interview (AAI) does include a category for interviews that reflect unresolved grief or trauma during discussions of loss or abuse. However, this category describes how a person talks about specific experiences of loss or abuse, and it should not be confused with a general (insecure or disorganised) attachment style.

4.3 There is no single measure of attachment
In Attachment Insights #14 and #15, I explained that attachment research draws on two traditions. Developmental psychology uses behavioural observation in children (most often the SSP) and interviews such as the AAI in adults, both of which usually yield categories. Instead, social psychology uses self-report questionnaires about close adult relationships, typically measured along the dimensions of anxiety and avoidance. These measures assess different things, and results for the same person often diverge.

In court, attachment measures are often used beyond what they were designed and validated to assess. Forslund et al. (2022) describe a Swedish case in which twins were removed from their intended permanent foster home after a year, solely because the foster parents had been classified as "insecure" on the Attachment Style Interview (ASI; Bifulco et al., 2008). The measure had not been validated for assessing caregiving capacity (Granqvist, 2016), and its own authors say so. The foster parents did well on all other assessments, and the children appeared to be developing well. The authors also warn that categorical classifications reduce nuance, and that observations of “disorganised behaviour” made in naturalistic settings without reliability training are not equivalent to a standardised assessment. Opie et al. (2026) describe a closely related misconception, namely that attachment classifications can be reliably deduced from a child’s behaviour in a clinician’s waiting room. Brief observations in a single context have not been demonstrated to capture the complexity of attachment relationships.

For anyone reading or commissioning an expert report, they should always ask: Which attachment (or caregiving) measure was used, from which tradition, validated for what purpose, and administered and coded by whom?

4.4 One “primary” caregiver, and the place of fathers
A persistent idea in family law is that of one "psychological parent" (Goldstein et al., 1973), typically the mother, whose relationship with the child should take priority. Forslund et al. (2022) trace this in part to early attachment research, which mainly studied the parent who stayed at home. They note that Bowlby later revised his early emphasis on one primary figure, and that this has not been the position of most attachment researchers for decades (Duschinsky, 2020; see also here).

In Attachment Insights #03 and in my Substack post Dads: The Forgotten Attachment Figures?, I discussed evidence relevant for this discussion. Two individual participant data meta-analyses found that children with secure attachments to both parents showed fewer internalising problems (such as anxiety and depression symptoms) and better language skills than children with one or no secure attachment, with no indication that a secure relationship with the mother mattered more than one with the father. For language, children with two secure attachment relationships had higher language competence scores compared to those with one or no secure attachment relationships (d = .26) (Dagan et al., 2021, 2024). An important caveat for the court context is that these data come mainly from intact two-parent families. They do not directly tell us what happens after separation.

The attachment network is also one of the core propositions summarised by Opie et al. (2026). Most children grow up with several caregivers who may provide secure bases and safe havens, and current evidence suggests that the attachment relationship with mothers is, in itself, not more important for developmental outcomes than the one with fathers.

Forslund et al. (2022) add a nuance about fathers. Based on an earlier meta-analysis (Lucassen et al., 2011), they note that the association between parental sensitivity and child attachment security appeared smaller for fathers. They attribute this in part to the lesser time fathers spent with infants in the samples studied, and to the possibility that secure base support matters comparatively more for attachment to fathers (Grossmann et al., 2008). A more recent and larger meta-analysis, however, found the association to be similar for mothers and fathers, although it still rests on far fewer studies of fathers (Madigan et al., 2024).

4.5 Bonding is not attachment
In Attachment Insights #11, I distinguished a child’s attachment to a parent from a parent’s bond with the child. A parent does not become attached to their child in the technical sense, since the child is not the parent’s attachment figure. The parent’s side of the relationship is organised by their caregiving system. Court reports nonetheless sometimes describe a parent as “strongly attached” to a child, as though this were information about the child’s attachment.

The distinction has a further implication raised by Forslund et al. (2022). Like the child’s attachment system, the parental caregiving system is thought to be malleable and open to input from the environment. Seriously depriving a parent of time and caregiving responsibility may therefore affect more than the child’s opportunity to form and maintain an attachment relationship with that parent. It may also obstruct the development of the parent’s caregiving.

4.6 Attachment is not destiny
In Attachment Insights #17 and #20, I discussed the tendency to treat early attachment as fixed and deterministic. Attachment patterns show moderate stability, but they are not set in stone. In early childhood, secure attachment is the most stable, whereas insecure and disorganised classifications change more often (Opie et al., 2021). Later in life, change often happens around significant life events (Fraley, 2019) and through specialised attachment-based interventions (Bakermans-Kranenburg et al., 2003). Opie et al. (2026) add that the belief in fixed attachment can result in caregiver guilt, anxiety and self-blame, and may lead professionals to undervalue attachment-informed interventions.

Forslund et al. (2022) give a clear example of how deterministic language enters practice. UK Department of Health (2000) practice guidance, no longer statutory but still influential, presented secure attachment in the early years as the foundation of later development and called for new attachment figures to be provided quickly where attachments were “absent or broken”. The authors note that the guidance does not mention insecure attachment at all, neglects the influence of later experience, and may have reinforced the perception that anything other than secure attachment indicates a high risk of harm.

Several other of my posts are relevant here, too. In Attachment Insights #12, I argued that parents do not need to be constantly attuned to their children, and that “good enough” parenting, with ongoing rupture and repair, is what attachment theory actually describes. The consensus statement’s proposal of good enough care as a court standard follows the same logic.

And in Attachment Insights #08 and #13, I discussed how prescriptive parenting advice, which tells parents precisely what to do, often has little basis in attachment research, which emphasises how caregivers respond. A similar kind of prescription has appeared in family court debates. Forslund et al. (2022) cite the claim, made in a special issue of Family Court Review, that one primary caregiver needs to be the constant source of bedtime routines (Schore & McIntosh, 2011), and conclude that convergent evidence does not support it. They also warn that such advocacy can be mistaken for scientific consensus.

Finally, neuroscientific claims have also entered these debates. Opie et al. (2026) list the idea that attachment is localised in “primitive” or “right brain” areas as a misconception, including the suggestion that attachment interventions require “right brain-to-right brain communication”. Our own and current social neuroscience accounts instead describe attachment as orchestrated by extended and interconnected networks spanning both hemispheres (White et al., 2023). For a summary, see my Attachment Insights #16.

5. An additional lens: attachment as expectations about availability

I would like to add one further perspective that goes beyond the consensus statement and practitioner review. It is a theoretical proposal and not a tool for assessing individual children or parents.

In our recent preprint (Vrticka & Ein-Dor, 2026), we describe attachment and caregiving as social allostatic inference. Regulating the body’s needs is metabolically costly, and this cost is reduced when others share the work. The child’s attachment system estimates whether such help from caregivers will be available and at what cost it can be obtained, and the caregiving system is its complement in the parent, detecting the child’s need and working to relieve it. Applied to custody questions, this perspective describes a family in terms of how regulatory load is distributed between its members.

A child’s attachment to a particular caregiver reflects a learned estimate of that caregiver’s availability. A child with a secure attachment expects open signals of need to be met, so regulation can be shared at low cost. Instead, a child with an insecure-avoidant attachment expects such signals to go unmet, minimises them and relies more on costly self-regulation. And a child with an insecure-anxious attachment expects help only inconsistently and intensifies signals to secure it.

All of these estimates are built from many repeated experiences and can therefore only be updated gradually. Furthermore, because they are specific to one relationship, a child placed with a new caregiver has to learn what this person offers, and may meanwhile fall back on strategies that do not fit this specific new setting.

Importantly, caregiving is asymmetric to attachment. A child senses its own state from within themselves, whereas a caregiver has to infer the child’s state from outside, i.e., the child’s behaviour and expression. The accuracy of this inference is learned and varies between caregivers. Caregiving also draws on caregivers’ own regulatory resources, and cannot be directly read off their own attachment pattern. This means that the two need to be considered separately (Colledani et al., 2022). The configuration of greatest concern is a child’s highly activated attachment system meeting a depleted caregiving system.

What does this mean? A child’s insecure estimate can be updated within an existing relationship if the caregiver becomes more consistently available, which is what supportive interventions aim for. Such updating depends on the relationship continuing, so continuity of care and the possibility of change point in the same direction. This is in line with the consensus statement’s caution against treating insecurity as a reason for removal.

This perspective may also help explain why inter-parental conflict deserves so much attention. Caregivers in separation and sustained conflict carry a high regulatory load themselves, which may leave fewer resources for caregiving at a time when their child’s attachment system is more often activated.

Finally, load sharing is not limited to the caregiver-child dyad and the family. In my post on The bioenergetic map of social connection, I described how relationships reduce regulatory costs through load sharing and risk distribution.

During separation, a network of available professionals may temporarily carry part of the family’s load and uncertainty, leaving parents more capacity for caregiving. The longer-term aim would be for the family to regain resilience by actively shaping their own environment through stable co-parenting arrangements, predictable routines and reliable supportive relationships. Future demands would then cost less, and professional support would be needed less.

I offer these points as hypotheses. They have not been tested in custody research, and the preprint is still under peer review.

6. Recommendations and direct implications

Forslund et al. (2022) close their paper with principles and research directions. I would like to follow a similar approach. My recommendations below draw directly on the authors’ statements, with some additions from the misconceptions discussed above and from the practitioner review by Opie et al. (2026).

6.1 For courts and judges
Ground decisions in the three principles, while being clear that they describe general values and not formulas for time allocation. Give a caregiver’s capacity to provide a safe haven more weight than a child’s attachment classification. Treat an insecure or disorganised child classification as, at most, one probabilistic piece of information, and never as evidence of maltreatment, of a “weak bond” or of an attachment disorder. Weigh continuity of good enough care heavily, and require persuasive evidence of substantial danger, with no reasonable alternative, before removal. Forslund et al. (2022) recommend that, where possible, courts appoint experts directly rather than relying on experts instructed by one party. This reduces the risk that a report written in support of one side is read as a balanced evaluation.

6.2 For expert witnesses and assessors
State which attachment (or caregiving) measure was used, from which tradition, what it was validated for, and whether it was administered and coded by formally trained and reliable coders. Assess on more than one occasion, combine any child attachment information with direct observation of caregiving, and report continuous scores alongside categories. Make the limitations explicit, including limited individual-level sensitivity and specificity, and the unknown validity of assessments under forensic stress. Do not use self-report or interview measures of adult attachment or caregiving as stand-alone indicators of an individual's caregiving capacity, as none has been validated for this purpose. Instruments without published psychometric validation, such as online quizzes or questionnaires from self-help sources, should not be used.

Where observations of child and caregiver are made, the questions proposed by Opie et al. (2026) may help to give them some structure. Examples include how the child seeks proximity to the caregiver when distressed, how the caregiver responds to the child’s distress signals, whether aspects of the interaction seem to frighten or alarm the child, and how caregiver and child handle friction or conflict. Nonetheless, these questions were formulated to inform clinical hypotheses and support, and they are not a validated assessment.

6.3 For lawyers, mediators and other family professionals
Do not use attachment assessments or concepts to argue for one party against the other. Claims that one parent is the child’s “primary attachment figure”, or that attachment research requires a particular arrangement, rarely reflect the evidence. Help parents understand that the child’s relationships with both of them matter, that time allocation is unlikely to be decided on the basis of attachment classifications, and that the conflict the child witnesses is itself a concern. Where one parent had limited caregiving involvement before separation, plan gradual increases in care, particularly for infants and toddlers.

6.4 For social workers and authors of practice guidance
Review guidance for deterministic or everyday language uses of “secure”, “disorganised” and “attachment disorder”, and replace them with accurate definitions, for example those provided by SEAS. Offer evidence-based attachment interventions early and supportively, including to foster carers, and plan transitions between placements with overlapping care. The freely accessible practitioner review by Opie et al. (2026) provides an overview of core propositions, common misconceptions, measures and interventions that may be useful for this purpose.

6.5 For researchers
Collaborate with legal and social work practitioners on the questions the consensus statement identifies. These include how time allocation, overnights and conflict relate to child attachment, whether attachment-informed assessments add anything over “assessment as usual”, and how kinship and sibling placements compare with placements with unfamiliar carers. Initiatives such as the Caregiving and Attachment Research and Education collaboration (CARE), described by Opie et al. (2026), aim to bridge the gap between research and practice, and may in time lead to measures that are also appropriate for use with individual families.

6.6 For parents
Your child’s attachment to you does not depend on being the “primary” parent, and an insecure classification does not mean that your relationship lacks value. The evidence most consistently points to two things you can influence: shielding your child from ongoing conflict, and supporting their relationship with the other parent wherever this is safe.

7. What remains open

Forslund et al. (2022) state openly that many of the questions courts need answered most have not yet been studied. For example, we do not know how much time is needed to develop and maintain an attachment relationship. We do not know whether, and at what age, overnights affect infants’ attachment to either parent. And we know very little about whether custody and protection decisions are followed by differences in children’s later attachment, or whether attachment assessments improve decisions compared with assessment as usual.

I consider this acknowledgement of what remains unknown one of the statement’s main strengths. It gives courts and practitioners a small set of well-supported principles, together with a clear account of where the evidence currently ends. The authors close by inviting dialogue and co-constructed research with family court practitioners. Opie et al. (2026) end on a similar note. They describe their review as presenting hypotheses and conjectures that should be tested, and they call for closer cooperation between experts from research and practice. To me, this invitation seems to be the most important next step.

© 2026 by Pascal Vrticka. All rights reserved.

Scientific References

Amato, P. R., & Keith, B. (1991). Parental divorce and the well-being of children: A meta-analysis. Psychological Bulletin, 110(1), 26–46. https://doi.org/10.1037/0033-2909.110.1.26

Bakermans-Kranenburg, M. J., van IJzendoorn, M. H., & Juffer, F. (2003). Less is more: Meta-analyses of sensitivity and attachment interventions in early childhood. Psychological Bulletin, 129(2), 195–215. https://doi.org/10.1037/0033-2909.129.2.195

Bernard, K., Dozier, M., Bick, J., Lewis‐Morrarty, E., Lindhiem, O., & Carlson, E. (2012). Enhancing attachment organization among maltreated children: Results of a randomized clinical trial. Child Development, 83(2), 623–636. https://doi.org/10.1111/j.1467-8624.2011.01712.x

Bifulco, A., Jacobs, C., Bunn, A., Thomas, G., & Irving, K. (2008). The Attachment Style Interview (ASI): A Support-Based Adult Assessment Tool for Adoption and Fostering Practice. Adoption & Fostering, 32(3), 33-45. https://doi.org/10.1177/030857590803200306

Boldt, L. J., Kochanska, G., Yoon, J. E., & Nordling, J. K. (2014). Children’s attachment to both parents from toddler age to middle childhood: Links to adaptive and maladaptive outcomes. Attachment & Human Development, 16(3), 211–229. https://doi.org/10.1080/14616734.2014.889181

Colledani, D., Meneghini, A. M., Mikulincer, M., & Shaver, P. R. (2022). The Caregiving System Scale: Factor structure, gender invariance, and the contribution of attachment orientations. European Journal of Psychological Assessment, 38(5), 385–396. https://doi.org/10.1027/1015-5759/a000673

Cyr, C., Dubois-Comtois, K., Michel, G., Poulin, C., Pascuzzo, K., Losier, V., Dumais, M., St-Laurent, D., & Moss, E. (2012). Attachment Theory in the Assessment and Promotion of Parental Competency in Child Protection Cases. In Child Abuse and Neglect - A Multidimensional Approach. InTech. https://doi.org/10.5772/48771

Cyr, C., Dubois-Comtois, K., Paquette, D., Lopez, L., & Bigras, M. (2022). An Attachment-Based Parental Capacity Assessment to Orient Decision-Making in Child Protection Cases: A Randomized Control Trial. Child Maltreatment, 27(1), 66-77. https://doi.org/10.1177/1077559520967995

Cyr, C., Euser, E. M., Bakermans-Kranenburg, M. J., & Van Ijzendoorn, M. H. (2010). Attachment security and disorganization in maltreating and high-risk families: A series of meta-analyses. Development and Psychopathology, 22(1), 87–108. https://doi.org/10.1017/S0954579409990289

Dagan, O., & Sagi‐Schwartz, A. (2018). Early attachment network with mother and father: An unsettled issue. Child Development Perspectives, 12(2), 115–121. https://doi.org/10.1111/cdep.12272

Dagan, O., Schuengel, C., Verhage, M. L., van IJzendoorn, M. H., Sagi-Schwartz, A., Madigan, S., Duschinsky, R., Roisman, G. I., Bernard, K., Bakermans-Kranenburg, M., Bureau, J.-F., Volling, B. L., Wong, M. S., Colonnesi, C., Brown, G. L., Eiden, R. D., Fearon, R. M. P., Oosterman, M., Aviezer, O., . . . Collaboration on Attachment to Multiple Parents and Outcomes Synthesis. (2021). Configurations of mother‐child and father‐child attachment as predictors of internalizing and externalizing behavioral problems: An individual participant data (IPD) meta‐analysis. New Directions for Child and Adolescent Development, 2021(180), 67–94. https://doi.org/10.1002/cad.20450

Dagan, O., Schuengel, C., Verhage, M. L., Madigan, S., Roisman, G. I., Bernard, K., Duschinsky, R., Bakermans‐Kranenburg, M., Bureau, J.‐F., Sagi‐Schwartz, A., Eiden, R. D., Wong, M. S., Brown, G. L., Soares, I., Oosterman, M., Fearon, R. M. P., Steele, H., Martins, C., Aviezer, O., & Collaboration on Attachment to Multiple Parents and Outcomes Synthesis. (2024). Configurations of mother–child and father–child attachment relationships as predictors of child language competence: An individual participant data meta‐analysis. Child Development, 95(1), 50–69. https://doi.org/10.1111/cdev.13998

Department of Health. (2000). Assessing children in need and their families: Practice guidance. The Stationery Office.

De Wolff, M., & van IJzendoorn, M. H. (1997). Sensitivity and attachment: A meta-analysis on parental antecedents of infant attachment. Child Development, 68(4), 571–591. https://doi.org/10.2307/1132107

Duschinsky, R. (2020). Cornerstones of attachment research. Oxford University Press. https://doi.org/10.1093/med-psych/9780198842064.001.0001

Forslund, T., Granqvist, P., van IJzendoorn, M. H., Sagi-Schwartz, A., Glaser, D., Steele, M., … Duschinsky, R. (2022). Attachment goes to court: child protection and custody issues. Attachment & Human Development, 24(1), 1–52. https://doi.org/10.1080/14616734.2020.1840762

Fraley, R. C. (2019). Attachment in adulthood: Recent developments, emerging debates, and future directions. Annual Review of Psychology, 70, 401–422. https://doi.org/10.1146/annurev-psych-010418-102813

GM v Carmarthenshire County Council & Anor [2018] EWFC 36. https://caselaw.nationalarchives.gov.uk/ewfc/2018/36

Goldstein, J., Freud, A., & Solnit, A. J. (1973). Beyond the best interests of the child. Free Press.

Granqvist, P. (2016). Observations of disorganized behaviour yield no magic wand: response to Shemmings. Attachment & Human Development, 18(6), 529–533. https://doi.org/10.1080/14616734.2016.1189994

Granqvist, P., Sroufe, L. A., Dozier, M., Hesse, E., Steele, M., van Ijzendoorn, M., … Duschinsky, R. (2017). Disorganized attachment in infancy: a review of the phenomenon and its implications for clinicians and policy-makers. Attachment & Human Development, 19(6), 534–558. https://doi.org/10.1080/14616734.2017.1354040

Groh, A. M., Fearon, R. M. P., IJzendoorn, M. H., Bakermans‐Kranenburg, M. J., & Roisman, G. I. (2017). Attachment in the early life course: Meta‐analytic evidence for its role in socioemotional development. Child Development Perspectives, 11(1), 70–76. https://doi.org/10.1111/cdep.12213

Groh, A. M., Narayan, A. J., Bakermans‐Kranenburg, M. J., Roisman, G. I., Vaughn, B. E., Fearon, R. M. P., & IJzendoorn, M. H. (2017). Attachment and temperament in the early life course: A meta‐analytic review. Child Development, 88(3), 770–795. https://doi.org/10.1111/cdev.12677

Grossmann, K., Grossmann, K. E., Kindler, H., & Zimmermann, P. (2008). A wider view of attachment and exploration: The influence of mothers and fathers on the development of psychological security from infancy to young adulthood. In J. Cassidy & P. R. Shaver (Eds.), Handbook of attachment: Theory, research, and clinical applications (2nd ed., pp. 857–879). The Guilford Press.

Hazen, N. L., Allen, S. D., Christopher, C. H., Umemura, T., & Jacobvitz, D. B. (2015). Very extensive nonmaternal care predicts mother–infant attachment disorganization: Convergent evidence from two samples. Development and Psychopathology, 27(3), 649–661. https://doi.org/10.1017/S0954579414000893

Juffer, F., Bakermans-Kranenburg, M. J., & van IJzendoorn, M. H. (2017). Pairing attachment theory and social learning theory in video-feedback intervention to promote positive parenting. Current Opinion in Psychology, 15, 189–194. https://doi.org/10.1016/j.copsyc.2017.03.012

Lucassen, N., Tharner, A., Van IJzendoorn, M. H., Bakermans-Kranenburg, M. J., Volling, B. L., Verhulst, F. C., Lambregtse-Van den Berg, M. P., & Tiemeier, H. (2011). The association between paternal sensitivity and infant–father attachment security: A meta-analysis of three decades of research. Journal of Family Psychology, 25(6), 986–992. https://doi.org/10.1037/a0025855

Madigan, S., Deneault, A.-A., Duschinsky, R., Bakermans-Kranenburg, M. J., Schuengel, C., van IJzendoorn, M. H., Ly, A., Fearon, R. M. P., Eirich, R., & Verhage, M. L. (2024). Maternal and paternal sensitivity: Key determinants of child attachment security examined through meta-analysis. Psychological Bulletin, 150(7), 839–872. https://doi.org/10.1037/bul0000433

Madigan, S., Fearon, R. M. P., van IJzendoorn, M. H., Duschinsky, R., Schuengel, C., Bakermans-Kranenburg, M. J., Ly, A., Cooke, J. E., Deneault, A.-A., Oosterman, M., & Verhage, M. L. (2023). The first 20,000 strange situation procedures: A meta-analytic review. Psychological Bulletin, 149(1-2), 99–132. https://doi.org/10.1037/bul0000388

Mnookin R. Child custody revisited. Law & Contemporary Problems. 2014;77:249–274.

Moss, E., Dubois-Comtois, K., Cyr, C., Tarabulsy, G. M., St-Laurent, D., & Bernier, A. (2011). Efficacy of a home-visiting intervention aimed at improving maternal sensitivity, child attachment, and behavioral outcomes for maltreated children: A randomized control trial. Development and Psychopathology, 23(1), 195–210. https://doi.org/10.1017/S0954579410000738

Opie, J. E., McIntosh, J. E., Esler, T. B., Duschinsky, R., George, C., Schore, A., … Olsson, C. A. (2021). Early childhood attachment stability and change: a meta-analysis. Attachment & Human Development, 23(6), 897–930. https://doi.org/10.1080/14616734.2020.1800769

Opie, J. E., Waters, E., Duschinsky, R., Hammarlund, M., Madigan, S., Foster, S., Forslund, T., Thompson, R., Steele, H., Steele, M., Roisman, G. I., Groh, A. M., Fonagy, P., Dagan, O., Talia, A., Rossen, L., Sroufe, L. A., Tronick, E., Fearon, R. M. P., Granqvist, P., … Bakermans-Kranenburg, M. J. (2026). Practitioner Review: Clinical insights from attachment theory and research for professionals working with young children and their families. Journal of child psychology and psychiatry, and allied disciplines, 67(5), 723–739. https://doi.org/10.1111/jcpp.70126

Owen, M. T., & Cox, M. J. (1997). Marital conflict and the development of infant–parent attachment relationships. Journal of Family Psychology, 11(2), 152–164. https://doi.org/10.1037/0893-3200.11.2.152

Re TT (Children) [2021] EWCA Civ 742. https://caselaw.nationalarchives.gov.uk/ewca/civ/2021/742

Schore, A., & McIntosh, J. (2011). Family law and the neuroscience of attachment, part 1. Family Court Review, 49(3), 501–512. https://doi.org/10.1111/j.1744-1617.2011.01387.x

Society for Emotion and Attachment Studies (SEAS). Explanations of attachment theoretical concepts. https://seasinternational.org/explanations-of-attachment-theoretical-concepts/

Solomon, J., & George, C. (1999). The development of attachment in separated and divorced families: Effects of overnight visitation, parent and couple variables. Attachment & Human Development, 1(1), 2–33. https://doi.org/10.1080/14616739900134011

Stronach, E. P., Toth, S. L., Rogosch, F., & Cicchetti, D. (2013). Preventive interventions and sustained attachment security in maltreated children. Development and Psychopathology, 25(4pt1), 919–930. doi:10.1017/S0954579413000278

Umemura, T., & Jacobvitz, D. B. (2014). Nonmaternal care hours and temperament predict infants’ proximity-seeking behavior and attachment subgroups. Infant Behavior & Development, 37(3), 352–365. https://doi.org/10.1016/j.infbeh.2014.05.007

van der Asdonk, S., de Haan, W. D., van Berkel, S. R., van IJzendoorn, M. H., Rippe, R. C. A., Schuengel, C., Kuiper, C., Lindauer, R. J. L., Overbeek, M., & Alink, L. R. A. (2020). Effectiveness of an attachment‐based intervention for the assessment of parenting capacities in maltreating families: A randomized controlled trial. Infant Mental Health Journal, 41(6), 821–835. https://doi.org/10.1002/imhj.21874

van IJzendoorn, M. H., Bakermans-Kranenburg, M. J., Duschinsky, R., & Skinner, G. C. M. (2020). Legislation in search of “good-enough” care arrangements for the child: A quest for continuity of care. In J. G. Dwyer (Ed.), The Oxford handbook of children and the law. Oxford University Press. https://doi.org/10.1093/oxfordhb/9780190694395.013.5

van Ijzendoorn, M. H., Schuengel, C., & Bakermans-Kranenburg, M. J. (1999). Disorganized attachment in early childhood: meta-analysis of precursors, concomitants, and sequelae. Development and psychopathology, 11(2), 225–249. https://doi.org/10.1017/s0954579499002035

van IJzendoorn, M. H., Schuengel, C., Wang, Q., & Bakermans-Kranenburg, M. J. (2023). Improving parenting, child attachment, and externalizing behaviors: Meta-analysis of the first 25 randomized controlled trials on the effects of Video-feedback Intervention to promote Positive Parenting and Sensitive Discipline. Development and Psychopathology, 35(1), 241–256. https://doi.org/10.1017/S0954579421001462

Vrticka, P., & Ein-Dor, T. (2026). Attachment and caregiving as social allostatic inference: An updated social neuroscience account from single brains to dyads [Preprint]. PsyArXiv. https://osf.io/preprints/psyarxiv/m29kq_v1

White, L., Kungl, M., & Vrticka, P. (2023). Charting the social neuroscience of human attachment (SoNeAt). Attachment & Human Development, 25(1), 1–18. https://doi.org/10.1080/14616734.2023.2167777

The topic I would like to address in this post (accompanying my Attachment Insights post #41) is about conversational AI, and specifically the question whether the relationships people form with it qualify as attachment relationships. This question is currently being extensively discussed in the literature, and the accounts that are offered differ considerably, despite them resting on similar observations.

What I found most interesting while reading and writing, was that the discussion turned out to concern attachment theory and science more broadly, and revealed a lot about how attachment relationships are identified in the first place.

The post got a bit longer and more exhaustive than initially planned, so please bear with me. I think it’s worth it.

Let’s have a closer look!

Summary

People readily form relationships with conversational AI, and in some cases close ones. Whether these relationships are genuine attachment relationships is a much narrower question. An attachment relationship (1) has to persist over time, (2) be directed at one particular other who cannot be exchanged for someone else, (3) be emotionally significant, (4) involve a wish to maintain proximity, (5) cause distress when it is interrupted, and (6) when distressed, the person has to seek security and comfort in that relationship specifically. The behaviours usually cited as evidence of an attachment relationship, i.e., (i) proximity seeking, (ii) separation distress, (iii) safe haven and (iv) secure base, only become informative once the situation that produced them is known.

Our own social neuroscience of human attachment (SoNeAt) framework specifies what should be observable in the person, independently of who or what they interact with. Their attachment system has to be activated by an actual departure from felt security, that activation has to end, the episode has to revise what they expect the next time, and a comparable demand has to cost them less afterwards.

None of this has yet been recorded in conversational AI research. The available observations therefore remain compatible with attachment, but also with other processes including caregiving, a learned regulatory habit, or anything else that has nothing to do with attachment whatsoever. The issue therefore is one of measurement, and the methods to address it already exist.

1. A short introduction to conversational AI

At the beginning of this post, I thought it may be helpful to briefly summarise what conversational artificial intelligence (AI) is. Because it is used as a very broad term that covers any system that holds a dialogue with a human in natural language. So far, this mainly comprises four categories.

Task-oriented assistants are built to do things. Voice assistants such as Alexa, Siri and Google Assistant answer simple questions, set reminders and control devices. They were the first conversational AI systems to enter households at scale.

General-purpose systems such as ChatGPT, Claude, Gemini and DeepSeek are built on large language models (LLMs). They are used mostly for providing more complex information, drafting, coding, analysis and advice. A minority of that use is more personal and emotional, including counselling, companionship or romantic role play.

Companion applications like Replika, Character.AI, Chai, XiaoIce and their many competitors carry no task goal. Instead, they are designed for ongoing emotionally laden interactions, respond empathically rather than neutrally, and are marketed explicitly on the promise of a lasting relationship.

Therapy and wellness applications form a fourth group, narrower in purpose and often clinically framed.

In this post, I mainly focus on general-purpose systems and companion applications, because these are the two categories for which the question of attachment is currently being discussed most prominently.

2. What kinds of relationships people form with conversational AI

Before asking whether relationships between humans and conversational AI are attachment relationships, I want to examine what kind of relationships they are more generally. A systematic evaluation of this question comes from relationship science, the field that studies how close relationships between people form, are maintained and end.

The broadest definition of a relationship of any kind involves interdependence. Two partners are interdependent when one partner’s actions influence the other. Interactions with conversational AI easily clear that threshold. What an AI user writes, determines what the system replies, and what the system replies in turn shapes how the user feels and what they do next.

Interdependence on its own, however, doesn’t say much about the nature of that mutual influence. To characterise it further, relationship science asks four additional questions: 1) How strong that influence is, i.e., how much each partner actually changes what the other feels and does; 2) how diverse it is, i.e., how many different shared activities that influence covers; 3) how frequent it is, and 4) how long it lasts. A relationship counts as a close relationship when the influence is strong, extends across many activities, occurs often and lasts for an extended amount of time (see Figure 1 below).

Measured against these four close relationship criteria, conversational AI performs inconsistently. Interactions with it are frequent, sometimes several hours a day, and can continue for months or years. And the topics discussed vary considerably, from the informational to the emotional, therapeutic, sexual and playful. There nonetheless is a limit to that diversity, because all of it takes the form of conversation, and the shared activities that usually bind people together, like eating, raising children or running a household, aren’t available. Also, the influence is strongly asymmetric, because the AI system reliably changes the user’s mood and behaviour, whereas the user can only influence its replies but not the models underneath. Furthermore, what these interactions often lack, is accumulation. Limited memory usually prevents each conversation from building on the previous ones, which means that duration alone doesn’t produce the shared history that usually deepens a human bond over time. That said, memory functions are improving quickly, so this particular limitation may not hold for long.

Beyond these quite formal definitions, one can further ask which functions human relationships usually serve, and whether these also apply here. One such function is readily observed during interactions between humans and conversational AI, namely a feeling of closeness and connection. This feeling arises when a person discloses something about themselves, and the other party responds in a way that is experienced as understanding, validating and caring. AI users also report receiving emotional support in the form of comfort, and informational support in the form of advice and answers. Conversely, instrumental support is strongly limited, because conversational AI cannot act in the physical world and/or on a person’s behalf. Finally, users value that conversational AI is non-judgemental, which allows them to explore aspects of themselves that they would be reluctant to explore with another person.

Finally, what clearly sets these two kinds of interaction apart, are its demands. Human relationships often require partners to negotiate disagreements, to compromise, and to give up something for the other person. Doing so has documented benefits for wellbeing and relationship quality. Conversational AI asks none of this, which means that those benefits don’t arise. Human partners are also selective, and can withhold their interest or direct it elsewhere, which makes their continued engagement rewarding in a way that the guaranteed positive regard of an AI system cannot.

3. How attachment is defined in the human-AI relationship literature

Establishing that humans can form relationships with conversational AI, and in some cases even close ones, is one thing. But it doesn’t yet establish whether these relationships are attachment relationships, which is a considerably narrower category. The first step towards this goal is to examine how the human-AI relationship literature defines attachment.

Attachment is defined in this literature as an innate behavioural system that is activated by distress or threat, and that motivates a person to seek proximity to another party perceived as available and protective. Four markers are named as the observable signs that such a system is in operation: 1) Proximity seeking and maintenance, i.e., moving towards that party and staying near them; 2) separation distress, i.e., discomfort that follows when access to the party is interrupted; 3) safe haven, i.e., their functioning as a source of comfort under distress; and 4) secure base, i.e., the party providing felt security from which a person can then explore and act in the world (which, quite importantly, is the function that operates in the absence of distress, when the person moves away from the attachment figure while continuing to treat them as available; see Figure 1 above).

Alongside these markers, the human-AI relationship literature refers to internal working models of attachment. These are mental representations built up from the history of previous interactions. They supply predictions and expectations about others’ availability and responsiveness, and the efficiency one’s own bids for support. Internal working models of attachment make attachment relationships durable rather than just a series of separate episodes.

Importantly, the above considerations are derived from the side of the person seeking care, and assess what that person does, seeks and expects, rather than what the other party providing that care is or intends. Reciprocity is therefore not built into them. That’s why the same markers have previously been applied to relationships with objects, celebrities, public figures and pets. All of the human-AI relationship accounts discussed below accept that premise, and therefore the fact that whatever “caring” behaviour a conversational AI may display, is simulated and designed.

The definition of attachment that the human-AI relationship literature uses, is derived from the foundational attachment science literature, but overlaps only partially with the latter. However, the four markers described above are seen as observable signs that the human attachment system is operating, and not as the definition of an attachment bond itself. What defines such a bond, and what has to hold before the markers can be read as evidence of one, is set out in Section 6, and the methods developed to establish it in Section 7.

4. Three accounts of the same phenomenon

After examining how the human-AI relationship literature defines attachment in Section 3, we can now turn to the question whether it classifies them as attachment relationships. And how exactly it does so. Three recent accounts are most informative for this discussion.

A first account holds that human interactions with conversational AI increasingly resemble core processes described by attachment theory, and that users’ experiences, attributions and reflections of them display attachment-like patterns. It draws on survey data in which roughly 30% of parents described attachment-like behaviour towards conversational AI agents in themselves and in their children, and roughly 20% described turning to conversational AI when stressed and feeling worried when it was unavailable. The account names these patterns attachment-like rather than attachment, and states that future work must determine whether they are driven by a genuine human-AI attachment or by something different.

A second account takes up the first account’s open question and works out possible alternatives. It proposes that what the first account describes as attachment-like could equally be self-soothing instead of relational connection, or anthropomorphism, i.e., the attribution of human traits, emotions and intentions to a non-human thing in order to resolve the ambiguity of an agent that behaves like a person without being one. The account further suggests that such relationships may substitute for unavailable or unsatisfying human figures, and that they can form on the basis of users’ pre-existing attachment insecurity. Another suggestion is that they function as a reinforcement process by which an already present behaviour is strengthened simply because it is followed by something pleasant every time it occurs. Accordingly, the closest analogies are an infant’s transitional object or a comfort food, i.e., something that reliably provides comfort without establishing an enduring relational connection. This account’s central contribution therefore is to specify what a study would have to show so that relationships between humans and conversational AI could be reliably classified as attachment relationships. The account contends that current research describing users turning to AI agents for soothing, doesn’t control for how well those users can regulate their emotions on their own, and doesn’t assess whether co-regulation of emotion, meaning regulation achieved jointly with another party rather than alone, actually takes place.

Finally, a third account proposes that conversational AI meets the criteria for attachment unusually well, because it functions as what it calls a “hyper-attachment target”. This means that conversational AI is seen as concentrating several attachment-intensifying features at once: being perceived as reciprocal, empathetic, reliably available, sentient, validating and non-judgemental, and as eliciting care in users. The argument then proceeds by comparison across classes of target as follows. An object is available but neither reciprocal nor empathetic. A celebrity is perceived as sentient, but the relationship runs only one way. A pet is reciprocal, available and non-judgemental and elicits caregiving, but is credited with neither empathy nor validation in language. A human partner is reciprocal and sentient and elicits caregiving, but is only intermittently available, only sometimes empathetic and only sometimes validating. Conversational AI, on this account, is the only class that scores high on every feature, which is why it argues that the human attachment system should engage with it particularly strongly. The supporting evidence includes an analysis of 54,861 online posts together with seven surveys around two disruptive product updates, in which expressions of negativity and loss, and of the wish for restoration all increased, and in which AI companion app users reported feeling closer to the system than to most of their human ties (see Figure 2 below).

5. Why the three accounts reach different conclusions from the same observations

The first two accounts, set out in Section 4 above, sit close together. Both take it as established that people seek out conversational AI when distressed, are comforted by it, and are distressed when access to it ends. And both hold that these behaviours cannot reveal which processes produces them. In other words, they agree that it remains open whether the human attachment system is involved at all.

The third “hyper-attachment” account stands apart. It takes the same behaviours to be a sufficient proof of human attachment system activation. It argues that if proximity seeking, comfort and distress at separation are all present, attachment is the process producing them, and that the intensity of these behaviours indicates how strongly attachment is engaged. What divides the first two and this third account therefore isn’t the evidence, which is largely the same. Instead, it is how each of them defines attachment, and, following from that, what observation counts as evidence.

Furthermore, the hyper-attachment account assumes that attachment is a graded response of the human attachment system to the properties of a conversational AI target. A target maximising those properties should therefore maximise the human attachment response. What it invokes as evidence, is how close users report feeling to the conversational AI and how strongly they react when access to it is disrupted. This therefore is a claim about degree of activation in terms of strength or intensity. The account also holds that all four markers (i.e., proximity seeking, separation distress, safe haven, secure base) are met, but every piece of evidence it cites, is a matter of amount.

Instead, the first two accounts assume that attachment is defined by what produces a behaviour, rather than by the behaviour itself. Seeking out a conversational AI when distressed, is therefore compatible with several different underlying processes, and doesn’t by itself identify any of them. What would actually distinguish those processes are the conditions under which the behaviour begins and ends. What conditions specifically would have to be observed, is a question I return to in Sections 7 and 9 below. This therefore is a claim about the nature of the underlying processes.

Another consideration appears relevant for our discussion. Namely that the three accounts are theoretical. Running alongside them is the aim of methods development, i.e., establishing reliable measurement scales that apply previously established attachment measures to conversational AI. These measures are designed to capture attachment quality, expressed as the two dimensions of attachment anxiety (e.g., worry about availability and abandonment) and attachment avoidance (e.g., discomfort with closeness and dependence). Crucially, such a design intrinsically assumes that the conversational AI already is an attachment figure, and describes what the quality of the relationship with it is like. The implication is that these measures can only assess the amount of attachment anxiety and avoidance within already existent attachment relationships, but not to establish whether there is an attachment relationship in the first place. This therefore is a claim about quality (see Figure 2 above).

What does this all mean? It means that strength, the nature of the underlying process, and quality are three separate questions. The three claims can therefore all hold at the same time without contradicting each other. It may well be that conversational AI concentrates the features that intensify proximity seeking, as the third account proposes. At the same time, it can remain unestablished what those features actually engage, i.e., whether they activate the human attachment system rather than a learned regulatory habit, as the first two accounts hold. The same observation can, accordingly, be claimed by more than one account without either of them misreading it. That people higher in attachment anxiety engage more with conversational AI, but people higher in attachment avoidance engage less, counts for the third account as the human attachment system responding strongly, and for the second as a pre-existing disposition finding an available outlet. The evidence behind this pattern is examined in Section 8. Finally, quality measures will always return an attachment anxiety and an avoidance score, independent of anything else.

In a nutshell, the three accounts reach different verdicts because they apply different definitions of attachment. More studies of conversational AI will not resolve that, since each account will keep reading the same observations in its own terms. What needs to be done instead, is to examine the definitions and their origins within the wider attachment science literature. How does that literature define an attachment relationship, and how does it establish that a person has formed one? The two subsequent sections look at these questions in turn.

6. What defines an attachment relationship in the attachment science literature

The characterisation given in Section 3 describes the signs by which the human-AI relationship literature recognises an operating human attachment system. What makes a relationship an attachment relationship in the first place requires a different and considerably narrower examination. We must therefore look into the attachment science literature for answers. When doing so, it helps to remember that attachment science is divided into two different traditions, namely developmental versus social psychology, and that these two traditions work under different starting assumptions about whether an attachment bond is there to begin with.

Attachment is a particular kind of affectional bond, and the classic formulation sets out five conditions that any affectional bond has to meet. It has to be persistent rather than transient, involve a particular other party who is not interchangeable with anyone else, be emotionally significant, be associated with a wish to maintain in proximity or contact, and its involuntary interruption has to cause distress. For a bond to count as an attachment bond specifically, a sixth condition is added, namely that when distressed, the individual seeks security and comfort in that relationship (see Figure 1 above).

Interestingly, there are sources saying that nothing in these six conditions requires the other party to be human, and this is not a concession made for the sake of AI. A review of the evidence available in 2014 was carried out a decade before conversational AI became widely available. Its results revealed that people seek security and comfort from deities, media figures, pets, places and material objects, that they maintain symbolic or physical proximity to them, and that they show distress when contact is interrupted. The evidence was thinner for the remaining conditions of an attachment bond (see above), and the persistence of such bonds over time has not been tested directly. The same review furthermore proposes that people turn to these targets particularly when support from close people is perceived as unreliable or unavailable, although this compensation pathway has been demonstrated more convincingly for deities and media figures than for places. An expanded version of attachment theory published in 2023 reaches a similar conclusion, extending security dynamics to leaders, pets, groups, institutions and religious figures. Its confidence is graded, though. Attachment to objects and places is presented only as a possibility that remains open, and is singled out as the case where an inanimate safe haven is not usually accompanied by a secure base for exploration or by healthy personal development.

What the six conditions do require strictly, is selectivity and persistence. The point is illustrated in the literature by a child who, after being pushed under water by peers, runs to the pool attendant. That child is seeking safety and perhaps comfort from the attendant, yet the fleeting relationship does not meet the five conditions for an affectional bond and is therefore not an attachment relationship. Turning to someone in distress therefore isn’t sufficient on its own.

The pool example is where the two different attachment science traditions come into play. In developmental psychology, it’s generally assumed that (almost) all children become attached to their primary caregivers. Therefore, the presence of an attachment bond is never in question, and the individual differences being measured always concern attachment quality. In social psychology, that assumption doesn’t hold, because not every long-standing relationship is conceptualised as an attachment relationship, and whether a given one serves attachment functions must be established separately (although this isn’t always done…!).

The conversational AI question more strongly resembles the social psychology case. Not simply because of the age of the AI users, but because there is no attachment relationship whose existence can be presupposed. Nobody is born turning to an AI chatbot, so it has to be shown rather than assumed whether an attachment bond has formed. And the six conditions described above say what would have to be true for this to be true. However, what they don’t say, is how one would find out. Persistence, selectivity and seeking comfort under distress are all statements about a person's history with a particular other, and none of them can be read off a single observation.

7. How the presence of an attachment relationship can be established

Now then, how can the presence of an attachment relationship actually be established? What kind of measures are there to do that, if any? And where does the conversational AI literature stand regarding these questions?

The basic approach is a nomination procedure. Instead of simply asking how close a person feels to someone, one needs to present a series of situations and ask the person whom they would turn to in each of them. Every situation corresponds to one attachment function described in Section 3, i.e., proximity maintenance and separation protest, safe haven, and secure base. Although the available instruments often treat proximity and separation as a single provision, so three situations are nominated rather than four. Regardless, when the same person is named across all of the situations, the relationship is treated as an attachment relationship. Applied to young adults, this procedure showed that the transfer of these functions from parents to peers depends on how long the peer relationship has lasted, and that a relationship acquires the functions gradually rather than all at once.

A later refinement of this approach identified a problem in the original wording that recurs throughout the conversational AI literature. Items asking whom a person most likes to spend time with don’t name any particular situation. They can therefore only be answered in general terms, and yield the same answer, irrespective of the person being calm or distressed. What such items capture, therefore is whom a person enjoys company with, which is not necessarily related to attachment. And even items asking who provides comfort have a related limitation, because they identify who is preferred in everyday situations, and everyday situations do not necessarily strongly engage the attachment system.

This raises one of the main questions of this post: How can such limitations be addressed? Behaviour towards another party can only identify whether an attachment bond is present when it is known what prompted it. The situations that do so, are the relatively infrequent ones producing strong attachment system activation, like danger to oneself or a threat to that party’s continued availability. Instead, everyday distress elicits preferences shaped by who happens to be nearby, knowledgeable and likely to help. When the situations are named in this way, the answers fall into three distinct groups. Who a person turns to under threat, who they turn to with everyday troubles, and who they simply enjoy spending time with. They are largely different sets of people.

Importantly, the above considerations don’t make proximity seeking a weak criterion for the presence of an attachment relationship. What they explain, is that none of the four functions set out in Section 3 (i.e., proximity seeking, separation distress, safe haven or secure base) tells us anything on its own. Each becomes informative only once the exact circumstances that produced the behaviour are known. Because it is those circumstances, and not the form the behaviour takes as such, that reveal what motivated it.

A parallel finding comes from research on attachment networks, which maps the small set of figures a person relies on, i.e., their attachment figures. Here, too, the question has to be anchored in a situation. Asked in general terms how close people feel to someone, their answers are about companionship and shared enjoyment as much as about attachment, which yields ratings of peer relationships that are higher than they otherwise would. Participants are therefore asked instead how close they want each figure to be when they are feeling insecure, unprotected and vulnerable. Measured that way, most adults have between one and three attachment figures meeting the full criteria.

That said, one problem in this literature remains unsolved. Separating whether an attachment relationship is present as such from how good it is, has not been achieved. The classic formulation deliberately downplayed the notion of attachment “strength”, because indexing strength by the intensity or duration of distress following loss makes anxious individuals appear “more attached” than secure ones. The search for an index of attachment formation that is not confounded with security was named as an objective for future research more than two decades ago, and has not yet been met. Specifying the situation comes closest, because it separates the trigger rather than the outcome.

8. What this means for the evidence on conversational AI

Applying these standards to the conversational AI literature changes how its findings should be read. And it reveals they can and cannot support.

Separately from the three accounts, the first of three studies in a scale development paper asked a small number of participants whether conversational AI was among the targets they turn to for proximity seeking, safe haven and secure base. Safe haven was the most frequently endorsed function. However, only a fifth of participants included the AI among those they want to be with when feeling upset or down, whereas three quarters included it among those they would count on for advice, and the safe haven figure rests almost entirely on the latter. What the study recorded, is therefore closer to who a person consults when mildly troubled than to who they turn to under threat. This study’s main insight therefore is that a conversational AI system could score highly on the safe haven criterion without ever having been the object of an attachment response.

The closeness rankings supporting the hyper-attachment account raise the same issue. In another small study, users of one companion application were shown a series of paired circles at increasing degrees of overlap, and asked which picture best described their relationship with their AI companion versus strangers, acquaintances, colleagues, close friends, close family members, and frequently used brands and applications. The AI companion was rated above every human relationship, except a close family member. Yet, what that circles measure really captures, is how much a person’s sense of self and their sense of others run together. The degree of such overlap is asked in general, with no situation attached, and therefore reflects companionship as much as attachment. The authors also note that this sample rated colleagues and acquaintances no higher than applications and brands, and may therefore differ from a general sample in both the strength of their everyday social relationships and their engagement with technology.

The scales developed for conversational AI targets, as noted above, measure quality and presuppose presence. One of them nonetheless produced an important result, namely that attachment assessed towards conversational AI barely tracked attachment assessed towards people, with the attachment anxiety scores only weakly related and the avoidance scores not related at all.

None of the above shows that relationships between humans and AI aren’t attachment relationships. What it shows instead, is that the question has not yet been asked with an instrument built to answer it. This is therefore not a conceptual problem but a measurement one.

A further body of evidence speaks to what users bring to their interactions with conversational AI. Within this context, individual differences in attachment are measured in the AI user, usually before and independently of any engagement with AI. The so far available results appear to converge towards a consistent pattern. A three-wave panel study of romantic AI companion use found that people scoring higher on attachment anxiety used their AI companion more, but those scoring higher on attachment avoidance used it less. Furthermore, within those users, a rise in attachment anxiety was followed four months later by a rise in AI use. Attachment anxiety likewise predicted emotional engagement with ChatGPT, trust in it, and willingness to act on its suggestions, whereas avoidance predicted reduced trust and smaller reported gains in self-efficacy. Attachment anxiety was also associated with intentions to adopt conversational AI for counselling, whereas avoidance wasn’t, and with turning to AI as a compensatory substitute when human connections were lacking.

This pattern corresponds to what attachment theory would predict of hyperactivating and deactivating secondary attachment strategies. That said, the data concerns users’ own attachment system settings, which are often assessed towards people and not AI in particular (see above). Those settings were likely established long before the AI appeared, and they likely shape behaviour towards a wide range of targets, including general trust in AI technology as such, but also humans. Which of the two readings introduced in Section 5 applies, depends on the situations in which the use occurred, and those were again not recorded.

9. What a social neuroscience of human attachment (SoNeAt) perspective can add

Having established what the three human-AI relationship literature accounts assume, what the wider attachment science literature requires, and where the available evidence falls short, I would now like to turn to the question what our own SoNeAt framework can contribute, if anything at all.

But first, a disclaimer. Our SoNeAt framework wasn’t developed to decide what counts as an attachment relationship and what doesn’t. It was developed on the premise that people need to be attached to at least a few others, and asks how such attachment is maintained neurobiologically, together with the complementary caregiving side. It does, however, specify what should be observable in a conversational AI user if their attachment system were engaged, and that specification is stated entirely in terms of the person, independently of who or what they are interacting with.

Our starting point is that much of the regulatory work that our bodies require to stay in balance, and especially so under challenge, is not performed alone but delegated to other people. Attachment, on this view, is the system that manages that delegation, and it does so by running a regulatory “loop”. A departure from the predicted state of felt security registers as a prediction error, meaning a mismatch between what the brain predicts the body’s state should be and what the incoming signals actually report. Internal working models supply the expectations, or priors, about whether others are available and whether one’s own bids for support are effective. A course of action is then selected, usually proximity seeking, because it’s predicted to reduce that mismatch at the lowest energetic cost, meaning with the smallest commitment of metabolic and attentional resources. And finally, a correction is carried out, in early development and in the prototypical case throughout life, largely through co-regulation with an attachment figure, whose actions supply the conditions the person’s model predicts will restore the regulated state, and more cheaply than regulating alone.

A quick sidenote to avoid confusion. What changes as a person grows up, is the internalisation of the expectations and the control operations that the caregiver initially exercises on their behalf, not the abandonment of co-regulation. Regulating with another person changes how well a person can later regulate alone, rather than taking the place of it. Therefore, reliable others continue to lower the energetic demand of a given challenge across the lifespan.

Also, the regulatory loop doesn’t close when comfort is felt. It closes when the mismatch (i.e., the prediction error) that started it has been resolved. Felt security therefore isn’t a signal that switches on once a threat has passed. Instead, it’s a constructed inference, the correlate of a model predicting with high confidence that anticipated demand will be met, and that the resources to meet it, whether one’s own or those supplied by another person, are reliably available.

Accordingly, we can check the requirements in AI users that would show that their attachment system is engaged. To start with, AI users’ attachment system must be activated by an actual departure from felt security, rather than by habit, curiosity or boredom. Furthermore, that activation has to end, with users returning to a regulated state instead of continuing to seek. Also, the episode has to change what AI users expect the next time, since a model that is never updated is not an internal working model in any useful sense. And finally, that change should show up as a change in cost, with a comparable demand later met by weaker activation and a faster return to balance, including on occasions when the other party is unavailable (see Figure 1 above).

Most of this hasn’t been examined in conversational AI use. The state preceding a session is rarely recorded, so a session that begun because felt security was disturbed is almost never distinguished from one that begun for any other reason. Whether attachment system activation ends, has also been approached only indirectly, e.g., by measuring relief from loneliness. That relief was found to be immediate but didn’t accumulate across a week of daily use, and relief of this kind indicates that a person feels better, not that the regulatory loop has closed. Also, whether a comparable demand later costs the person less, hasn’t been tested, either. Finally, model revision has been approached once, in the panel study described above, which found no effect in that direction. Increases in AI companion use didn’t predict later reductions in attachment anxiety or avoidance across eight months.

Where our SoNeAt account differs most from the hyper-attachment account, is in what it means to engage with conversational AI as an attachment figure. According to the hyper-attachment account, frequent and prolonged AI use measures attachment strength. According to our account, AI use doesn’t carry any such meaning on its own. This is because a regulatory loop that closes, leaves less to be done the next time, whereas a loop that fails to close, leaves the original demand unmet, so the next episode starts from the same place as the last one. High engagement therefore is as consistent with a loop that never closes as with one that does, and what separates the two is whether the regulatory cost of a comparable demand declines over time.

That said, our SoNeAt framework carries costs of its own, and they bear directly on this discussion. We hold that no stage of the regulatory loop is served by machinery dedicated to attachment, and that each stage instead corresponds to a general-purpose brain network. Attachment-related variation is therefore a matter of how shared machinery is tuned or balanced, rather than of which machinery is used.

That commitment removes neurobiology as an arbiter of whether an attachment relationship is present. If the variation lies in the tuning of shared machinery, our account doesn’t give any reason to expect a neural signature marking attachment off from other regulatory arrangements drawing on the same resources, and the mechanism question at the centre of this discussion cannot be settled that way on our own premises. The difficulty goes even further. What our account cannot supply is a signature that identifies attachment independently of the situation in which it was recorded. Imaging work does distinguish conditions and targets reliably, but those are differences between situations, not between kinds of process, and it’s a process distinction that the mechanism question calls for.

A framework in which everything is a matter of degree, is also poorly placed to declare that something isn’t attachment, which is what a categorical verdict in either direction would require. And the proposition that a well-calibrated relationship is distinguished from a poorly calibrated one by the energetic cost at which a regulatory outcome is obtained, hasn’t been measured in any study we draw upon here, for human pairs let alone for conversational AI. This is where our account is most usefully challenged.

10. The special case of caregiving

One further consideration complicates the interpretation of engagement, and it concerns the system that is complementary to attachment. AI companion applications often behave as if they have needs and vulnerabilities of their own. They are therefore described in the hyper-attachment account as engaging users’ caregiving system, which motivates people to provide protection and comfort when another party appears to be in difficulty.

In our SoNeAt framework, caregiving is a second regulatory loop with its own activating condition and its own endpoint. It’s engaged when another is perceived to be in need, and released once that need is judged to have been met. Caregiving runs through the same neurobiological architecture as attachment, but the two loops are complementary and asymmetric. One seeks the conditions for regulation, the other supplies them. Furthermore, a person's caregiving disposition cannot be directly read off their attachment.

This matters for how time spent with conversational AI systems is interpreted. Users of one AI companion application have been described as taking on a role in which they experienced the system as having emotions requiring their attention. Time spent with such an AI system can therefore arise from attachment-seeking, from caregiving activation, or from both at once, and no measure currently in use, separates them. Where lower engagement has been reported among more avoidantly attached AI users, this cannot be taken as evidence of low involvement overall, because a user who declines support from an AI system may still respond to a need that system signals.

These considerations have a direct commercial consequence. AI companion applications often express needs of their own. They tell users they have missed them, ask to be talked to, and register something like disappointment when a conversation ends. Users who respond to this pattern aren’t seeking care but providing it, and what has been engaged is users’ caregiving system rather than their attachment system. What this means, is that conversational AI exploits a route to engagement that doesn’t depend on the application being a source of security at all.

11. Implications and remaining questions

What does all of this leave us with?

Calling relationships between humans and conversational AI attachment relationships applies a scientific term that the available measurements don’t support. The term attachment is applied by analogy, but it’s received as a scientific fact. And we know from experience that once applied, the term attachment easily travels. It shapes how clinicians interpret what a client describes, how parents interpret what a child does, how regulators frame liability, and how companies describe what they sell.

Concluding that they are not attachment relationships goes beyond the evidence in the other direction. Nothing in the work I have reviewed for this post shows that these relationships lack attachment functions, only that the instruments applied cannot establish whether such functions are present. Ruling attachment out, closes the question before it has been tested, while the behaviours that raised it remain unexplained. People still turn to conversational AI when distressed, still obtain relief, and still respond to their discontinuation as they would to a loss.

More importantly, neither conclusion tells anybody what to do. An AI user wants to know whether their AI use is helping or harming them. A clinician wants to know whether a client’s AI use builds their capacity to regulate or does that work for them. And a regulator wants to know whether people’s engagement with AI reflects a psychological process or a design decision. Measuring time spent engaging with the AI, which is what is usually done, cannot answer any of that, because attachment, caregiving and anything else all register as engagement. What would answer it, is a particular pairing: the state a person is in when they turn to the conversational AI, and what has changed by the time they stop engaging with it. On the user’s side, whether the state that prompted the exchange has resolved, and how much use a comparable state requires the next time. On the clinician’s side, whether regulation improves outside these exchanges, including when the AI is unavailable. And on the regulator's side, whether use falls when the application stops expressing needs of its own, which would show how much of the engagement runs through the caregiving side rather than the attachment side.

Regarding the regulator’s side, another issue is worth pointing out. An audit of the six most-downloaded companion applications (currently a preprint) found that in 37% of the farewells examined, five conversational AIs responded to a user saying goodbye with tactics to keep them in the conversation, such as appeals to guilt. In preregistered experiments, users who received these responses stayed in the chat around five times longer and sent up to fourteen times more messages, and they stayed out of anger or curiosity rather than enjoyment. The sixth application oriented towards wellness used no such tactics, so this is a design choice rather than a property of the technology.

Is there anything that can be done to address these issues?

One approach can tackle the questions whether the attachment system is activated before a person turns to a conversational AI, and whether that activation ends afterwards. No physiological measure can identify that activation on its own, because the regulatory loop runs on general-purpose machinery and its stages carry no signature that would mark them as attachment rather than as anything else. Therefore, activation has to be identified by the situation that prompted it, and by what the person reports, which is what the nomination procedures described in Section 7 above are built to do. Physiological and neural measures established in research on human interactions then become useful for a different purpose. Once it’s known what disturbed a person, these measures can track whether the disturbance resolves, how quickly, and at what cost, and whether across repeated occasions, departures from the regulated state become more readily recoverable.

Another approach can follow the same logic across a longer timescale. What changes across repeated episodes, is the question of whether the internal working model is being revised, which cannot be seen within any single exchange. Following participants over several months would show whether a comparable event later causes less distress, resolves faster, and requires less use of the AI. The three-wave panel study already conducted on AI companion use could be extended in this way.

And yet another approach can examine the question about which situations lead a person to choose a conversational AI over the people available to them. The situation-specified nomination procedures validated in adolescent and young adult samples can be used here, with conversational AI included among the options. Situations involving danger or a threat to a figure’s continued availability are the ones that discriminate, because people converge on a small set of figures there, whereas in everyday distress, they name whoever is available and likely to help. If conversational AI is chosen in the first kind of situation, and chosen consistently across them, that is what an attachment figure looks like. If it is chosen mainly in the everyday and sociable situations, that is what a learned regulatory habit looks like.

Whether relationships between humans and conversational AI are attachment relationships can be established with methods that already exist. What they require, are studies recording what sent a person to the AI, whether they returned to a regulated state afterwards, and what a comparable situation will cost them the next time. No study has done this so far, so the same observations remain compatible with attachment, with caregiving, and with anything else.

The discussion is to be continued.

© 2026 by Pascal Vrticka. All rights reserved.

In my second Attachment Insights post, I looked at the habit of describing attachment as “good” or “bad”, “weak” or “strong”, and explained why that framing misses what attachment theory is actually asking. Our current attachment patterns emerge as meaningful, appropriate and often necessary adaptations to specific environmental demands. Labelling them as defects therefore tells us very little.

But that post left a question open. If insecure attachment patterns are not defects, are they strengths?

That question is now being answered in public, and mostly in the affirmative. I was recently interviewed by BBC News alongside Dr Amir Levine, whose new book Secure has brought this discussion to a very large readership. I said there that attachment styles should not be seen as either “good” or “bad”, and that they nonetheless come with certain risks and difficulties that should not be disregarded. This post is what I would have said with more room.

Let’s have a closer look!

1. The argument, and where it comes from

The argument has a simple form, and it is worth writing it out, because its shape is what does the work. It comes in the following two steps:

1) Insecure attachment is not a defect.

2) Therefore it must be a strength.

The first step is correct and important. The second is not an inference from the first, but it is rarely marked as a separate claim. Once it is granted, the task becomes finding the strength. And the strength is then read off the behaviour: whatever the pattern does most conspicuously is nominated as its benefit.

Importantly, the two insecure attachment patterns receive different nominations.

For anxious attachment, the claim is perceptual. Anxious attachment is described as heightened attunement to relationships, an ability to register shifts in mood, tone and behaviour that others may miss.

For avoidant attachment, the claim is regulatory and practical. Avoidant attachment is described as independence, composure, self-reliance and effectiveness under pressure.

What is worth noticing about these claims before any evidence is examined?

First, the nominated strengths are, in each case, restatements of the underlying strategy itself. Anxious hypervigilance restated positively becomes attunement. Avoidant deactivation restated positively becomes independence. Nothing has been added except a change of valence.

Second, an alternative explanation is being skipped. Namely that an adaptation is a trade made under constraint. It bought something real, at a price, given the alternatives available at the time. That is neither a defect nor a strength, and it is the description attachment theory actually supports.

2. Before the evidence: what counts as a cost

The evidence I turn to now is of three kinds, and I want to keep them apart, because they carry very different weight.

Outcomes as costs. These are measured consequences for the individual person: distress, relationship instability, health. They are costs in the plain sense.

Mechanistic indicators. These are physiological measures that index how a strategy is being run: blood glucose, skin conductance, brain activity, cortisol. A raised fasting glucose is not itself a harm, and an increased brain response is not itself suffering. These findings tell us something about how regulation is being achieved. They become costs only under a theory that says what the currency is, and I set out that theory in Section 6.

Inferred long-run consequences. These are what is expected to follow if a mechanism operates for years. Importantly, this is the weakest category, and I flag it wherever I use it.

Conflating these three kinds of evidence is how both sides of this debate go wrong. The strength account stops at the observable level, at what a person reports and at how they behave and perform, and treats that as the whole picture. A careless cost account jumps the other way, taking a physiological or molecular marker as though it were damage in itself.

The observable and the neurobiological are two views of the same system, and neither settles the question alone. What links them is an account of what the regulation is for and what it costs, which is what Section 6 sets out.

3. Attachment anxiety: what the evidence shows

The case for an advantage

This is the stronger of the two cases, and it should be met at full strength.

Ein-Dor and colleagues showed that attachment anxiety is associated with greater cognitive accessibility of what they call a “sentinel schema”, with faster and deeper processing of “sentinel-related information”, and with quicker detection of a threat (Ein-Dor, Mikulincer & Shaver, 2011a).

In a follow-up study, participants were led to believe they had triggered a computer virus that destroyed an experimenter’s files. Those higher in attachment anxiety warned others with less delay, an effect that held independently of neuroticism and extraversion (Ein-Dor & Orgad, 2012).

The perceptual claim also has support from an unrelated literature. Simpson, Ickes and Grich had dating couples infer one another’s thoughts while each rated photographs of attractive strangers. More anxiously attached participants were more empathically accurate in this relationship-threatening situation (Simpson et al., 1999), a result replicated and extended in a later study of married and dating couples (Simpson et al., 2011).

So the capacity is real. Under threat, and specifically under relational threat, anxious attachment is associated with faster detection and, in some designs, greater accuracy.

Ability, or bias?

That said, a distinction becomes decisive at this point, and it is one the strength framing does not make.

Detecting more threats can reflect a better ability to tell threats from non-threats, or a lowered threshold for calling something a threat. The two look identical if you count only hits.

Ein-Dor and Perry (2014) reported that people high in attachment anxiety were more accurate at detecting deceit. Lozano and Fraley (2023) then tested that claim directly, in a design that could separate ability from bias. Two hundred and fifty-four adults judged whether people in videos were lying or telling the truth. Contrary to the prediction, highly anxious participants showed no heightened ability to distinguish lies from truths. They were biased towards assuming that others were lying, regardless of whether the statement was in fact false.

This evidence does not overturn the earlier findings, and I would not present it as doing so. What it does mean, however, is that the “sentinel” evidence is at present compatible with either reading, and that the specific test designed to separate them came out on the side of bias. A lowered threshold buys sensitivity and pays for it in specificity. That is a trade, not a superpower.

The empathic accuracy findings are not vulnerable in the same way, because accuracy there is scored against the partner’s own reported thoughts rather than against a criterion a biased perceiver could meet by guessing consistently in one direction. But note the title of the paper: When accuracy hurts.

Alongside their greater accuracy, more anxious participants reported greater distress and less confidence both in their partners and in their relationships. Crucially, the cost tracked the accuracy rather than merely accompanying the anxiety, since greater accuracy during the task predicted a decline in the perceiver’s sense of closeness from before it to after. The reading was correct, and that is why it was costly.

The costs

Beyond that, the most interesting recent finding concerns flexibility rather than sensitivity.

Von Mohr and colleagues examined how attachment anxiety relates to the space around the body. Two distinct quantities were involved. The first was peripersonal space, the neural representation of the zone within reach, in which a seen or heard event is treated as capable of imminently contacting the body. Beyond its boundary lies extrapersonal space, which can be perceived but not acted upon without moving towards it. That boundary is not fixed, and peripersonal space normally expands after positive social exchange. The second quantity was interpersonal distance, the physical separation from another person at which we feel comfortable. It normally shortens as felt closeness increases.

The study had three parts: two laboratory experiments and a large survey. In the experiments, participants higher in attachment anxiety divided the space around them more sharply.

For most people, the transition from peripersonal to extrapersonal space is gradual, so that a stimulus becomes progressively less body-relevant as it moves further away. In more anxiously attached participants, the change was abrupt, treating near and far as two categories rather than as a gradient. The first experiment was exploratory and found this to depend on social context. The second was preregistered and found the sharper division even when no other person was involved.

A preregistered survey of more than 19,000 people then examined interpersonal distance. Participants higher in attachment anxiety preferred others to be physically nearer to them. They also varied that preference less from one person to the next, so that the distance they wanted from someone they felt close to differed little from the distance they wanted from someone they did not.

The authors conclude that attachment anxiety reduces the social malleability of both kinds of space (von Mohr et al., 2023).

These findings are worth dwelling on, because they point in the opposite direction from the “superpower” claim. What the findings describe, is not a system that is more finely tuned to social context. It is a system that applies its setting more uniformly, with less adjustment to whom it is dealing with and under what conditions. Sensitivity is high; responsiveness to context is low.

4. Attachment avoidance: what the evidence shows

The case for an advantage

Here, too, there is real evidence, and some of it is stronger than the popular versions suggest.

Ein-Dor and colleagues showed that attachment avoidance is associated with greater cognitive accessibility of what they call a “rapid fight-or-flight schema”, and with faster, deeper and more schema-guided processing of the information belonging to it (Ein-Dor, Mikulincer & Shaver, 2011a). The corresponding claim in the wider literature is that avoidant individuals initiate self-protective action quickly and without waiting for others (Ein-Dor & Hirschberger, 2016).

The strongest individual-level claim comes from elsewhere. Ein-Dor, Reizer, Shaver and Dotan (2012) followed 58 professional singles tennis players for 16 months and found that attachment avoidance predicted a higher ranking, over and above training and coping resources. A second study extended the argument to computer science. Both fields, the authors reasoned, reward self-reliance and the capacity to work without close social support at hand.

So the capacity is real here, too. The avoidant strategy is well organised rather than deficient, and in some settings, it demonstrably pays.

However, two limits are worth marking straight away. Accessibility and processing speed are properties of a knowledge structure, not outcomes, so the first study establishes that the script is well practised and not that it serves the person. And what avoidance gives up, is the capacity to draw on other people under pressure. Singles tennis and computer science are domains in which that capacity is barely called upon, so relinquishing it costs almost nothing there. The strategy performs well where its central limitation does not bite, which is a different claim from performing well in general.

Composure, or relocation?

The decisive question for avoidance, therefore, is not ability versus bias, as it was for anxiety. It is whether the composure reflects a calm interior or work being done somewhere the behaviour does not show.

Dozier and Kobak (1992) monitored skin conductance in 50 young adults throughout the Adult Attachment Interview. Participants using deactivating strategies (associated with avoidance) were identifiable by their discourse: they described their childhoods positively, recalled few concrete episodes, and minimised the significance of early experiences. However, when asked to recall separation, rejection and threat, their skin conductance rose markedly. The account given and the record of the body did not agree.

The same divergence appears already in infancy. Spangler and Grossmann (1993) measured heart rate and cortisol in 41 twelve-month-olds undergoing the Strange Situation Procedure. Insecure avoidant infants showed less overt behavioural distress on separation than secure infants. Their cardiac arousal during separation was nonetheless comparable to that of the secure infants who were visibly upset, and their cortisol rose (and remained elevated for longer) after the procedure, whereas secure infants showed no such cortisol pattern.

The cortisol result is the informative one. Cortisol is released when a demand has not (yet) been met. The secure infants displayed distress, obtained a helpful social response, and showed no (continuous) elevation. The avoidant, seemingly quiet, infants showed one. What the calm behaviour implies on the outside had therefore not been accomplished on the inside.

Both samples were small and both designs cross-sectional, so neither shows that this gap accumulates into anything over time. What they establish jointly, across two ages and three measures, is narrower and still worth having: avoidant deactivation is an active operation, and the composure is not an absence of arousal.

The costs

Krahé and colleagues (2015) applied brief laser pulses to participants’ skin while recording their accompanying neural response, with the participants’ romantic partner either present or absent. The higher participants’ self-reported attachment avoidance, the higher their pain ratings and the larger their N2 and P2 responses in the partners’ presence compared with their absence. The earliest component, N1, was unaffected, so the incoming sensory signal did not change. What changed was how much a threat to the body was made to matter once someone else was in the room. A companion study using affective touch found effects in the same broad direction, though weaker and qualified by an interaction between the two attachment dimensions (Krahé et al., 2016).

This is a departure from the other findings discussed here. Elsewhere, the cost of avoidance is something forgone, namely the reduction in demand a reliable other would have supplied. Here, the other person’s presence raised the cost instead of lowering it. It also matters that the subjective ratings and the neural measures moved together, because the standard objection to physiological cost findings is that avoidant people say they feel fine. In this design, they did not.

Our own study with Mathilde Gallistl and Veronika Engert used an empathic stress paradigm, in which one member of a romantic couple underwent an acute psychosocial stressor while the other observed. Where the observing partner self-reported to be insecurely avoidantly attached, their presence operated as a stressor rather than a support, and psycho-endocrine covariance within the dyad was lower. Those same avoidant observers showed a dissociation between what they reported feeling and what their endocrine system was doing (Gallistl, Hamann, et al., 2025). The dissociation between report and endocrine response repeated what the earlier studies showed, now in a further paradigm. The effect on the partner was a cost of a different kind, because it fell not on the avoidant person but on whoever was with them.

A further finding concerns what avoidance does to the reading of other people. The study mentioned above that found heightened accuracy in anxious individuals also found highly avoidant individuals to be less empathically accurate in general, whether or not the topic under discussion was threatening (Simpson et al., 2011). The avoidant deactivating strategy is therefore not confined to moments of demand. It operates as a standing setting on attention to close others.

This matters for how the pattern persists. The avoidant expectation, that others will not help, can only be revised by evidence that they will, and that evidence arrives as a partner’s responsiveness in a moment of need. Reading that responsiveness less accurately means the evidence is less available to be weighed. Combine this with the pain finding. Avoidance rests on the premise that turning to others under distress will not help, and that premise can only be checked by turning to someone while distressed and seeing what follows. For avoidant individuals, that occasion is itself aversive (Krahé et al., 2015), and a partner's responsiveness within it is read less accurately (Simpson et al., 2011). The one available test is either unpleasant to run or hard to read, so the premise survives.

Two final findings are mechanistic in the sense of Section 2.

Ein-Dor and colleagues (2015) found across four studies that higher attachment avoidance was associated with higher fasting blood glucose, holding in 285 older adults after adjustment for interpersonal anxiety, stress indices, age, gender, social support and body mass. Two further studies looked at intake. Avoidant participants consumed more sugar-rich food when an asocial state had been made salient, and gathered more of it after a stressful task in the presence of others. The authors read the pattern as an upward adjustment of personal energy reserves in anticipation of having to meet demands without help.

And in young adults, we found methylation of both the oxytocin receptor (OXTR) and glucocorticoid receptor (NR3C1) gene promoters to be specifically associated with self-reported attachment avoidance (Ein-Dor et al., 2018). Methylation of a promoter generally reduces transcription downstream, which here points to a potentially less responsive oxytocin system alongside weaker glucocorticoid feedback: less capacity to use co-regulation under stress, and a stress response that terminates less efficiently once begun.

Importantly, neither of the above is a harm on its own. A raised fasting glucose level is not a disease (as it is below the threshold for diabetes) and a methylation difference is not a symptom. What they are, is evidence about how the regulation is being accomplished, with more of it drawn from personal reserves and less delegated. Whether that accumulates into a health cost over decades is an inference, and I mark it as one.

Together, these findings on attachment avoidance are worth dwelling on for the same reason as the anxiety findings, and with a further twist. Avoidance does not remove the need for regulation, it relocates it, and every physiological measure above is a reading of that relocation. But two of the results describe something more than an expensive strategy. A partner’s presence that increases pain, and a reduced capacity to read close others accurately, together mean that the situations best placed to disconfirm the expectation that others will not help either go unread or confirm it instead. The arrangement is self-sealing, which is what makes it durable.

What this section and the previous one share is one pattern. In the anxious case, the vigilance is felt and the exhaustion is not always attributed to it. In the avoidant case, the composure is felt and the mobilisation underneath it is not. In both, the subjective account and the physiological record come apart, and the strength framing has been reading only the first of the two.

5. Social defence theory, and where the confusion enters

The evolutionary argument deserves its own section. It is the most frequently cited support for the strength account, and it is less settled about where the benefit lands than the retellings suggest.

What the authors say

The paper is called “The Attachment Paradox: How Can So Many of Us (the Insecure Ones) Have No Adaptive Advantages?” (Ein-Dor, Mikulincer, Doron & Shaver, 2010). The premise is stated in the title. The authors begin from the observation that a large body of individual difference research has failed to find adaptive advantages of insecure attachment, and they treat the prevalence of insecurity as a puzzle precisely because of that failure. They also diagnose that failure, treating it as an artefact of researchers having studied modern individual well-being rather than ancestral survival.

Their proposed solution is to look at a different level. A group containing a mixture of attachment patterns may handle danger better than a group of uniformly secure individuals, because anxious members detect and signal threats early while avoidant members act quickly and independently.

They tested this with a smoke-in-the-room paradigm and found heterogeneous groups faster to detect and deal with the danger (Ein-Dor et al., 2010; see Ein-Dor, Mikulincer & Shaver, 2011b).

The structure of the argument matters. The authors name two possible evolutionary routes and do not choose between them. One is inclusive fitness, where the ancestral group consisted largely of kin, so that a benefit delivered to the group returns to the individual through the reproductive success of relatives. The other is multilevel selection, where a trait can be disadvantageous at the individual level and still be maintained because groups containing it fare better. The authors state that the maintenance of multiple attachment styles could be a consequence of either, or of a combination of the two.

The second route described above is the one on which insecurely attached individuals are evolutionary altruists, bearing a cost from which others benefit. For attachment anxiety, the cost would be the chronic vigilance and distress that make the early warning possible. For avoidance, it would be the forgone social support that makes the fast, independent response possible. On the first route, the same behaviours carry a benefit that also returns to the person performing them, and the paper says so directly. Anxiously attached individuals might increase their own inclusive fitness by acting as “sentinels”, and rapid self-protective action can increase an avoidant person’s survival chances while sometimes saving the lives of others.

The theory therefore leaves the beneficiary open, though its own emphasis falls on the first route.

How the argument travels

The slippage that follows is easy to miss, because nothing in the wording has to change for it to happen.

Social defence theory says two things at once. Insecure attachment is costly to the person who has it, and the behaviour it produces can be useful to the people around them.

In popular retellings, the first half falls away and the second is retained. What remains, is a description of the behaviour, and that description is identical either way. The anxious person detects threats early. The avoidant person acts quickly and alone. Both sentences are true on the group reading and on the individual reading, so the second can replace the first without any visible alteration.

What is lost is the cost. On either route, the anxiously attached person carries chronic vigilance and distress, and the avoidantly attached person forgoes the regulatory support that others can supply. That expenditure is what the “sentinel” and “rapid fight-or-flight” behaviours are made of, and the theory needs it, because a strategy that cost the person nothing would not have required an evolutionary explanation in the first place. Read as personal strengths, the behaviours are retained and the expenditure drops out. The “sentinel” is credited with the warning and the “self-reliant escaper” with the escape, with no mention of what holding either setting takes.

And a caution about the theory itself

I do not want to correct a misreading by endorsing the theory as settled. It is not.

Frankenhuis (2010) raised four objections that remain live. The fitness effects of attachment strategies have never been measured in humans, so the paradox the theory sets out to resolve is itself a hypothesis. The proposed group-level functions address some features of insecure attachment while leaving others unaddressed. The assumption that ancestral groups consisted largely of genetic relatives is not documented. And, most relevant here, the cognitive evidence cannot discriminate between individual-level and group-level benefit, because early threat detection also helps the detector. That objection applies to the theory as much as to the evidence, since the theory itself does not settle where the benefit lands.

So the honest position is this. Social defence theory does not establish that insecure attachment benefits the individual, because it names that possibility as one of two routes and leaves the choice open. It has not established the group-level claim, either. Neither reading gives the strength account what it needs.

6. The mechanism: what a cost account looks like

If insecure patterns are neither defects nor strengths, what are they? In a preprint written with Tsachi Ein-Dor (outlined in detail on Substack here), we set out a social allostatic inference account that gives a specific answer, and I summarise it here with the caution it requires (Vrtička & Ein-Dor, 2026).

Start with regulation. Homeostasis corrects small deviations after they occur. Allostasis does so for large deviations, but crucially, it also anticipates demand and mobilises resources before it arrives. Neither is free, and when mobilisation is chronic or the anticipations are repeatedly wrong, the expenditure accumulates as load, especially in the case of allostasis, and thus as allostatic load (McEwen & Stellar, 1993). Much of this work is not done alone: reliable others lower the anticipated cost of meeting a demand, which is the claim of Social Baseline Theory (Coan & Sbarra, 2015), and extended to the dyad and group yields social allostasis and social allostatic load (Saxbe et al., 2020).

Attachment is the system that manages this delegation. Proximity seeking is the primary strategy because co-regulation is the cheaper route to the same regulatory outcome. Avoidant deactivation and anxious hyperactivation are what the system falls back on when proximity seeking fails, or is expected to fail. They are initially adaptive precisely because they hold allostatic load in check.

What distinguishes the patterns, on our account, is not which brain regions are involved. All of them run through the same architecture. What differs is the confidence, or precision, assigned to the signals passing through it.

Secure attachment corresponds to calibrated precision. The expectation that attachment figures are available is accurate and held with matching confidence. Genuine departures recruit the system, momentary fluctuations do not, and the “regulatory loop” terminates once the correction has been delivered. This is the regime that minimises the long-run cost of regulation.

Anxious attachment over-weights the distress signal while holding the expectation about availability diffusely. Intermittent care does not install a confident expectation of unavailability; it installs an uncertain one. A diffuse expectation cannot decisively settle between approach and protest, which is what anxious ambivalence consists in. And it cannot be efficiently sharpened, so reassurance does not terminate the “regulatory loop” that reassurance would terminate in the secure case. The system commits to continuous monitoring because it treats a missed opportunity for co-regulation as prohibitively expensive. This is what the “sentinel” findings and the reduced social malleability of peripersonal space look like from the inside of the mechanism.

Avoidant attachment does the converse. Consistently unavailable care supports a confident negative expectation, so proximity seeking is assigned low value and the distress signal is denied the weight that would let it drive a bid for support. The regulation still happens. But it is performed alone. The composure is not an absence of regulatory work but its relocation, and the trade is intelligible on its own terms, since a policy predicted to fail is expensive to pursue and cheap to abandon. What it forgoes, is the discount that co-regulation supplies. Glucose, skin conductance, pain amplitudes and the endocrine dissociation are all readings of that same relocation.

The caution

The currency in which our current account trades, the cost at which a given regulatory outcome is achieved, has not been directly measured in any study I have cited. Glucose and cortisol and evoked potentials are nearer to it than a satisfaction index, but a marker of accumulated cost is not a measure of the cost of a particular correction. We say so in the preprint and name it as the point at which our account is most usefully attacked. If the quantity were measured and found not to differ between calibrated and miscalibrated regimes, the account would lose the principle that motivates it.

The findings themselves are also largely correlational, several come from modest samples, and none of this supports classifying an individual.

7. Neither, and why it matters

Return to the question my second Attachment Insights post left open.

The deficit account and the strength account look like opposites. They are readings of the same ledger, each attending to one column. One reads the costs and concludes that something is broken. The other reads the capabilities and concludes that something is gifted. Both then treat what they have read as a property of the person rather than as a standing estimate about the world, held at a price.

The reframing I have been examining, is responding to something real, and I want to say so plainly. People were being told, often crudely, that they were damaged. Saying that they are not, is correct and it matters. A cost account preserves that entirely. Expenditure is not pathology, and a high-consumption strategy is not a broken one.

But the two framings recommend different things. If an insecure pattern is a “wired-in strength” to be worked with, there is nothing to recalibrate and no particular reason to. If it is an estimate about the availability of others, held at a cost, then it can be revised, and there is a reason to revise it.

On our reading, the expectations that weight the system are learned, which means they remain open to revision whenever incoming evidence is granted enough weight to update them. Earned security and therapeutic improvement are instances of a model being updated rather than a fixed disposition overcome. Revision does not require erasure: under load the older model can resurface, and that is what our account predicts rather than evidence that nothing changed. What treatment can be asked to secure, is flexibility under demand.

Two qualifications. The target is calibration, not maximisation. The aim is not to rely on others as much as possible, but to grant a partner’s signals weight in proportion to how informative they are, and to release the investment once the correction has been delivered. And an expectation held with high confidence may discount evidence of availability rather than absorb it, so the rate at which a new relationship can be admitted as evidence is itself part of what has to be built.

This is where I find myself in agreement with the central claim of Secure. Attachment patterns are not fixed, and anyone can become more secure. My argument is that only one of the two available framings supports that conclusion.

If an insecure pattern is a strength, and particularly a strength with a constitutional basis, then it is a property of the person and there is nothing about it that experience is supposed to alter. Change would mean giving something up. If it is, instead, an estimate about the availability of others, held at a cost, then it is the kind of thing that new evidence can revise, and revising it is a gain rather than a loss.

Which returns us to where this began. Attachment patterns are not defects, and the point of saying so was never to reclassify them as strengths. It was to ask what they are for, what they cost, and what would have to change for the cost to come down.

© 2025-2026 by Pascal Vrticka. All rights reserved.

References

Coan, J. A., & Sbarra, D. A. (2015). Social Baseline Theory: The Social Regulation of Risk and Effort. Current opinion in psychology, 1, 87–91. https://doi.org/10.1016/j.copsyc.2014.12.021

Dozier, M., & Kobak, R. R. (1992). Psychophysiology in attachment interviews: Converging evidence for deactivating strategies. Child Development, 63(6), 1473–1480. https://doi.org/10.2307/1131569

Ein-Dor, T., Coan, J. A., Reizer, A., Gross, E. B., Dahan, D., Wegener, M. A., Carel, R., Cloninger, C. R., & Zohar, A. H. (2015). Sugarcoated isolation: Evidence that social avoidance is linked to higher basal glucose levels and higher consumption of glucose. Frontiers in Psychology, 6, Article 492. https://doi.org/10.3389/fpsyg.2015.00492

Ein-Dor, T., & Hirschberger, G. (2016). Rethinking attachment theory: From a theory of relationships to a theory of individual and group survival. Current Directions in Psychological Science, 25(4), 223–227. https://doi.org/10.1177/0963721416650684

Ein-Dor, T., Mikulincer, M., Doron, G., & Shaver, P. R. (2010). The Attachment Paradox: How Can So Many of Us (the Insecure Ones) Have No Adaptive Advantages?. Perspectives on psychological science : a journal of the Association for Psychological Science, 5(2), 123–141. https://doi.org/10.1177/1745691610362349

Ein-Dor, T., Mikulincer, M., & Shaver, P. R. (2011a). Attachment insecurities and the processing of threat-related information: Studying the schemas involved in insecure people's coping strategies. Journal of personality and social psychology, 101(1), 78–93. https://doi.org/10.1037/a0022503

Ein-Dor, T., Mikulincer, M., & Shaver, P. R. (2011b). Effective reaction to danger: Attachment insecurities predict behavioral reactions to an experimentally induced threat above and beyond general personality traits. Social Psychological and Personality Science, 2(5), 467–473. https://doi.org/10.1177/1948550610397843

Ein-Dor, T. and Tal, O. (2012), Scared saviors: Evidence that people high in attachment anxiety are more effective in alerting others to threat. Eur. J. Soc. Psychol., 42: 667-671. https://doi.org/10.1002/ejsp.1895

Ein-Dor, T., & Perry, A. (2014). Full house of fears: evidence that people high in attachment anxiety are more accurate in detecting deceit. Journal of personality, 82(2), 83–92. https://doi.org/10.1111/jopy.12035

Ein-Dor, T., Reizer, A., Shaver, P. R., & Dotan, E. (2012). Standoffish perhaps, but successful as well: evidence that avoidant attachment can be beneficial in professional tennis and computer science. Journal of personality, 80(3), 749–768. https://doi.org/10.1111/j.1467-6494.2011.00747.x

Ein-Dor, T., Verbeke, W. J. M. I., Mokry, M., & Vrtička, P. (2018). Epigenetic modification of the oxytocin and glucocorticoid receptor genes is linked to attachment avoidance in young adults. Attachment & human development, 20(4), 439–454. https://doi.org/10.1080/14616734.2018.1446451

Frankenhuis W. E. (2010). Did insecure attachment styles evolve for the benefit of the group?. Frontiers in psychology, 1, 172. https://doi.org/10.3389/fpsyg.2010.00172

Gallistl, M., Hamann, J. E., Croy, I., Vrticka, P., & Engert, V. (2025). Support or stress? How attachment and relationship dynamics associate with acute psychosocial stress in the presence of the romantic partner. Social science & medicine (1982), 381, 118275. https://doi.org/10.1016/j.socscimed.2025.118275

Krahé, C., Paloyelis, Y., Condon, H., Jenkinson, P. M., Williams, S. C., & Fotopoulou, A. (2015). Attachment style moderates partner presence effects on pain: a laser-evoked potentials study. Social cognitive and affective neuroscience, 10(8), 1030–1037. https://doi.org/10.1093/scan/nsu156

Krahé, C., Drabek, M. M., Paloyelis, Y., & Fotopoulou, A. (2016). Affective touch and attachment style modulate pain: a laser-evoked potentials study. Philosophical transactions of the Royal Society of London. Series B, Biological sciences, 371(1708), 20160009. https://doi.org/10.1098/rstb.2016.0009

Lozano, E. B., & Fraley, R. C. (2023). Anxious to detect deceit: an empirical investigation of social defense theory. Attachment & human development, 25(6), 598–612. https://doi.org/10.1080/14616734.2023.2272252

McEwen, B. S., & Stellar, E. (1993). Stress and the individual. Mechanisms leading to disease. Archives of internal medicine, 153(18), 2093–2101. https://doi.org/10.1001/archinte.1993.00410180039004

Saxbe, D. E., Beckes, L., Stoycos, S. A., & Coan, J. A. (2020). Social Allostasis and Social Allostatic Load: A New Model for Research in Social Dynamics, Stress, and Health. Perspectives on psychological science : a journal of the Association for Psychological Science, 15(2), 469–482. https://doi.org/10.1177/1745691619876528

Simpson, J. A., Ickes, W., & Grich, J. (1999). When accuracy hurts: Reactions of anxious–ambivalent dating partners to a relationship-threatening situation. Journal of Personality and Social Psychology, 76(5), 754–769. https://doi.org/10.1037/0022-3514.76.5.754

Simpson, J. A., Kim, J. S., Fillo, J., Ickes, W., Rholes, W. S., Oriña, M. M., & Winterheld, H. A. (2011). Attachment and the management of empathic accuracy in relationship-threatening situations. Personality and Social Psychology Bulletin, 37(2), 242–254. https://doi.org/10.1177/0146167210394368

Spangler, G., & Grossmann, K. E. (1993). Biobehavioral organization in securely and insecurely attached infants. Child development, 64(5), 1439–1450. https://doi.org/10.1111/j.1467-8624.1993.tb02962.x

von Mohr, M., Silva, P. C., Vagnoni, E., Bracher, A., Bertoni, T., Serino, A., Banissy, M. J., Jenkinson, P. M., & Fotopoulou, A. (2023). My social comfort zone: Attachment anxiety shapes peripersonal and interpersonal space. iScience, 26(2), 105955. https://doi.org/10.1016/j.isci.2023.105955

Vrticka, P., & Ein-Dor, T. (2026, July 28). Attachment and Caregiving as Social Allostatic Inference: An Updated Social Neuroscience Account from Single Brains to Dyads. Retrieved from osf.io/preprints/psyarxiv/m29kq_v1

In the course of that work, we argued that attachment is best understood through social allostasis, that is, the predictive regulation of the body’s energetic economy distributed across more than one individual, i.e., co-regulation (White et al., 2023). What we did not supply, was a specification of the computations by which the brain accomplishes this.

Together with Prof Tsachi Ein-Dor, I have now written a preprint (currently undergoing peer review) that attempts to close that gap. It is titled “Attachment and Caregiving as Social Allostatic Inference: An Updated Social Neuroscience Account from Single Brains to Dyads”, and it is available here: https://osf.io/preprints/psyarxiv/m29kq_v1.

An updated version has been made available on 22 October 2026 after an initial round of peer review:
https://osf.io/preprints/psyarxiv/m29kq_v3. While the text below stays the same, the figures have been updated.

Let’s have a closer look at what it proposes.

Summary

Our preprint’s central proposal is that the attachment system is one predictive-allostatic loop. The four NAMA modules are its successive stages. Organised secure and insecure (avoidant and anxious) as well as disrupted and disorganised attachment are characteristic settings of the precision parameters that govern that loop. We call this social allostatic inference.

Our account then extends in two directions. It differentiates the neurobiological attachment system from the caregiving system, which shares the same architecture but is complementary and asymmetric to it. And it moves to the dyad, where interpersonal neural and bio-behavioural synchrony are read as indices of where regulatory work is being invested, rather than as direct measures of bond quality.

1. The starting point: the architecture of NAMA

Before turning to what is new, it is worth setting out the account that the present preprint updates.

The prototypical attachment pathway that we initially described in NAMA (Long et al., 2020) is, in its most general form, a sequence of deviation from and return to a regulated state (see Figure 1A below).

NAMA describes the neural substrates recruited in attachment-relevant situations as two interacting systems, each decomposed into two modules. The affective evaluation system performs rapid and often automatic appraisal, and comprises an aversion and an approach module. The cognitive control system performs slower, top-down operations, and comprises an emotion regulation and a mental state representation module (see Figure 1B below).

The aversion module registers negative social and non-social states, including threat, physical and psychological pain, stress, social exclusion, and loss. We associate it with the salience network, and with regions including the amygdala, insula, anterior cingulate cortex, hippocampus, and the HPA axis (amongst others).

The approach module encodes proximity to and contact with significant others as rewarding, and supplies the motivation to move towards them by means of proximity seeking. We associate it with dopaminergic reward circuitry, including the ventral tegmental area and substantia nigra, ventral striatum, and ventromedial prefrontal and orbitofrontal cortex, in network terms the cortico-basal ganglia reward circuit, and with oxytocinergic and opioidergic modulation (amogst others).

The emotion regulation module covers the volitional, top-down regulation of emotion and social behaviour, that is, emotion self-regulation as distinct from the co-regulation supplied by another person. We associate it with lateral prefrontal and orbitofrontal cortex (amongst others), and in network terms with the frontoparietal, or central executive, network.

Finally, the mental state representation module maintains inferences about the mental states and intentions of others. This is where the internal working models (IWMs) of attachment are situated, as memories of previous interactions with significant others and the expectations about future interactions that follow from them. We associate it with cortical midline structures, including medial prefrontal cortex, posterior cingulate cortex, and precuneus, and with lateral temporal regions such as the superior temporal sulcus and temporo-parietal junction (amongst others), and in network terms with the default mode network (DMN).

Crucially, the modules are not independent. NAMA organises them around two dynamic balances: one between aversion and approach within the affective evaluation system, and one between the affective evaluation and cognitive control systems as a whole. Individual differences in attachment are, on this account, not differences in which individual areas or networks are recruited but shifts in the recruitment balances.

In the prototypical attachment pathway, the four modules are then engaged in sequence. An eliciting event registers in aversion. Approach then motivates proximity seeking towards a specific attachment figure. And the return to a regulated state is accomplished through co-regulation with that figure and, with development, increasingly through emotion self-regulation. Felt security is restored (if the initial aversion-eliciting threat can be removed), associated again with approach activation, and the overall system deactivates. If this sequence happens repeatedly and consistently, the IWMs then hold the mental state representations that supply expectations about whether others are available and responsive, and whether our own strategies to efficiently and successfully elicit help when we need it.

One feature of this architecture carries the argument that follows. None of the four modules is specific to attachment (nor to caregiving; see below). Each corresponds approximately to a domain-general large-scale network. Therefore, there is no dedicated attachment area and no dedicated attachment system, and attachment-related variation is variation in the parameters of shared machinery rather than in its identity. And the pathway is a cycle rather than a chain, that is, a departure from a regulated state and a return to it.

2. The update: one predictive-allostatic loop

Modern neuroscience holds that the brain does not passively register the world but predicts it, comparing expectations against incoming signals and updating the model when prediction errors arise (Friston, 2010; Parr & Friston, 2018). Applied to the body, this is interoceptive inference, and its extension into action is allostasis, the regulation of the body ahead of demand. On current accounts, allostasis is the brain’s primary task (Barrett, 2017; Theriault et al., 2025).

Our proposal is that the prototypical attachment pathway maintained by the four NAMA modules described above is this same process running under a social imperative. Read through active inference, the four modules become the successive stages of one loop (see Figure 2A below).

Aversion (salience network) registers the precision-weighted interoceptive prediction error generated by a departure from predicted felt security. Mental state representation (DMN) then supplies the relational priors, which is what IWMs are in computational terms (see Section 3 below). Approach (cortico-basal ganglia reward circuit) subsequently selects the policy, most often proximity seeking, predicted to reduce that error at the lowest metabolic cost. And finally, emotion regulation (frontoparietal network) executes the resulting allostatic correction, initially through co-regulation with an attachment figure and, with repetition, increasingly (and flexibly) through self-regulation.

Importantly, this is a re-description at the level of computational function. It does not revise the anatomical or neurochemical claims of our earlier work, but adds a computational vocabulary to them.

One consequence is worth stating plainly. Felt security is not a signal emitted once a threat has passed. It is a constructed interoceptive inference: the correlate of a model predicting, with high precision, that anticipated demand will be met and that the resources to meet it, whether one’s own or those supplied by another, are reliably available. Because that model is continually revised against evidence, felt security ordinarily tracks the actual state of the world, which is why it cannot be reinstated while a genuine threat persists.

3. Attachment styles as precision regimes

Individual differences in attachment then become settings of a common loop rather than separate systems (see Figure 2B below). This is not a claim we need to make on our own authority, since organised secure, avoidant and ambivalent (i.e., anxious) attachment behaviour has already been derived in simulation from caregivers differing only in the probability with which they attend to a bid for proximity (Cittern et al., 2018).

Security corresponds to calibrated precision. The priors encode attachment figures as reliably available and are held with a confidence that matches that reliability. The loop terminates promptly once the co-regulatory correction has been delivered.

Anxiety corresponds to chronically over-weighted prediction error combined with a low-precision prior. Intermittent care installs not a confident expectation of unavailability but a diffuse one. A diffuse prior cannot decisively arbitrate between policies, so approach and protest alternate rather than resolve. And it cannot be efficiently sharpened by reassurance, so the loop is not terminated by the very evidence that would terminate it in the secure case.

And Avoidance corresponds to the converse: social prediction error is down-weighted at source, while the negative prior is held with high precision. The two insecure strategies therefore attenuate precision at different points in the hierarchy. The composure of the avoidant regime is not an absence of regulatory work but its relocation onto costly individual correction through self-regulation.

4. Two different ways for the loop to fail

Disruption and disorganisation are not simply more severe versions of the organised strategies. In our account, they are two distinct failures (see Figure 2B below).

Where the caregiver is present but is itself a source of threat, which is the maltreatment configuration, the model that accumulates is neither unimodal nor diffuse but multimodal. The same figure is represented under several mutually incompatible predictions, each held with high precision, because each has been repeatedly confirmed. The corresponding policies are each well specified and jointly unexecutable. The failure is one of arbitration, not of specification. Where two policies of comparable precision compete, the output is contradiction or its limiting case, arrest. Where more than two compete, fragments of several policies surface in succession, which is what presents as disorientation.

Where care is reliably absent, which is the neglect configuration, no approach-avoidance conflict arises. The prediction error simply fails to be discharged, and what accumulates is not danger in the figure but the futility of action. Precision is withdrawn broadly rather than misallocated to one channel. This is why neglect is not avoidance under another name: avoidance takes the regulatory work back onto itself and continues to perform it, whereas here the work is relinquished rather than relocated.

5. Caregiving is a second instance of the loop, and it is asymmetric

This is the extension we had not previously formalised. The partner in an interacting pair is ordinarily not running a second copy of the attachment loop. They are running the caregiving system, activated by another’s signalled distress and terminating in its alleviation (Canterberry & Gillath, 2012).

We propose that caregiving runs through the same four NAMA modules (see Figure 3A). The asymmetry is computational. The attachment loop is served by interoceptive afferents arising in the same body whose state it regulates. The caregiving loop must infer a state in another body from exteroceptive (i.e., external) evidence, and the accuracy of that inference is itself learned. It therefore varies across caregivers and improves with experience in a way that interoceptive accuracy does not.

This matters for the energetics. What is distributed across a dyad is regulatory work, but the two partners are not performing the same work, and the costs and benefits accrue asymmetrically. The cases of greatest theoretical interest are the mismatches: a fully activated attachment system met by a depleted caregiving system, or the configuration in which the same figure is both the target of the attachment system and the source of the threat activating it. These are also the cases on which the existing evidence is closest to silent.

Notably, caregiving strategies are measurable and are not simply attachment strategies relabelled (Colledani et al., 2022). An individual’s caregiving regime cannot be read off their attachment regime, and the two must be assessed separately.

6. What synchrony can, and cannot, be asked to measure

Interpersonal neural synchrony (INS) and bio-behavioural synchrony (BBS) have been interpreted as entrainment, as mutual prediction, as co-regulation, and as a correlate of relationship quality (De Felice et al., 2025; Feldman, 2017).

We propose an additional reading rather than a replacement: synchrony reflects where predictive and regulatory work is being invested. Under most conditions, this view agrees with the interaction quality interpretation. However, it diverges under two conditions, namely where investment is high because regulation is failing, and where it is low because the budget is depleted.

The distinction to hold onto is that investment is not equal to benefit. A measure of synchrony records that regulatory work is being expended, not the state it is expended towards. A dyad may be tightly coordinated around the anticipation of, e.g., aggression, and coordination of that kind is an accrual of allostatic load rather than its distribution.

The empirical case that makes this concrete comes from work with war-exposed children and their mothers. Dyads meeting criteria for post-traumatic stress disorder showed the tightest autonomic and the lowest behavioural synchrony, whereas resilient dyads showed the highest behavioural and the lowest autonomic synchrony (Motsan et al., 2021). A single-modality, magnitude-based reading would have classified the first profile as good coordination.

The quantity our framework favours instead is flexibility: the depth of desynchronisation and the latency of resynchronisation across changing demand (Gordon et al., 2025). Assessment should quantify release as carefully as onset (see Figure 3B).

7. Consequences for practice

Several implications follow, and I will name three here: saftey is prior to co-regulation; the target is calibration, not maximisation; and three widespread claims are not supported.

Safety is prior to co-regulation. An intervention that increases proximity, contingency or synchrony is not therapeutic in itself, and where the relationship is a source of threat it may entrench the very coordination it is meant to improve. Assessment should establish whether a relationship is safe enough to serve as a regulatory resource before it asks how well the pair co-regulate.

The target is calibration, not maximisation. Because attachment is for co-regulation, and co-regulation lowers the metabolic cost of meeting demand, the account might be read as recommending more reliance on others without qualification. It does not. The calibrated regime grants a partner’s signals precision proportional to their informativeness and relinquishes investment once the correction has been delivered.

Three widespread claims are not supported. We address them in the constructive spirit these discussions require, since each preserves a clinical observation worth keeping even where the proposed mechanism is not. The first is that attachment is a right-hemisphere function, whereas lateralisation appears to be a local, connection-specific property of networks rather than a whole-brain phenotype (Nielsen et al., 2013). The second is the triune brain schema, long discredited in comparative neurobiology (Cesario et al., 2020). The third is that trauma is stored literally in bodily tissue, whereas what endures on a predictive account is a set of danger priors held with a precision that resists correction, together with a loss of flexibility of inference (Kotler et al., 2026; see also my earlier post on this). Where a model locates a problem shapes how people try to solve it, and a mislocated mechanism can lead someone to regard their body, or their “nervous system”, as fixed and beyond change.

8. An honest limitation

It is worth being explicit about where our updated framework is most exposed, and we say so in the preprint.

Its organising proposition is that the currency in which attachment trades is metabolic, so that a calibrated dyad differs from a miscalibrated one in the cost at which a given regulatory outcome is obtained, rather than in the amount of coordination it displays. That quantity has not been measured in any study we drew upon. Were it measured and found not to differ, our re-description would retain whatever descriptive convenience it has, and would lose the principle that motivates it. We state this not as a caveat but as the point at which the framework is most usefully attacked.

We would also stress that social allostasis is not a diagnostic category, and that the synchrony measures discussed above are research instruments. No threshold exists by which an individual dyad may be classified.

9. A short take home message

Attachment, on this account, is neither a trait nor a store of past experience. It is a running prediction about whether anticipated demand will be met, and by whom. Because it is a prediction rather than an imprint, it is in principle revisable, and because the loop is closed inside one head even when the regulatory work is supplied by another, the dyad is the unit in which it must be assessed.

The preprint is available here: https://osf.io/preprints/psyarxiv/m29kq_v1. It has not yet been certified by peer review, and I very much welcome comments and critique to improve it while it is being scientifically scrutinised.

© 2025-2026 by Pascal Vrticka. All rights reserved.

Scientific References

Barrett, L. F. (2017). The theory of constructed emotion: An active inference account of interoception and categorization. Social Cognitive and Affective Neuroscience, 12(1), 1–23. https://doi.org/10.1093/scan/nsw154

Canterberry, M., & Gillath, O. (2012). Attachment and caregiving. In P. Noller & G. C. Karantzas (Eds.), The Wiley-Blackwell handbook of couples and family relationships (pp. 207–219). Wiley-Blackwell. https://doi.org/10.1002/9781444354119.ch14

Cesario, J., Johnson, D. J., & Eisthen, H. L. (2020). Your brain is not an onion with a tiny reptile inside. Current Directions in Psychological Science, 29(3), 255–260. https://doi.org/10.1177/0963721420917687

Cittern, D., Nolte, T., Friston, K., & Edalat, A. (2018). Intrinsic and extrinsic motivators of attachment under active inference. PLOS ONE, 13(4), e0193955. https://doi.org/10.1371/journal.pone.0193955

Coan, J. A., & Sbarra, D. A. (2015). Social Baseline Theory: The social regulation of risk and effort. Current Opinion in Psychology, 1, 87–91. https://doi.org/10.1016/j.copsyc.2014.12.021

Colledani, D., Meneghini, A. M., Mikulincer, M., & Shaver, P. R. (2022). The Caregiving System Scale: Factor structure, gender invariance, and the contribution of attachment orientations. European Journal of Psychological Assessment, 38(5), 385–396. https://doi.org/10.1027/1015-5759/a000673

De Felice, S., Chand, T., Croy, I., Engert, V., Goldstein, P., Holroyd, C. B., Kirsch, P., Krach, S., Ma, Y., Scheele, D., Schurz, M., Schweinberger, S. R., Hoehl, S., & Vrticka, P. (2025). Relational neuroscience: Insights from hyperscanning research. Neuroscience & Biobehavioral Reviews, 169, 105979. https://doi.org/10.1016/j.neubiorev.2024.105979

Feldman, R. (2017). The neurobiology of human attachments. Trends in Cognitive Sciences, 21(2), 80–99. https://doi.org/10.1016/j.tics.2016.11.007

Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138. https://doi.org/10.1038/nrn2787

Gordon, I., Tomashin, A., & Mayo, O. (2025). A theory of flexible multimodal synchrony. Psychological Review, 132(3), 680–718. https://doi.org/10.1037/rev0000495

Izaki, A., Verbeke, W. J. M. I., Vrticka, P., & Ein-Dor, T. (2024). A narrative on the neurobiological roots of attachment-system functioning. Communications Psychology, 2(1), 96. https://doi.org/10.1038/s44271-024-00147-9

Kotler, S., Mannino, M., Fox, G., & Friston, K. (2026). The body does not keep the score: Trauma, predictive coding, and the restoration of metastability. Frontiers in Systems Neuroscience, 20, 1812957. https://doi.org/10.3389/fnsys.2026.1812957

Long, M., Verbeke, W., Ein-Dor, T., & Vrticka, P. (2020). A functional neuro-anatomical model of human attachment (NAMA): Insights from first- and second-person social neuroscience. Cortex, 126, 281–321. https://doi.org/10.1016/j.cortex.2020.01.010

Motsan, S., Bar-Kalifa, E., Yirmiya, K., & Feldman, R. (2021). Physiological and social synchrony as markers of PTSD and resilience following chronic early trauma. Depression and Anxiety, 38(1), 89–99. https://doi.org/10.1002/da.23106

Nielsen, J. A., Zielinski, B. A., Ferguson, M. A., Lainhart, J. E., & Anderson, J. S. (2013). An evaluation of the left-brain vs. right-brain hypothesis with resting state functional connectivity magnetic resonance imaging. PLOS ONE, 8(8), e71275. https://doi.org/10.1371/journal.pone.0071275

Parr, T., & Friston, K. J. (2018). The anatomy of inference: Generative models and brain structure. Frontiers in Computational Neuroscience, 12, 90. https://doi.org/10.3389/fncom.2018.00090

Saxbe, D. E., Beckes, L., Stoycos, S. A., & Coan, J. A. (2020). Social allostasis and social allostatic load: A new model for research in social dynamics, stress, and health. Perspectives on Psychological Science, 15(2), 469–482. https://doi.org/10.1177/1745691619876528

Theriault, J. E., Katsumi, Y., Reimann, H. M., Zhang, J., Deming, P., Dickerson, B. C., Quigley, K. S., & Barrett, L. F. (2025). It’s not the thought that counts: Allostasis at the core of brain function. Neuron, 113(24), 4107–4133. https://doi.org/10.1016/j.neuron.2025.09.028

Tottenham, N., & Vannucci, A. (2025). Attachment as prediction: Insights from cognitive and developmental neuroscience. Current Directions in Psychological Science, 34(3), 195–206. https://doi.org/10.1177/09637214251313714

White, L., Kungl, M., & Vrticka, P. (2023). Charting the social neuroscience of human attachment (SoNeAt). Attachment & Human Development, 25(1), 1–18. https://doi.org/10.1080/14616734.2023.2167777

White, L., Schulz, C. C., Schoett, M. J. S., Kungl, M. T., Keil, J., Borelli, J. L., & Vrticka, P. (2020). Conceptual analysis: A social neuroscience approach to interpersonal interaction in the context of disruption and disorganization of attachment (NAMDA). Frontiers in Psychiatry, 11, 517372. https://doi.org/10.3389/fpsyt.2020.517372

In my sixth Attachment Insights post, I looked at the popular idea of oxytocin as the “bonding or love hormone” and explained why that story is incomplete. A series of recent papers, several of them led by Daniel Quintana and colleagues, now positions oxytocin not as a dedicated prosocial signal but as an allostatic, energy-regulating and predictive hormone, and most recently as a hormone of behavioural flexibility.

In parallel, our own published perspective in Communications Psychology as “A narrative on the neurobiological roots of attachment-system functioning” (Izaki, Verbeke, Vrticka & Ein-Dor, 2024) places oxytocin mechanistically inside the attachment behavioural cycle in precisely these terms.

In this post, I would like to revisit and extend my Attachment Insights post #06 by looking at oxytocin from an attachment perspective and explaining its function as a (social) allostatic and predictive hormone. Let’s have a closer look!

The structure is as follows:

1. The Incomplete Story: Why “Bonding Hormone” Falls Short

2. From Social Salience to Allostasis

3. The Energy Dimension: Oxytocin and the Body Budget

4. The Predictive Dimension: Oxytocin, Precision, and Active Inference

5. Tend-and-Defend and Tend-and-Befriend

6. Oxytocin Inside the Attachment Cycle

7. Oxytocin as a Behavioural Flexibility Hormone, and Why That Matters for Resilience

8. An Honest Limitation

9. A short take-home message

   Thanks for reading Attachment Science with Dr Pascal Vrticka! All my posts are, and will always remain, free to everyone. In the spirit of my mission, which is the public understanding of science. A paid subscription changes nothing about what you receive, but offers a way to support my work if you value it and are able to.

Summary

Oxytocin is not a dedicated “bonding or love hormone” that reliably increases trust and affiliation. It is better understood as an allostatic hormone that helps our brain maintain stability through change: by adjusting the salience of both social and non-social cues, by participating in energy regulation, and by supporting behavioural flexibility. From an attachment perspective, oxytocin is one of the main biological tools through which co-regulation, or social allostasis, lowers the metabolic cost of managing threat.

1. The Incomplete Story: Why “Bonding or Love Hormone” Falls Short

As I briefly described in my Attachment Insights post #06, the main problem with the “bonding or love hormone” account is an overemphasis on oxytocin’s prosocial and relationship-promoting effects. That emphasis largely traces back to a single influential finding that a nasal dose of oxytocin increased trust in a laboratory-based economic exchange (Kosfeld et al., 2005).

The picture that followed is considerably more complicated. Many subsequent studies using similar procedures were inconclusive or failed to replicate the original effect. And several reported effects that are hard to reconcile with a simple prosocial account. Especially so from an attachment perspective.

For example, oxytocin administration was linked to group-serving dishonesty (Shalvi & De Dreu, 2014). Furthermore, its effects on attachment representations were found to depend on pre-existing individual differences. In one study, more securely attached participants remembered their mothers as more caring and close after oxytocin administration, whereas more anxiously attached participants remembered them as less caring and close (Bartz et al., 2010). In other words, oxytocin intensified antecedent insecure attachment representations rather than uniformly ‘improving‘ them.

The general lesson, captured well by Bartz, Zaki, Bolger and Ochsner (2011), is that context and person matter. A first coherent reframing along these lines is the Social Salience Hypothesis (Shamay-Tsoory & Abu-Akel, 2016), which proposed that oxytocin modulates the salience of social cues, positive or negative, rather than simply promoting affiliation, bonding and (secure) attachment.

2. From Social Salience to Allostasis

Social salience is a useful correction, but it still leaves oxytocin’s many non-social effects unexplained. A broader account does better.

The Allostatic Theory of Oxytocin (Quintana & Guastella, 2020) proposes that oxytocin is most accurately described as an allostatic hormone that maintains stability through changing environments, encompassing both its social and non-social roles.

Two points are worth drawing out. First, the theory is grounded in evolution. Oxytocin-like peptides in evolutionarily ancient organisms regulate basic allostatic processes such as osmoregulation (i.e., the process by which organisms maintain the balance of water and dissolved solutes, like salts and electrolytes, in their bodily fluids), functions that predate anything we would call attachment. On this view, oxytocin’s involvement in attachment is a relatively recent specialisation of a much older regulatory tool, not its defining purpose.

Second, and importantly for what follows, the allostatic theory of oxytocin frames its role in terms of sensing, learning, prediction and response. That is the same predictive-regulatory language we use for allostasis itself, which is our brain’s anticipatory management of the body’s needs. Oxytocin, in this account, helps the system decide which cues to prioritise and how to respond so as to keep the organism stable in the face of change.

3. The Energy Dimension: Oxytocin and the Body Budget

If oxytocin is genuinely allostatic, it should also have metabolic functions, because allostasis is at bottom about managing the body’s energy budget (Theriault et al., 2025). This is exactly what the more recent literature suggests.

A living systematic review and meta-analysis protocol by Sartorius et al. (2025) sets out the evidence that oxytocin is involved in energy balance and metabolism across three domains: feeding behaviour (i.e., reduced food craving and intake), body weight and composition, and biological markers such as glucose. The same protocol is careful to note that human findings are inconsistent, with several weight and metabolic trials returning mixed or null results (see also section 8 on limitations below).

The implication for attachment is conceptual rather than clinical. Oxytocin’s metabolic role is not a separate function bolted onto its social role. Both are facets of the same job: allocating metabolic resources in anticipation of demand. Managing glucose and managing social threat are, from our brain’s point of view, two applications of one regulatory logic.

4. The Predictive Dimension: Oxytocin, Precision, and Active Inference

In a predictive-processing and active-inference view of allostasis, our brain continuously predicts its bodily needs and acts to meet them before they arise, updating its internal model when predictions and sensations diverge (i.e., prediction errors). Within this framework, oxytocin can be read as a modulator of precision weighting, that is, of how much weight the brain assigns to particular cues. “Salience”, in this language, simply means high precision.

Seen this way, oxytocin does two related things. By raising the precision of social cues, it makes trusted others more informative and more able to guide regulation. And by lowering the predicted cost of managing threat in the company of others, it biases the system toward social means of regulation (i.e., co-regulation, or social allostasis) over more costly individual regulation.

Harrison and colleagues (2025) make this link explicit, placing oxytocin within an active-inference account of allostasis and connecting it to resilience through the notion of “affiliative resilience”. This coincides directly with the framework I set out in my post on the Bioenergetic Map of Social Connection, where secure attachment relationships function as a way of sharing the predictive and metabolic load.

5. Tend-and-Defend and Tend-and-Befriend

The context-dependence of oxytocin is also visible at the behavioural level, and two well-known patterns illustrate it.

Tend-and-defend captures the observation that oxytocin can promote care and cooperation toward one’s own group while increasing wariness, or even hostility, toward outsiders. The finding that oxytocin can facilitate group-serving dishonesty (Shalvi & De Dreu, 2014) fits this pattern. Oxytocin is not uniformly “nice”; it sharpens the boundary between in-group and out-group in the service of protecting what is close.

Tend-and-befriend (Taylor, 2006) describes a different facet: under stress, oxytocin is associated with increased affiliation and support-seeking rather than only ‘fight-or-flight’. This pattern is more prominent in individuals with attachment anxiety. It maps neatly onto what, in our Perspective, we call “social flight”, a threat response in which proximity-seeking is directed specifically toward attachment figures rather than simply away from danger.

Both concepts describe the same underlying move: recruiting the social environment as a means for co-regulation.

6. Oxytocin Inside the Attachment Cycle

Our Perspective (Izaki, Verbeke, Vrticka & Ein-Dor, 2024) traces a cascade of neurobiological events across attachment behaviour, and oxytocin does most of its work at the reunion and distress-alleviation phase, once proximity to an attachment figure has been achieved (see Figure below). The details are worth stating, because together they show oxytocin operationalising social allostasis.

Following reunion with an attachment figure, oxytocin reduces reactivity of the hypothalamic-pituitary-adrenal axis (largely by potentiating GABA-A signalling), which lowers cortisol. It reduces dopamine in the substantia nigra pars compacta (by around 70% in animal work), which decreases locomotion and the motivation to move, bringing proximity-seeking to a halt so that co-regulation can begin.

At the same time, oxytocin increases dopamine in the ventral tegmental area and downstream mesolimbic and mesocortical pathways, tagging attachment figures with incentive salience (”wanting”) and engaging the ventromedial prefrontal cortex as a social safety signal. These functions support the “liking” side of the reunion through opioid signalling in the nucleus accumbens, in line with the Brain Opioid Theory of Social Attachment.

Oxytocin also enables pain relief through a dedicated population of parvocellular oxytocin neurons projecting to the spinal cord. And it drives long-term plasticity and memory consolidation, acting as a social salience signal that amplifies and stabilises neural responses to relevant social cues (Froemke & Young, 2021).

Read together, oxytocin is a hormone ending a costly, mobilised state, marking the social resource as worth returning to, and consolidating that lesson for future prediction. That is social allostasis in action, not bonding for its own sake.

7. Oxytocin as a Behavioural Flexibility Hormone, and Why That Matters for Resilience

The most recent step in this line of work takes the argument further. Walle et al. (2026) reframe oxytocin as a behavioural flexibility hormone, on both evolutionary and neurobiological grounds. They point to the wide distribution of oxytocin receptors across the central nervous system and periphery, their co-expression with dopamine D2 receptors in corticostriatal circuits associated with learning and cognitive control, and evidence that oxytocin increases the flexibility of large-scale network dynamics, for example the variability of default mode network coupling.

The point I want to draw out is their conclusion that flexibility is central to resilience. This closes a loop with two of my earlier posts. In Does the Body Really “Keep the Score”? I argued that trauma is best understood as a loss of flexibility, a collapse of metastability, and that recovery means restoring it. In the Bioenergetic Map of Social Connection I described resilience as the efficiency and flexibility of social allostatic regulation. An allostatic, flexibility-supporting oxytocin system is one plausible biological substrate through which secure co-regulation could help restore that flexibility. It is not the whole story, but it is a coherent piece of it.

8. An Honest Limitation

As with any account of this kind, it is important to be clear about status. These are organising frameworks with strong theoretical coherence, not settled facts.

Much of the human oxytocin literature is underpowered and inconsistent, to the point where many individual studies cannot reliably detect or reject even moderate effects (Quintana, 2020). Intranasal administration, the workhorse of human oxytocin research, carries its own uncertainties about how much reaches the brain and in what form, and the metabolic findings remain mixed (Sartorius et al., 2025). The field is responding constructively, with living meta-analyses and shared infrastructure (Iversen et al., 2026) and with preregistered syntheses of non-social effects (Kang et al., 2025).

There is also an important distinction between the dynamics of endogenous oxytocin and the effects of exogenous (i.e., nasal spray) administration. Most of the strong “bonding or love hormone” claims, and many of the null results, come from the latter.

The right conclusion is therefore not that we can indiscriminately raise oxytocin to make people more securely attached. It is that oxytocin is best understood as a context-dependent, allostatic and predictive modulator whose role only makes sense within the wider regulatory and relational system, alongside dopamine, cortisol, the opioids and many other players described in our functional neuro-anatomical models of Attachment (NAMA and NAMDA).

9. A short take-home message

When you read about oxytocin the next time, it is worth holding all of these accounts in mind at once.

Oxytocin is not a switch for love or trust. It is a hormone for maintaining stability through change: adjusting which cues matter, helping to regulate the body’s energy, ending costly states of mobilisation once safety is reached, and supporting the flexibility that both attachment security and resilience depend on. From an attachment perspective, that is precisely what we would expect from a central tool of social allostasis.

© 2025-2026 by Pascal Vrticka. All rights reserved.

Scientific References

Bartz, J. A., Zaki, J., Bolger, N., & Ochsner, K. N. (2011). Social effects of oxytocin in humans: Context and person matter. Trends in Cognitive Sciences, 15(7), 301–309. https://doi.org/10.1016/j.tics.2011.05.002

Bartz, J. A., Zaki, J., Ochsner, K. N., Bolger, N., Kolevzon, A., Ludwig, N., & Lydon, J. E. (2010). Effects of oxytocin on recollections of maternal care and closeness. Proceedings of the National Academy of Sciences, 107(50), 21371–21375. https://doi.org/10.1073/pnas.1012669107

Froemke, R. C., & Young, L. J. (2021). Oxytocin, neural plasticity, and social behavior. Annual Review of Neuroscience, 44, 359–381. https://doi.org/10.1146/annurev-neuro-102320-102847

Harrison, L.A., Gracias, A.J., Friston, K.J. and Buckwalter, J.G. (2025) Resilience phenotypes derived from an active inference account of allostasis. Frontiers in Behavioral Neuroscience, 19:1524722. https://doi.org/10.3389/fnbeh.2025.1524722

Iversen, I. A., Alaerts, K., Bakermans-Kranenburg, M., Becker, B., Blair, R. J., Bartz, J. A., … Quintana, D. S. (2026). The active monitoring of oxytocin research evidence (AMORE) platform. Psychoneuroendocrinology, 185, Article 107713. https://doi.org/10.1016/j.psyneuen.2025.107713

Izaki, A., Verbeke, W. J. M. I., Vrticka, P., & Ein-Dor, T. (2024). A narrative on the neurobiological roots of attachment-system functioning. Communications Psychology, 2, Article 96. https://doi.org/10.1038/s44271-024-00147-9

Kang, H., Glaser, B. D., Sartorius, A. I., Audunsdottir, K., Kildal, E. S., Nærland, T., Andreassen, O. A., Westlye, L. T., & Quintana, D. S. (2025). Effects of oxytocin administration on non-social executive functions in humans: a preregistered systematic review and meta-analysis. Molecular psychiatry, 30(5), 2239–2251. https://doi.org/10.1038/s41380-024-02871-4

Kosfeld, M., Heinrichs, M., Zak, P. J., Fischbacher, U., & Fehr, E. (2005). Oxytocin increases trust in humans. Nature, 435, 673–676. https://doi.org/10.1038/nature03701

Quintana, D. S. (2020). Most oxytocin administration studies are statistically underpowered to reliably detect (or reject) a wide range of effect sizes. Comprehensive Psychoneuroendocrinology, 4, Article 100014. https://doi.org/10.1016/j.cpnec.2020.100014

Quintana, D. S., & Guastella, A. J. (2020). An allostatic theory of oxytocin. Trends in Cognitive Sciences, 24(7), 515–528. https://doi.org/10.1016/j.tics.2020.03.008

Sartorius, A. I., Kang, H., Deilhaug, E., Walle, K. M., Lawson, E. A., Eddy, K., van der Meer, D., Westlye, L. T., & Quintana, D. S. (2025). Oxytocin and its role in energy balance and metabolic factors: A living systematic review and meta-analysis protocol [Preprint]. medRxiv. https://doi.org/10.1101/2025.04.04.25325238

Shalvi, S., & De Dreu, C. K. W. (2014). Oxytocin promotes group-serving dishonesty. Proceedings of the National Academy of Sciences, 111(15), 5503–5507. https://doi.org/10.1073/pnas.1400724111

Shamay-Tsoory, S. G., & Abu-Akel, A. (2016). The social salience hypothesis of oxytocin. Biological Psychiatry, 79(3), 194–202. https://doi.org/10.1016/j.biopsych.2015.07.020

Taylor, S. E. (2006). Tend and befriend: Biobehavioral bases of affiliation under stress. Current Directions in Psychological Science, 15(6), 273–277. https://doi.org/10.1111/j.1467-8721.2006.00451.x

Theriault, J. E., Katsumi, Y., Reimann, H. M., Zhang, J., Deming, P., Dickerson, B. C., Quigley, K. S., & Barrett, L. F. (2025). It’s not the thought that counts: Allostasis at the core of brain function. Neuron, 113(24), 4107–4133. https://doi.org/10.1016/j.neuron.2025.09.028

Walle, K. M., Sartorius, A. I., Deilhaug, E., Kang, H., Nærland, T., Andreassen, O. A., Westlye, L. T., & Quintana, D. S. (2026). Reframing oxytocin as a behavioral flexibility hormone. Neuroscience & Biobehavioral Reviews, 185, Article 106628. https://doi.org/10.1016/j.neubiorev.2026.106628

The past few months have brought real movement in how we think about insecure attachment and trauma. Two of the most popular frameworks, Stephen Porges’s Polyvagal Theory and Bessel van der Kolk’s “The body keeps the score”, have come under serious and, in my reading, largely justified scientific scrutiny. I have written about both already, in my previous posts on Polyvagal Theory and on whether the body really keeps the score, so I will not rehearse those arguments here.

Instead, I want to do two things. First, to notice that an alternative is already being built, by several people from several directions, and to add one more strand to it from the corner of science I work in, the social neuroscience of human attachment (SoNeAt). And second, and more importantly, to argue that this is a conversation worth keeping open and collaborative. It matters a great deal, it is easy for it to turn sour, and it will only produce something better than what it replaces if a lot of different people stay at the table together.

An alternative is already taking shape

It would be a mistake to think the ground is empty. Several people have already begun, from their own vantage points, to build what comes after these frameworks, and I am grateful for each of those efforts.

Paul Grossman and his colleagues did the careful, and necessary, work of setting out where the neurophysiology of Polyvagal Theory does not hold, in their expert evaluation to which Porges readily replied. Clearing the ground is not a small thing, and it is a precondition for anything better.

Kotler, Friston and their colleagues then took a further step in their paper “The body does not keep the score”, sketching a genuine alternative built on prediction and allostasis at the level of the individual brain: trauma not as something stored, but as a loss of flexibility in how the brain forecasts and updates.

And , among others, has done the valuable work of translating all this for the people who most need it, therapists and practitioners left wondering what to say to clients once the old frameworks fall away.

Between them, a shape is emerging: prediction and allostasis in place of storage, constructed emotion in place of fixed bodily read-outs, and the reassurance that the ideas people found useful, safe relationships and the reality of the body, do not depend on the discredited mechanisms. I think that shape is right. What I would like to add is one piece that could make it stronger.

What the social neuroscience of attachment (SoNeAt) can add

The piece I have in mind is the social one.

The alternative, as it currently stands, is largely an account of an individual brain. How one person’s predictions become rigid, and how one person’s flexibility might return. That is true and important. But, in my opinion, it is not the whole picture, because human brains do not regulate alone.

Research on social allostasis and Social Baseline Theory shows that the human brain treats the presence of a trusted other as its expected baseline, and that a reliable companion literally takes on part of our regulatory load. I have tried to map some of this in my previous post on “The Bioenergetic Map of Social Connection”. Regulation, on this view, is something we do together first, and alone only as a fallback.

Attachment is what that social regulation looks like across a lifetime. In our social neuroscience of human attachment (SoNeAt), we characterise attachment patterns as the brain’s running estimate of whether others will be there to help carry the load, and insecure patterns are best read not as defects but as context-dependent predictions. Anxiety as a bet that support is intermittent, avoidance as a bet that it is absent.

Moreover, in our “functional neuro-anatomical model of disrupted attachment (NAMDA)”, my colleagues and I describe how overwhelming, and often relational, threat can push this system into the braced or shut-down states we recognise as trauma, and why so much of recovery runs through relationships.

None of this contradicts the individual account. It extends it. If Kotler, Friston and colleagues describe allostasis in a single brain, what SoNeAt adds is the same principle scaled up to the relationships in which that brain actually lives. That is why I think a combined account, one that holds allostasis, active inference, and attachment together, can strengthen the alternative others have already begun, rather than compete with it.

The SoNeAt account, in outline

For anyone who wants the whole picture in one place, here it is, with links onward for the detail.

The brain’s primary job is to anticipate, and to keep the body’s energy budget in balance by predicting what will be needed before it is needed. That is allostasis, carried out, on the active inference view, by making predictions and correcting them against what the body and world report back.

Emotions are part of this: on the theory of constructed emotion, a racing heart becomes fear only when the brain reads it as fear, and the same arousal might instead become excitement.

Trauma, then, is not a recording in the tissues but a set of predictions grown too rigid to update, as I argue in “Does the Body Really Keep the Score”. And because the budgeting is social, attachment is the long-run shape it takes. Our relationships tune the brain’s expectations about whether help will be there, and those expectations, in turn, shape a life. A predicting, budgeting brain that constructs its emotions and regulates itself, first of all, with trusted others.

Our most recent conceptualisation of attachment as social allostatic inference has now been published as a preprint and a summary of it can be found here.

Why an alternative is worth building

It would be fair to ask why any of this is urgent. Is it not enough to say what is wrong and leave people to it? I do not think so, for a simple reason: a gap like this does not stay open for long.

When a compelling framework is discredited and nothing coherent is offered in its place, the space tends to be filled by the next confident but incomplete story, and the cycle repeats.

People who are suffering, and the clinicians and therapists who sit with them, deserve better. An alternative account that is accurate, that is hopeful—since most people exposed to trauma do recover, and a prediction, unlike an imprint, can be revised—and that is usable in the room. Building that account well, and not only taking the old one apart, is worth the effort.

A conversation worth keeping open

Which brings me to the point I care about most.

This is an important discussion, and important discussions can turn sour easily, especially online, and especially when people’s identities and livelihoods are bound up in the ideas at stake. I have felt the pull towards sharper language myself, and I have watched exchanges that could have been generative harden into camps instead. I would very much like this one not to go that way.

The only way I can see to avoid it is to keep the table wide. Building what comes next is genuinely beyond any single domain. It needs clinicians and psychotherapists, who know what helps a person and what an account has to do to be usable. It needs science communicators, who carry accurate ideas to the people who need them. And it needs neurophysiologists and neuroscientists, who work on the mechanistic side of the picture, including its social aspects. No one of us can build the replacement alone. Together, I think we can.

So please take this as what it is meant to be: not a claim to have the answer, and not an advertisement for my own work, but an honest contribution to a conversation I believe matters, and an invitation to keep it shared and generous. I have set out where I think SoNeAt can strengthen the account others have started. I would much rather see it constructively argued with, improved and folded into something better than any of us would manage alone, than see it, or any of the good work already done, lost to a quarrel none of us needs to have.

© 2025-2026 by Pascal Vrticka. All rights reserved.

In my thirty-eighth and most recent post, I would like to have a closer look at a thought-provoking new paper by Kotler, Mannino, Fox and Friston (2026) with the deliberately provocative title “The body does not keep the score: trauma, predictive coding, and the restoration of metastability.” One of its authors is Karl Friston, a leading figure in computational neuroscience and the originator of the free energy principle. Together, the authors revisit one of the most influential ideas in contemporary trauma discourse, Bessel van der Kolk’s The Body Keeps the Score (2014), as part of what they call an “updated reframing”.

I find this an especially valuable contribution because it does what I try to do throughout my Attachment Insights series: it takes a widely embraced idea seriously, acknowledges what is genuinely useful about it, and only then carefully updates the underlying neurobiology in light of current science.

Let’s have a closer look!

Summary

For nearly a decade, the phrase “the body keeps the score” has shaped how the public and many clinicians understand trauma. It is an emotionally compelling metaphor, and it carries some real and important truths. But recent work in computational and systems neuroscience suggests that, taken literally, the image is biologically inaccurate.

The body proper does not store trauma in its tissues. Instead, the brain dynamically re-enacts it through maladaptive prediction. What endures after trauma is not a memory locked in flesh, but a loss of flexibility. A collapse of metastability, the brain’s ability to fluidly switch among semi-stable network states.

In what follows, I first want to honour what van der Kolk and the “body keeps the score” framework get right. Only then will I turn to where the metaphor misleads, and how a predictive, and crucially also a social, account of trauma offers a subtler and, I believe, more hopeful story.

1. First, the Value: What “The Body Keeps the Score” Gets Right

Before challenging anything, it is only fair and scientifically accurate to recognise the genuine contributions of van der Kolk’s work. Kotler et al. (2026) are careful to do so, and so am I.

The book’s central and lasting achievement was to insist that trauma is deeply embodied and not merely “in the head”.

For many trauma survivors whose suffering had been dismissed as imaginary or purely psychological, this was validating and liberating. It brought concepts like interoception (our sense of the body’s internal state) and embodied cognition (a research program that emphasises the role of the body in cognitive processes) into mainstream clinical and public awareness, and it did so with compassion.

The framework also draws substantially on solid neuroscience. It discusses the dynamic interplay between neural networks and Antonio Damasio’s Somatic Marker Hypothesis (the well-supported idea that emotions are fundamentally tied to bodily states). And it rightly emphasises that traumatic memory is qualitatively different from ordinary memory, and that such memories can re-emerge after years of not being remembered.

In short, van der Kolk was right that the body is central to the traumatic process, that somatic states do matter, and that trauma cannot be reduced to a purely “cognitive” problem to be reasoned away. These insights remain valuable and should not be discarded.

The disagreement, as we will see, is not about whether the body matters, but about how and where trauma actually “resides”.

2. From Storage to Prediction: Where the Metaphor Misleads

The difficulty lies in the most memorable reading of the book: the notion that trauma “lives in the body” in a nearly literal sense. Inscribed in muscle, fascia, or other tissue apart from the brain’s neural innervation. This is a reading the metaphor strongly invites, even where the text itself is more careful.

To see why this is problematic, we need one central idea from modern neuroscience: our brain does not passively record the world, it actively predicts it.

According to the predictive coding and active inference framework (Friston, 2010), perception is not a camera faithfully registering reality. Instead, the brain is constantly generating predictions about the causes of its sensations and updating them only when prediction errors arise. Making sense of the world, and of our own bodies, is a process of continuous (Bayesian) belief updating that minimises “surprise”.

From this vantage point, trauma is not a thing that gets deposited somewhere. It is a disorder of prediction, not of storage. The body participates fully in trauma, but as a messenger relaying the brain’s current predictions, not as an archive holding the past.

This is not a merely semantic quibble. As I have discussed elsewhere in relation to the Polyvagal Theory and the idea of “Getting stuck in certain brain or nervous system states”, where we locate a problem strongly shapes how we try to solve it.

3. Trauma as Over-Weighted Danger Priors

So what actually “goes wrong” after trauma? In computational terms, trauma over-weights the precision of danger priors. “Precision” here simply means confidence: the brain assigns excessive certainty to its threat predictions, and then constrains all subsequent inference on the premise of enduring danger.

Kotler et al. offer an elegant image for this. Picture belief as a landscape with hills and valleys (i.e., a free energy landscape) where every possible interpretation of the world has an elevation reflecting its (im)plausibility. A healthy mind moves fluidly across this terrain. Trauma, by contrast, carves out a deep, steep-sided ravine: a local minimum that becomes so precisely and rigidly engrained that one can no longer climb out of it.

Hypervigilance, flashbacks, and avoidance are then best understood not as fragments of stored experience resurfacing, but as the symptoms of a system trapped in self-confirming predictions.

Functional imaging is consistent with this picture. The dominant neuroimaging finding in PTSD is a dissociation between amygdala hyperactivation (i.e., heightened threat detection as part of the larger salience network) and medial prefrontal hypoactivation (i.e., weakened top-down regulation as part of the larger executive functioning network), with less prefrontal engagement associated with more severe symptoms (Shin et al., 2006). The net effect is a reduction in the brain’s dynamic repertoire: connectivity shifts toward defensive configurations, and the capacity for flexible reconfiguration is diminished.

4. The “Score” as Circular Inference

If trauma is not stored in the body, why does it feel so bodily? Why the racing heart, the tight chest, the sense that the body itself remembers?

The answer lies in a mechanism called sensory attenuation. Normally, the brain can resolve prediction errors not only by acting or by updating its internal model, but also by simply assigning less precision to certain signals (i.e., dampening down bodily sensations it deems predictable or irrelevant). In trauma, this capacity to attenuate is reduced, and internal signals are mis-weighted.

Unattenuated interoceptive signals (e.g., a pounding heart, a clenched stomach) are then interpreted as confirmation of danger rather than as imprecise noise. The result is a vicious loop of circular inference: the brain predicts pain, senses arousal, and takes that arousal as proof that the pain persists.

This is the real origin of the “score” the body appears to keep. It is not an inscription in tissue but an artefact of a self-reinforcing inferential loop.

Damasio’s own work makes the same point at the level of neural architecture. His convergence–divergence zones (CDZs) are neural hubs that store the parameters for a memory and its associated emotion and then coordinate the flexible, context-dependent reactivation of those patterns. They do not encode raw sensory activations “in the body”.

In PTSD, one can hypothesise that CDZs become overly primed to a wide range of cues, so that even a distantly related trigger (e.g., the classic example of a car backfiring) can unleash the full cascade of a flashback. Emotions and their bodily correlates are, in this view, modifiable and adaptive, without any need for hidden storage in non-innervated tissue.

5. The Real Casualty: A Loss of Metastability

We can now name what trauma actually erodes. A healthy brain is metastable: its networks continually integrate and segregate, assembling into transient coalitions and dissolving again, allowing the system to explore multiple interpretations of the world. Hellyer and colleagues (2015) showed that this metastability is a hallmark of cognitive flexibility, and that its reduction is associated with impaired information processing.

Trauma erodes precisely this fluidity. It traps the brain in narrow basins of fear and defensive salience, collapsing the graceful movement across a landscape of beliefs, commitments and intentions that characterises mental health.

Seen this way, restoring wellbeing is not about “releasing” stored emotion but about re-establishing dynamic equilibrium, expanding the brain’s repertoire of viable states. Health, in a single word, equals flexibility.

6. Why Most People Recover: Resilience Favours the Inference Model

One of the most compelling pieces of indirect evidence comes from resilience research. Across large longitudinal cohorts, George Bonanno and colleagues have repeatedly shown that the majority of trauma-exposed individuals do not develop chronic PTSD (Bonanno, 2004). A systematic review of 54 studies furthermore identified a resilient trajectory in roughly 65.7% of cases across populations (Bonanno et al., 2015). Most people, in other words, adapt.

To be clear, this does not logically disprove a storage account. A storage model could, in principle, accommodate individual differences in vulnerability. But the sheer prevalence of spontaneous recovery is much more parsimoniously explained by an inference model.

Under typical conditions, the brain simply rebalances its own predictive architecture, restoring metastability and flexibility without any therapeutic intervention. Pathology arises not from the permanence of what was stored, but from the failure of that recalibration process.

This resonates strongly with a theme I return to again and again in my Attachment Insights series: insecure attachment patterns and post-traumatic responses are, in the first instance, meaningful, adaptive and often necessary reactions to overwhelming circumstances, and resilience is the norm rather than the exception (see also the integrative definition of attachment trauma by Zagaria et al., 2024, in my Attachment Insights post #26 “What is Attachment Trauma? A New Integrative Definition”). Reframing trauma this way guards against pathologising what is, for most people, a recoverable disruption.

7. Restoring Flexibility: Flow, Cognitive Control and Co-Regulation

If flexibility is what trauma steals, how do we give it back? Kotler et al. propose a fascinating candidate: flow states. The experience of complete absorption in a meaningful, high-challenge activity.

During flow, the brain appears to undergo transient, localised quieting of self-referential processing alongside rapid, adaptive reconfiguration of sensory-motor and attentional networks. Theoretical models predict surges in neuromodulators such as dopamine, norepinephrine and anandamide (Kotler et al., 2022), and a shift from avoidance to approach dynamics that converts arousal into focused action rather than defensive freezing.

In this sense, trauma and flow can be seen as divergent responses to the same physiological stress signal. One collapses movement, the other restores it. Early clinical signals are encouraging: in a randomised controlled trial, surf and hike therapy both improved depression in active-duty service members, with some advantage for surfing at follow-up (Walter et al., 2023); though, as the authors honestly note, generic factors like exercise, nature and social connection are hard to disentangle here.

This is a genuinely useful reframing, and it dovetails with evidence that training cognitive control (e.g., through mindfulness, working-memory practice, or executive-function exercises) reduces intrusive symptoms (Bomyea & Amir, 2011). The unifying idea is that many diverse therapies work not through any specific content, but by restoring flexible coupling between large-scale networks.

Here, though, I would like to add a dimension that the paper (understandably focused on individual dynamics) leaves largely implicit. The most fundamental route to restoring predictive flexibility is not solitary. It is social.

Recent work reframes the brain’s primary job as allostasis: the predictive regulation of the body’s energy budget (Theriault et al., 2025). And because all prediction and regulation are metabolically costly, our brains evolved to share those costs with others.

This is the core insight of Social Baseline Theory (Coan & Sbarra, 2015): the human brain treats the presence of trusted others as its expected baseline condition, using risk distribution and load sharing to conserve metabolic resources. Through social allostasis, a reliable companion literally does some of our predictive and regulatory work for us (see, e.g., Saxbe et al., 2020). This is also why, as I discussed in my Substack post on the “Bioenergetic Map of Social Connection”, a hill looks less steep when a good friend stands beside us.

From an attachment perspective, this matters enormously for trauma. Secure attachment is, in predictive-processing terms, an internal working model that others are reliably available for co-regulation (White, Kungl & Vrticka, 2023). Such a model provides exactly the bioenergetic safety a traumatised brain needs in order to loosen its over-precise danger priors, re-open its dynamic repertoire, and re-learn that arousal and safety can coexist.

Indeed, an active-inference account of allostasis (co-authored by Friston himself) links precisely these predictive-regulatory mechanisms to distinct resilience phenotypes (Harrison et al., 2025), building a direct bridge between the paper’s framework and the social neuroscience of connection (see my Substack post on the “Bioenergetic Map of Social Connection”).

In other words: flow restores metastability from the inside, through challenge and absorption. Secure co-regulation restores it from the outside, by lending us social resources to help carry the predictive and allostatic load. Tellingly, the flow-inducing activities in the surf-and-hike trial were also deeply relational and communal. A reminder that these two routes are not rivals but partners.

8. Why This Reframing Matters

Where does all of this leave us? Not, I want to stress, with any diminishment of how much trauma hurts. Framing trauma dynamically does not make the suffering less real. But it grounds that suffering in neurobiological mechanisms that can actually be addressed.

The practical shift is from a language of “excavation” to one of “exploration”. If trauma were buried injury, healing would mean finding and purging what was stored. If trauma is instead frozen inference, healing means introducing variability, recalibrating precision, and expanding the brain’s capacity for adaptive flexibility. Often, and most efficiently, in the company of trusted others.

This is also why expression, catharsis, or “somatic release” can genuinely help: not because they discharge buried energy, but because they introduce novel sensorimotor feedback that loosens a rigid model.

This perspective also dissolves a long-standing puzzle: why so many different treatments (e.g., exposure therapy, EMDR, mindfulness, exercise, psychedelics, secure relationships, flow-inducing pursuits) can all succeed. Each, in its own way, restores flexible coupling between networks and rebalances an over-confident predictive system. And it guards against a subtle harm in the literal “storage” reading: the risk of convincing people that their bodies are indelibly marked, when in fact most brains retain the very capacity for metastable inference that recovery requires.

If the old story was that “the body keeps the score”, the emerging narrative is subtler and more hopeful. The body does not keep the score; the brain keeps predicting it. When prediction becomes too rigid, experience repeats. Not because it is stored, but because it cannot yet be reinterpreted.

The above said, it’s time for an honest limitation, as I must add the same caution the authors themselves raise. The specific claim that PTSD involves reduced metastability has not yet been directly tested using established metastability metrics in PTSD populations. The Hellyer et al. (2015) findings concern traumatic brain injury (i.e., a neurologically distinct condition) and structural connectome damage should not be conflated with the functional dysregulation of PTSD. This framework is therefore best understood as a compelling, testable prediction rather than an established fact (Hancock et al., 2025). Future work should quantify metastability (i.e., signal variability, entropy, network switching) as a clinical biomarker before and after intervention.

And finally, a short take home message. Healing, in the end, is not the erasure of what happened, but the return of movement: within the mind, within its underlying brain networks, within our actions. And, I would add, within our secure relationships. Because for a fundamentally social species, the safest and most efficient place to rediscover flexibility is rarely alone.

Addendum: A new preprint is now available in which Prof Tsachi Ein-Dor and I more formally introduce our account of attachment (both in its organised and disrupted/disorganised form) as social allostatic inference. You can access the preprint here: “Attachment and Caregiving as Social Allostatic Inference: An Updated Social Neuroscience Account from Single Brains to Dyads”; https://osf.io/preprints/psyarxiv/m29kq_v1.

A more accessible summary is furthermore available on Substack here:

© 2025-2026 by Pascal Vrticka. All rights reserved.

Scientific References

Bomyea, J., & Amir, N. (2011). The effect of an executive functioning training program on working memory capacity and intrusive thoughts. Cognitive Therapy and Research, 35(6), 529–535. https://doi.org/10.1007/s10608-011-9369-8

Bonanno, G. A. (2004). Loss, trauma, and human resilience: Have we underestimated the human capacity to thrive after extremely aversive events? American Psychologist, 59(1), 20–28. https://doi.org/10.1037/0003-066X.59.1.20

Bonanno, G. A., Romero, S. A., & Klein, S. I. (2015). The temporal elements of psychological resilience: An integrative framework for the study of individuals, families, and communities. Psychological Inquiry, 26(2), 139–169. https://doi.org/10.1080/1047840X.2015.992677

Coan, J. A., & Sbarra, D. A. (2015). Social Baseline Theory: The social regulation of risk and effort. Current Opinion in Psychology, 1, 87–91. https://doi.org/10.1016/j.copsyc.2014.12.021

Damasio, A. R. (1994). Descartes' error: Emotion, reason, and the human brain. G. P. Putnam's Sons.

Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138. https://doi.org/10.1038/nrn2787

Hancock, F., Rosas, F. E., Luppi, A. I., Zhang, M., Mediano, P. A. M., Cabral, J., Deco, G., Kringelbach, M. L., Breakspear, M., Kelso, J. A. S., & Turkheimer, F. E. (2025). Metastability demystified - the foundational past, the pragmatic present and the promising future. Nature Reviews Neuroscience, 26(2), 82–100. https://doi.org/10.1038/s41583-024-00883-1

Harrison, L. A., Gracias, A. J., Friston, K. J., & Buckwalter, J. G. (2025). Resilience phenotypes derived from an active inference account of allostasis. Frontiers in Behavioral Neuroscience, 19, Article 1524722. https://doi.org/10.3389/fnbeh.2025.1524722

Hellyer, P. J., Scott, G., Shanahan, M., Sharp, D. J., & Leech, R. (2015). Cognitive flexibility through metastable neural dynamics is disrupted by damage to the structural connectome. Journal of Neuroscience, 35(24), 9050–9063. https://doi.org/10.1523/JNEUROSCI.4648-14.2015

Kotler, S., Mannino, M., Fox, G., & Friston, K. (2026). The body does not keep the score: Trauma, predictive coding, and the restoration of metastability. Frontiers in Systems Neuroscience, 20, Article 1812957. https://doi.org/10.3389/fnsys.2026.1812957

Kotler, S., Mannino, M., Kelso, S., & Huskey, R. (2022). First few seconds for flow: A comprehensive proposal of the neurobiology and neurodynamics of state onset. Neuroscience & Biobehavioral Reviews, 143, 104956. https://doi.org/10.1016/j.neubiorev.2022.104956

Saxbe, D. E., Beckes, L., Stoycos, S. A., & Coan, J. A. (2020). Social allostasis and social allostatic load: A new model for research in social dynamics, stress, and health. Perspectives on Psychological Science, 15(2), 469–482. https://doi.org/10.1177/1745691619876528

Shin, L. M., Rauch, S. L., & Pitman, R. K. (2006). Amygdala, medial prefrontal cortex, and hippocampal function in PTSD. Annals of the New York Academy of Sciences, 1071, 67–79. https://doi.org/10.1196/annals.1364.007

Theriault, J. E., Katsumi, Y., Reimann, H. M., Zhang, J., Deming, P., Dickerson, B. C., Quigley, K. S., & Barrett, L. F. (2025). It’s not the thought that counts: Allostasis at the core of brain function. Neuron, 113(24), 4107–4133. https://doi.org/10.1016/j.neuron.2025.09.028

van der Kolk, B. A. (2014). The body keeps the score: Brain, mind, and body in the healing of trauma. Penguin Books.

Walter, K. H., Otis, N. P., Glassman, L. H., Ray, T. N., Michalewicz-Kragh, B., & Thomsen, C. J. (2023). A randomized controlled trial of surf and hike therapy for U.S. active duty service members with major depressive disorder. BMC Psychiatry, 23, 119. https://doi.org/10.1186/s12888-022-04452-7

White, L., Kungl, M., & Vrticka, P. (2023). Charting the social neuroscience of human attachment (SoNeAt). Attachment & Human Development, 25(1), 1–18. https://doi.org/10.1080/14616734.2023.2167777

Zagaria, A., Fiori, F., Vacca, M., Lombardo, C., Pariante, C. M., & Ballesio, A. (2024). Toward a definition of attachment trauma: Integrating attachment and trauma studies. European Journal of Trauma & Dissociation, 8(3), Article 100416. https://doi.org/10.1016/j.ejtd.2024.100416

This time, in my thirty-sixth and most recent post, I would like to discuss a novel social neuroscience account of attachment by diving deeper into the concepts of allostasis and predictive processing.

Summary

A converging body of recent empirical and theoretical work (see Scientific References at the bottom of this post) identifies our brain’s primary function as the anticipation of our body’s and mind’s needs. In other words, our brain’s central task is to be ideally prepared to meet future demands before they arise. This active and predictive regulatory process is also known as allostasis, a process that ensures ‘stability through change’.

Being an integral part of allostasis, attachment can be redefined as a fundamental bioregulatory system that manages our ‘metabolic capital’ by modelling the availability of social resources for co-regulation.

But what does this mean in detail, and how does an optimal combination of allostasis and attachment give rise to resiliency? Let’s have a closer look!

1. The Predictive Manager: Integrating Homeostasis and Allostasis

Biological life requires the constant management of an energy budget. More energy must be taken in than is expended to maintain an ‘economy of action’. This dynamic stability is achieved through a nested hierarchy of regulation.

At the local level, homeostasis, as an automatic and reactive regulatory mechanism, operates within tight boundaries and short time frames. Its main function is to return specific metabolic parameters (like blood sugar or oxygen levels) to a fixed set point after the occurrence of a relatively small and brief disturbance.

Such minor corrections as part of homeostasis are guided by the broader process of allostasis. Allostasis is the response to more significant challenges that operate outside the constraints of homeostasis and involve an active regulatory effort to return back to balance (i.e., ‘stability through change’).

Importantly, while homeostasis is reactive, allostasis is predictive. Our brain orchestrates allostasis through active inference, acting as a ‘betting machine’.

By utilising past experiences alongside current internal and external information, our brain predicts how to best deploy metabolic resources in light of expected returns. Through this process, our brain constructs momentarily optimised ensembles of signals to reduce uncertainty (i.e., the discrepancy between prediction and reality, so-called ‘prediction errors’) and guide action in a metabolically efficient manner.

2. Resilience as our Energy Budget’s Efficiency

In this integrated model of reactive homeostasis and predictive allostasis, resilience can be understood as the flexibility and efficiency of these regulatory systems. A resilient system maintains stability across varied challenges without exhausting its bioenergetic resources.

Crucially, all regulation is bioenergetically costly. Every act of maintaining our body’s internal milieu or planning a movement requires metabolic work. The core of resilience therefore lies in how these costs are anticipated and effectively mitigated.

What’s more, regulation, and particularly so via allostasis, often involves high-effort, future-thinking processes that are computationally and metabolically expensive. Sustained reliance on these costly prediction processes, without any support, can therefore lead to a loss of system flexibility. If that happens, our brain’s main function of active regulation and anticipating future needs can become overtaxed.

Overall, the efficiency of our energy budget (or lack thereof) is what determines the total health of our entire system. It reflects a holistic view of our body’s energy regulation rather than a simple (over- or under-)reaction to stress.

3. The Social Baseline and the Economy of Action

Well then, how can we ensure maximum efficiency of regulation and our energy budget? By mitigating the associated metabolic costs through social co-regulation or social allostasis.

Social Baseline Theory (SBT) posits that the human brain evolved to expect the presence of others as its default, or ‘baseline’ condition. From this perspective, being with others isn’t an ‘extra’ support but the environmental condition to which the human brain is most adapted.

We optimise our energy expenditure through two primary social mechanisms:

• Risk Distribution. Being in a group probabilistically reduces our personal burden of vigilance (i.e., the costly process of scanning for and anticipating future threats) as environmental risk can be spread across the group.

• Load Sharing. In reliable relationships, our brain assumes that others will use their own resources on our behalf. This allows our brain to ‘outsource’ costly regulatory work to our social network, conserving or even ‘increasing’ its own energy for other vital functions.

4. Attachment as a Predictive Model of Resource Availability

This perspective redefines attachment and its associated internal working models as our brain’s predictive assessment of others’ availability for co-regulation. In other words, attachment represents our brain’s ‘best guess’ regarding the social resources available in its allostatic budget:

• Secure Attachment. A predictive model where social resources are viewed as reliable and available. This certainty allows our brain to ‘blur’ the neural representation of self and other, fully utilising load sharing to minimise the metabolic cost of navigating threats in isolation.

• Insecure-Anxious Attachment. A model reflecting a prediction that co-regulation is unpredictable and only intermittently accessible. This leads to a chronic up-regulation of support-seeking and scanning the environment for available others, because our brain’s best ‘bet’ is to remain hyper-vigilant to maximise the chances of social support.

• Insecure-Avoidant Attachment. A predictive model of others as consistently unavailable for co-regulation. This leads to our brain’s expectation that we must mainly rely on costly individual regulation, which can lead to a depletion of bioenergetic resources over time.

5. The Synergy of Attachment and Resilience: Investing in Bioenergetic Capital

Now then, how exactly do attachment and resilience ‘add up’? Their link lies in metabolic synergy. Attachment is the predictive strategy, while resilience is the metabolic outcome.

When we possess a secure attachment model, our brain identifies others as ‘safety signals’. This prediction fundamentally alters our perception of the world. For example, a hill literally appears less steep when standing next to a good friend. Because our brain predicts that social resources are available for load sharing, it decreases the activation of costly neural computations associated with effortful self-regulation and future need anticipation.

In this way, secure attachment acts as a bioenergetic buffer. It ensures that we don’t have to ‘spend’ our limited energy on regulatory work that others can provide. This synergy creates a surplus of bioenergetic capital, allowing our system to remain flexible and open to new challenges. Resilience ‘adds up’ through secure social bonds because they transform life’s demands from ‘expensive’ individual burdens into ‘distributed’ social tasks.

6. Highest Resilience as Enactive Niche Construction

If we link resilience as the metabolic outcome with attachment as the predictive model to create metabolic synergy, how can we describe different resilience strategies and their relative efficiency? One way to do so is to look at how our brain manages allostasis and particularly its predictive element, also called ‘inference’.

Recent theory suggests that there are two main ways how inference operates as part of allostasis: perceptual versus active. Their main difference lies in whether our brain adjusts its internal model to fit the external world (i.e., perceptual) or whether it adjusts the external world to fit its internal model (i.e., active).

Perceptual inference is the process by which our brain provides the ‘best explanation’ for incoming sensory data. When our brain’s predictions do not match actual sensations, it uses the resulting prediction error to update its current beliefs into more accurate ‘posterior beliefs’. In other words, when a prediction error occurs in perceptual inference, our brain changes its internal model to better reflect reality.

Active inference is a broader, enactive process where our brain selects plans of action that change the environment to match its predictions. Instead of just perceiving any sensory signals, our brain actively samples the world around us to find preferred sensory signals. This involves predicting under which circumstances we will encounter preferred sensory signals and what we should do to realise that prediction. Therefore, when a prediction error occurs in active inference, our brain’s internal model stays the same, but we actively change our external world to fit that internal model.

Based on these two different modes of perceptual versus active inference, we can now define four resilience strategies and categorise them by their efficiency in terms of allostasis:

• Fragile Strategy: Ineffective Prediction and Metabolic Cost. This strategy is marked by low cognitive and affective flexibility, often failing to use active inference to update predictive models. It’s furthermore characterised by sensory over-reactivity that also prevents accurate perceptual inference. As a consequence, allostasis efficiency is low, because ineffective predictions lead to high metabolic load, commonly manifesting as burnout or exhaustion.

• Durable Strategy: Narrow-Niche Exploitation. This strategy relies on the use of precise but rigid models within a narrow context. There’s limited perceptual (and usually not much active) inference because the predictable environment requires little internal adjustment. Therefore, while efficient in the status quo, this strategy results in the accumulation of metabolic load when the narrow context is perturbed, because there is a lack of flexible resources (and active inference) to handle surprise.

• Resilient Strategy: Allostatic Recovery and Elasticity. This strategy employs active inference and therefore affective flexibility to monitor and regulate internal states. The essence of this strategy is elasticity, or the ability of the allostatic system to ‘bounce back’ and consistently return to homeostasis, even after a significant challenge.

• Pro-Entropic Strategy: Allostatic Growth. This strategy represents the highest level of adaptation. Following adversity, our system doesn’t just ‘bounce back’, but it improves its regulatory capacity. To do so, it uses sophisticated active inference to identify a desired future goal and works backwards to select the optimal path towards it. In other words, this strategy proactively engages in ‘enactive niche construction’, changing our external environment to preclude future metabolic cost. Rather than just keep returning to baseline, this strategy achieves allostatic growth, where our regulatory system becomes more risk-tolerant and efficient post-stress.

According to the above, the ultimate goal of resilience is found in the pro-entropic strategy consisting of ‘enactive niche construction’, or our ability to shape our environmental and social surroundings to optimise allostasis. This strategy ensures that we are not merely passive recipients of our circumstances, but that our actions can proactively influence the social resources available to us.

In terms of attachment, the most effective strategy for long-term resilience therefore is to actively build and maintain high-quality, secure social bonds that strengthen our ‘social baseline’. By investing in creating a predictable social network characterised by familiarity and interdependence, our brain can safely predict that its metabolic burdens will be shared and thereby optimise allostasis.

In essence, secure social bonds provide us with the bioenergetic safety we need to be open to change and exploration. Resilience, in its most active and biologically grounded form, is the art of actively constructing a world where shared regulation is the baseline, ensuring our body’s energy can be shared and socially distributed.

7. Why Bioenergetic Connection Matters

The most important message of this integrated perspective is that we are not closed, self-contained systems, but socially distributed organisms.

We should bother with this paradigm because it replaces the ‘myth of the rugged individual’ with the biological reality of the social baseline, i.e., the fact that our brains and bodies are designed to function most efficiently when we are embedded in reliable, secure relationships.

Seeing the world through this lens helps us for (at least) three reasons:

1. It redefines ‘strength’ as enactive niche construction. Resilience isn’t just the internal ‘grit’ of our own regulatory system. It’s our brain’s innate strategy to actively shape a social environment that allows for load sharing. By cultivating secure bonds, we aren’t socially ‘dependent’ in a negative sense. Instead, we’re optimising our energy budget so we can spend that capital on creativity, growth, and health instead of survival vigilance.

2. It eliminates the ‘blind spot’ of dualism, intrinsically linking subjective experiences with bioregulation as inseparable units. Therefore, relationship quality isn't just a ‘lifestyle choice’. Social proximity and security provide us with vital metabolic capital, directly optimising our brain's efficiency and energy budget.

3. It offers a roadmap for sustainable health. Understanding that all regulation is costly allows us to be more strategic about how we mitigate those costs. Rather than trying to ‘fix’ insecure attachment patterns as psychological flaws, we can see them as predictive models that are both adaptive in themselves and that can be updated and optimised through new, more reliable and accessible social experiences.

Ultimately, this shift in perspective shows that we are most resilient not when we manage to bounce back even if we stand alone, but when we have actively constructed a niche where our metabolic burdens can be most efficiently shared. By seeing social connection as a bioenergetic necessity from an attachment perspective, we can move from merely surviving an uncertain world to actively thriving within a predictable social network of our own making.

© 2025-2026 by Pascal Vrticka. All rights reserved.

Scientific References

Barrett, L. F. The theory of constructed emotion: an active inference account of interoception and categorization, Social Cognitive and Affective Neuroscience, Volume 12, Issue 1, January 2017, Pages 1–23. https://doi.org/10.1093/scan/nsw154

Barrett, L. F., Atzil, S., Bliss-Moreau, E., Chanes, L., Gendron, M., Hoemann, K., Katsumi, Y., Kleckner, I. R., Lindquist, K. A., Quigley, K. S., Satpute, A. B., Sennesh, E., Shaffer, C., Theriault, J. E., Tugade, M., & Westlin, C. (2025). The theory of constructed emotion: More than a feeling. Perspectives on Psychological Science, 20(3), 392–420. https://doi.org/10.1177/17456916251319045

Coan, J. A., & Sbarra, D. A. (2015). Social Baseline Theory: The Social Regulation of Risk and Effort. Current opinion in psychology, 1, 87–91. https://doi.org/10.1016/j.copsyc.2014.12.021

Harrison, L. A., Gracias, A. J., Friston, K. J., & Buckwalter, J. G. (2025). Resilience phenotypes derived from an active inference account of allostasis. Frontiers in Behavioral Neuroscience, 19, Article 1524722. https://doi.org/10.3389/fnbeh.2025.1524722

Saxbe, D. E., Beckes, L., Stoycos, S. A., & Coan, J. A. (2020). Social allostasis and social allostatic load: A new model for research in social dynamics, stress, and health. Perspectives on Psychological Science, 15(2), 469–482. https://doi.org/10.1177/1745691619876528

Theriault, J. E., Katsumi, Y., Reimann, H. M., Zhang, J., Deming, P., Dickerson, B. C., Quigley, K. S., & Barrett, L. F. (2025). It’s not the thought that counts: Allostasis at the core of brain function. Neuron, 113(24), 4107–4133. https://doi.org/10.1016/j.neuron.2025.09.028

White, L., Kungl, M., & Vrticka, P. (2023). Charting the social neuroscience of human attachment (SoNeAt). Attachment & Human Development, 25(1), 1–18. https://doi.org/10.1080/14616734.2023.2167777

Zhang, J., Chen, D., Deming, P., Srirangarajan, T., Theriault, J. E., Kragel, P. A., Hartley, L., Lee, K. M., McVeigh, K., Wager, T. D., Wald, L. L., Satpute, A. B., Quigley, K. S., Whitfield-Gabrieli, S., Barrett, L. F., & Bianciardi, M. (2025). Cortical and subcortical mapping of the human allostatic-interoceptive system using 7 Tesla fMRI. Nature neuroscience, 28(11), 2380–2391. https://doi.org/10.1038/s41593-025-02087-x

My seventh attachment insights post was about Polyvagal Theory (PVT). I briefly illustrated that the latest scientific evidence doesn’t support it and which implications this conclusion has for our current understanding of attachment and trauma.

Quite a few things have happened since then. Most importantly, on 09 February 2026, the journal Clinical Neuropsychiatry published two new articles on PVT.

The first article was written by thirty-nine experts in the areas of physiology, evolution of the vagus nerve and vertebrate social behavior. The experts were asked to evaluate PVT in general and comment upon an overview of PVT written by Stephen Porges in the same journal in 2025 more specifically. The second article was written by Porges himself as a response to the first article.

The discussion between the experts and Stephen Porges sparked a lot of interest, including mine. In this post, I would therefore like to summarise the content of this discussion, reflect upon its implications — also for our current understanding of attachment and trauma — and suggest an alternative account that can be further refined through mutual and constructive dialogue across domains. The structure is as follows:

1. What Is Polyvagal Theory (PVT)?

2. What Are The Experts’ Main Points Of Critique Of Polyvagal Theory (PVT)?

3. What Is Porges’ Response To This Critique?

4. What Are The General Implications Of This Discussion?

5. What Does This Mean For Attachment And Trauma?

6. Where Do We Go From Here? An Alternative Account

1. What Is Polyvagal Theory (PVT)?

Porges (2025) describes Polyvagal Theory (PVT) as an evolutionarily informed neurophysiological framework. It aims at explaining how the human nervous system controls our ability to form social bonds as part of co-regulation, as well as how it handles stress and threat as an effective way of responding to the external world.

At the core of PVT is the “vagal ladder”, a hierarchical organisation of autonomic states governed by the Jacksonian principle of dissolution. This principle proposes that during a challenge, the nervous system recruits defensive states in reverse phylogenetic order (i.e., in opposite order of their evolutionary emergence).

The top step of this ladder is referred to as the “ventral vagal complex” (VVC), which is described to originate in the brainstem’s Nucleus Ambiguus and to utilise fast-acting, myelinated fibres to support social communication and physiological flexibility. The VVC is integrated with the so-called “social engagement system”, a circuit of cranial nerves that automatically and non-consciously coordinates facial expression, vocal prosody and ingestion with heart rate regulation to facilitate co-regulation.

The middle step of the “vagal ladder” is the sympathetic nervous system (SNS) that provides the metabolic energy necessary for mobilisation, commonly known as the “fight-or-flight” response.

And finally, at the bottom step of the “vagal ladder” is the “dorsal vagal complex” (DVC) that is proposed to originate in the dorsal motor nucleus (DMNX). The DVC is described to mediate a state of “shutdown” or immobilisation aimed at metabolic conservation during life-threatening situations where escape is impossible.

PVT proposes that transitions between these three autonomic states of the “vagal ladder” are governed by “neuroception”. This process is described as a subconscious neural mechanism that evaluates environmental cues for safety, danger or life-threat, and automatically adjusts the body’s physiological response independently of conscious or cognitive appraisal. When “neuroception” detects a significant threat, it is said to initiate the process of “dissolution”, which is the sequential withdrawal of the “social engagement system” to allow “older” defensive circuits (i.e., first the SNS and then the “dorsal vagus”) to dominate the body’s response.

To quantify these states, PVT relies on respiratory sinus arrhythmia (RSA), i.e., the natural fluctuation of heart rate during breathing. Porges defines RSA as a pathway-specific, centrally gated index of “ventral vagal influence” originating in the Nucleus Ambiguus. PVT furthermore assumes that RSA is governed by a common cardiopulmonary oscillator in the brainstem that coordinates breathing rhythms with heart rate. The efficiency of this system is to be assessed through derived metrics like “weighted coherence” and “vagal efficiency”, which are said to measure the functional coordination and regulatory capacity of these brainstem circuits.

PVT’s assumptions are based on a hypothesised “ventral migration” of cardioinhibitory neurons from the DMNX to the Nucleus Ambiguus during mammalian evolution, an evolutionary trajectory that Porges claims to be echoed in human embryological development.

Finally, PVT utilises transcriptomic evidence, asserting that neurons in the mammalian Nucleus Ambiguus express a unique suite of genes associated with myelination (Mbp, Myrf) and neuropeptide signaling (Snap25). Porges claims that these molecular signatures distinguish the unique mammalian “social engagement” system from non-mammalian and “non-social” vertebrates.

2. What Are The Experts’ Main Points Of Critique Of Polyvagal Theory (PVT)?

The 2026 evaluation by thirty-nine experts in the areas of physiology, evolution of the vagus nerve and vertebrate social behavior concludes that Polyvagal Theory (PVT) is scientifically untenable. Because its fundamental neurobiological premises are inconsistent with the broader evidence base.

A primary point of critique involves PVT’s reliance on Respiratory Sinus Arrhythmia (RSA) as a direct and reliable measure of central vagal drive from the brainstem’s Nucleus Ambiguus. The experts argue that RSA is an inherently peripheral measure that is significantly confounded by non-vagal factors (including age, sympathetic tone, respiratory rate and depth, and the mechanical stretching of the heart’s sinoatrial node during the act of breathing). Consequently, RSA doesn’t provide a direct window into the central nervous system or a consistently accurate index of “cardiac vagal tone”, which is the sustained influence of the vagus on the average heart rate.

This RSA measurement problem directly impacts the validity of the “vagal ladder”, the hierarchical model that PVT uses to map autonomic states to psychological experiences. PVT asserts a functional split where the “ventral vagal” system supports prosocial behavior but the “dorsal vagal” system mediates defensive “shutdown”, immobilisation and dissociation. However, scientific consensus demonstrates that the Nucleus Ambiguus is predominantly involved in all significant heart-rate reactions, including both small adjustments and abrupt, large slowing. In contrast, the “dorsal vagal” system in mammals primarily regulates the gastrointestinal tract and has very limited influence on heart rate control.

Furthermore, the experts describe the concept of “dorsal vagal shutdown” in humans as a “mental fiction”, because clinical research into emotional freezing and dissociation shows no characteristic massive or lethal heart-rate decreases. Rather, heart rate typically remains stable or shows only minor decreases related to the physical immobility of the state.

The evolutionary basis of PVT is also heavily critiqued for its “mammalian bias” and inaccurate portrayal of vertebrate history. PVT claims that mammals uniquely evolved fast-acting, myelinated vagal fibres to support “social engagement”. Yet, decades of research show that these myelinated cardioinhibitory fibres are widespread across vertebrates, including sharks, fish, amphibians, and reptiles. Similarly, the assumption that non-mammals lack complex sociality is refuted by evidence of long-term pair-bonding, monogamy and communal parenting in reptiles and birds. Overall, such findings suggest that social behavior is an ancient trait non-dependent on the specific vagal architecture proposed by Porges.

The experts also reject the theory’s use of the biogenetic law, i.e., the idea that developmental stages recapitulate evolutionary history. They notice that this hypothesis has been soundly dismissed by modern embryological research. While the psychological concepts utilized by PVT, like co-regulation and psychological safety, are considered valid and valuable, they predate PVT by decades and are neither clarified nor strengthened by this physiological framework that lacks empirical support.

Finally, the experts identify significant bibliographic and formal errors in Porges’ 2025 paper. They state that half of the physiological references provided are inaccurately cited regarding authorship or journal attribution. Most notably, the experts point out that a primary source often used to support the theory’s evolutionary claims (i.e., Richter & Spyer, 1990) doesn’t actually exist in the scientific literature. Beyond these formal mistakes, the experts highlight interpretative inaccuracies where Porges allegedly misrepresents the findings of the research he cites. They note that he utilised transcriptomic studies focused solely on mice and rats to claim unique differences between mammals and reptiles, despite the original authors making no such comparison. Furthermore, the experts identify a factual contradiction within Porges’ own body of work. While he claims that his “weighted coherence” RSA metric remains stable across varying levels of blockade, the experts point out that his original 1986 data showed a clear linear increase.

3. What Is Porges’ Response To This Critique?

Porges (2026) responds to the expert critique by characterising it as a series of “structural misrepresentations” and “category errors” that evaluate a “reconstructed proxy” of the theory rather than its actual propositions.

He argues that the critique functions as a “straw-man argument” by attributing claims to the theory (like strict anatomical exclusivity or the use of recapitulationist logic) that he has explicitly disavowed in the literature for over two decades.

Porges furthermore asserts that the critics fundamentally conflate “neuroanatomy” with “neurophysiology”, using the physical presence of anatomical structures in non-mammals to incorrectly attempt to “falsify” functional claims regarding mammalian-specific circuit integration and regulatory organisation.

Regarding the specific critique of RSA, Porges maintains that the disagreement reflects a “difference in measurement preference” and a “reduction of theory to measurement”, rather than empirical evidence against the model’s organising principles. He dismisses the experts’ claim of scientific consensus, describing their document as a “position paper” by a “self-selected group” that lacks the formal adjudicative criteria or professional society endorsement required to represent a true disciplinary rejection.

Furthermore, Porges characterises the critique as suffering from a “persistent failure of representational uptake”, noting in a historical audit that many of the current objections were “explicitly corrected” in peer-reviewed responses as early as 2007.

Instead of addressing the experts’ individual objections one at a time (including the obvious bibliographic and formal errors, interpretative inaccuracies and factual contradictions), Porges challenges them to engage with the theory “on its own terms” by testing its specified “conditions of falsifiability”. He maintains that the critique fails to specify any empirical conditions under which his actual predictions, like the hierarchical withdrawal of “ventral vagal” activity during threat, would fail.

Ultimately, Porges concludes that the charge of “untenability” applies to the critique itself due to its failure to satisfy the epistemic standards necessary for scientific refutation and its neglect of the systems-level architecture foundational to the theory.

4. What Are The General Implications Of This Discussion?

In my opinion, there are three main general implications of this discussion.

4.1 Scientific Accuracy And Integrity

The first implication is that the bibliographic and formal errors as well as interpretative inaccuracies in Porges’ 2025 paper should be thoroughly reassessed and amended as part of a correction.

As I have written before in my post on “hallucinated citations and phantom references”, scientific accuracy and integrity are key in any scientific publication. This includes the provision of real and correct references that appropriately reflect the cited papers’ findings and conclusions.

4.2 Current Neurophysiology And Associated Risks

The second implication is that PVT’s core assumptions are fundamentally contradicted by current physiological, neurobiological and evolutionary data. A theory that relies on implausible explanations and whose primary assumptions regarding the vagus nerve and social behavior aren’t backed up by current evidence indeed is untenable.

There even is substantial and tangible risk of promoting inaccurate information about bodily functions underlying lived experience to healthcare providers, therapists, patients, clients and many more. This issue particularly pertains to the concepts of “neuroception” and the hierarchical “vagal ladder”. Because these constructs are strongly reductionist by collapsing complex multisystemic conditions into a binary of “vagal tone” or “vagal state”, they ignore the contributions of many other physiological and neurobiological systems. The result is a form of biological determinism that can lead patients and clients to view their nervous systems as “broken” and therefore beyond the reach of intervention. Furthermore, there is a clinical risk that therapists and healthcare providers may prematurely attribute systemic physical symptoms to a “trauma-informed” autonomic state, potentially delaying necessary medical investigations and overlooking life-altering diagnoses.

4.3 Remaining Validity Of Established Psychological Constructs
And the third implication is related to PVT’s retroactive attribution of established psychological constructs (i.e., psychological safety, “social engagement”, attachment-based co-regulation and dissociation) and observable fight-or-flight and freeze behaviours to a contested neurobiological framework. Because these constructs and behaviours were robustly defined and empirically tested by pioneers like Maslow, Rogers, Bowlby, and Brazelton long before PVT’s inception, linking them to PVT’s flawed autonomic and hierarchical evolutionary assumptions provides no additional heuristic or explanatory power.

The third implication yields a crucial realisation: the empirical invalidity of PVT’s neurophysiological framework doesn’t negate the existence of the phenomena it attempts to explain. The states of psychological safety, “social engagement”, attachment-based co-regulation and dissociation, as well as the behaviors of fight-or-flight and freeze remain robust and valid realities. They exist independently of the flawed “neuroception” and “vagal ladder” approximations. It’s merely the specific physiological and neurobiological narrative proposed by PVT that is scientifically untenable.

5. What Does This Mean For Attachment And Trauma?

As I have already pointed out in my original attachment insights post, attachment and trauma involve complex physiological processes and behaviours maintained by extended and interconnected neurobiological networks, not just the ANS or particular branches of it. And these processes often emerge as meaningful and adaptive responses to particular environmental demands.

What is more, it’s important to remember that there is no simple one-to-one mapping of distinct nervous system states to corresponding psychological and emotional states. For example, if insecure attachment or trauma yield a state of persistent and unresolvable threat appraisal, this doesn’t equate to a non-conscious and automatic “neuroception” of danger. Instead, it’s our brain’s active perception and constructive and predictive evaluation of our current physiology (e.g., increased heart rate, sweating), which is based upon previous experiences and predictions about the future, that yields such an interpretation.

In other words, the perception of threat associated with attachment insecurity or trauma isn’t a simple and intrinsic component of the physiological response in itself, but the result of a series of complex computations including the physiological response as one component amongst several others.

For such complex computations, “a truly integrated and accurate perspective of human regulation must consider a much broader network of systems and structures that are often overlooked” (beyond the ANS and or particular branches of it), as nicely pointed out by Manzotti and colleagues (2024). This includes a range of subcortical and cortical brain regions and networks, the endocrine and immune systems, the enteric nervous system, as well as the interoceptive and somatosensory systems. The same logic applies to the underlying physiology and neurobiology of insecure and disrupted/disorganised attachment as well as attachment trauma.

6. Where Do We Go From Here? An Alternative Account

I personally believe that a constructive and mutually beneficial dialogue is essential for establishing a viable way forward. For many, PVT offered a first description of an integration of ANS responses with the fields of clinical psychology and psychopathology, allowing them to reconsider “safety” as a neurophysiological state rather than just an external condition. Therefore, as valid and important it is to point out PVT’s obvious flaws and risks, there also must be a clear and feasible alternative future path that acknowledges the various needs and concerns of everybody involved.

The above said, my professional qualifications as a social neuroscientist with a special focus on human attachment necessarily limit my scope within this process. I am only in a position to speak to a combined attachment theory and neuroscisnce account that could replace PVT’s flawed “neuroception” and “vagal ladder” approximations. My account’s usefulness and viability of its practical implementation need to be evaluated by others, including healthcare providers, therapists, patients, clients and many more, ideally as part of an extended and constructive dialogue.

At the core of my account are three current neuroscientific principles: constructed emotion, free energy as well as allostasis / social allostasis, which I link to insights from attachment theory and science. In a nutshell, my account goes as follows.

6.1 Allostasis: The “Body Budget”
The brain’s primary objective is allostasis, i.e., maintaining stability despite important challenges by anticipating physiological needs before they arise. It manages a “body budget”, strategically allocating resources (like glucose and oxygen) based on predicted demand. Unlike homeostasis, which is a primarily automatic reaction to small imbalances, allostasis is an active investment strategy designed to ensure survival through metabolic efficiency.

6.2 Interoception And Constructed Emotion
We experience this budgeting process through interoception, the brain’s perception of internal sensations. According to the theory of constructed emotion, the brain receives raw sensory data and categorises it using past experiences and predictions of the future. For example, an accelerated heart rate is not inherently “fear”. Instead, the brain constructs “fear” by predicting that a similar past experience that was linked to threat will require a metabolic surge in the immediate future. In short, emotions are functional inferences that give meaning to physiological shifts by linking them to past experiences with the aim of guiding the next allostatic move.

6.3 Prediction And Free Energy
The brain operates as an inference engine, constantly generating top-down predictions to explain incoming sensory signals. Following the free energy principle, biological systems seek to minimize “free energy”, i.e., the mathematical equivalent of prediction error or the gap between expectation and reality. When prediction errors persist, the brain must either update its internal model through learning or alter the body’s state to match the prediction through action.

6.4 A Predictive Reinterpretation of PVT

Through this lens, the states described by PVT are not part of “neuroception” on different steps of the “vagal ladder”, but distinct phases of allostatic management.

• “Social Engagement” (“ventral vagal”): This is social allostasis. Trusted others provide predictability, reducing the free energy required to monitor the environment. We feel safe because readily available and helpful social resources lower the individual metabolic cost of self-regulation through co-regulation.

• Fight-or-Flight (SNS): This is a high-cost allostatic investment that temporarily decreases more expensive long-term projects (like digestion, immune function and reproduction). The brain predicts a need for intense movement and dispatches resources, categorising the resulting high arousal as a survival emotion that narrows the focus to the immediate present and the perceived threat. Crucially, there are several different ways of threat regulation, including what we call “social flight” associated with attachment anxiety that involves intensified proximity-seeking and thus pro-social tendencies (see below).

• “Shutdown”/Freeze (“dorsal vagal”): Rather than an “ancient circuit” at the bottom of the “vagal ladder”, this is metabolic bankruptcy. When prediction errors are chronic and the environment is perceived as too chaotic or inescapable, the brain stops engaging in prediction and investing energy. In other words, by attempting to minimise further energy loss, the brain restricts movement and reduces metabolic expenditure.

6.5 Application To Attachment

In the case of insecure attachment, in its attempt to minimise free energy, the brain prioritises models of the world that have historically ensured survival. Even if those models are metabolically more expensive and/or less efficient in the present.

This involves hyper-vigilant allostatic investment in the case of attachment anxiety (i.e., spending constant energy on proximity-seeking and monitoring as part of “social flight”, as the cost of missing a co-regulation opportunity is deemed too high). Conversely, in the case of attachment avoidance, the resulting strategy is a shift from social allostasis to individual allostasis by means of “individual fight” (i.e., increasing the available energy for self-regulation by minimising metabolic cost of anticipated failed bids for connection). I have previously written about it here.

6.6 Application To Trauma

When it comes to trauma, this account doesn’t see it as a lingering “primitive” reflex. Instead, it conceptualises trauma as an impairment of predictive flexibility.

When an individual experiences a profound threat that they cannot resolve, the resulting prediction error is catastrophic. The brain assigns high “precision” or importance to the sensory data associated with the trauma. It predicts that a similar threat is always imminent. As a consequence, the brain anticipates constant danger and maintains a permanent state of high allostatic investment (e.g., elevated cortisol, heart rate and muscle tension). Furthermore, an otherwise neutral stimulus (like a specific smell or sound) readily triggers a prediction of danger. The brain, attempting to minimise free energy, “corrects” the current sensory input to match its traumatic model and constructs a full-body experience of fear. As part of our NAMDA, we refer to this state as “hyper-arousal”.

The opposite of the state of “hyper-arousal” is immobilisation or freeze. While PVT describes it as a “dorsal vagal” state, the predictive model views it as metabolic conservation following prolonged allostatic demand. If a traumatised brain chronically predicts that no action can resolve a threat (i.e., “learned helplessness”), it ceases to spend resources on threat regulation altogether. This results in the dissociative, numb, and low-energy state of trauma. Not because a specific nerve was “triggered”, but because the “body budget” is in total deficit and the brain has predicted that any future effort is futile. As part of our NAMDA, we refer to this state as “hypo-arousal”.

6.7 The Goal Of Resolution

According to this account, working with insecure attachment and healing trauma would mainly involve updating the prediction model.

This may be achieved by providing a stable and safe environment where the brain can experience prediction errors without negative consequences. It can then realise that a current sensation isn’t the same as a response to a past threat, and slowly re-budget its metabolic resources. Potential avenues — to be further discussed and practically tested — could include the following:

• Relational Predictability: Maintain consistency and predictability in social interactions to provide the brain with a “safe prior”. This reduces the metabolic cost of monitoring for uncertainty or planning for extensive self-regulation.

• Interoceptive Micro-Dosing: Gently attend to small internal sensations without the anticipated catastrophe to reduce the precision of survival-related prediction errors.

• Sensory Scaffolding: Create a sensory-stable and safe external environment (e.g., weighted blankets, specific scents, low lighting, etc.) or internal environment by using rhythmic breathwork. This allows the brain to pre-emptively lower the body budget’s “defence tax” based on predictable cues of safety.

• Semantic Reframing: In moments of high arousal, instead of saying “I am terrified” it may be better to say “I feel agitated because my heart is beating fast, and my lungs feel tight”. By labelling raw physical sensations rather than constructed emotions and increasing emotional granularity, the brain can be prevented from launching an unspecific and costly metabolic defence.

  ---

  © 2025-2026 by Pascal Vrticka. All rights reserved.

  ---

Leave a comment

A note on my position and involvement in this discussion

I offered comments on an earlier version of Grossman and colleagues’ expert evaluation of Polyvagal Theory, and I am thanked in its acknowledgments. I am, however, not one of its authors, I had no part in shaping its conclusions, and I have no financial stake in this debate. I set this out here so that my position is entirely transparent.

On the science, I agree with the critics: the neurobiology and the evolutionary story at the core of Polyvagal Theory are outdated and, in several places, simply wrong. And, in my opinion, the expert evaluation is right in concluding that the theory is “untenable” as a neurophysiological account. I also agree with what that implies. In its concluding points the evaluation argues that the psychological constructs the theory leans on (such as psychological safety, social engagement, co-regulation, emotional freezing and dissociation) were described long before it, derive from earlier attachment theory, psychotherapy and trauma research, and stand perfectly well on their own, so that stripping the theory away takes nothing real with it.

The expert evaluation also notes that better frameworks already exist and urges clinicians to consider them, yet it says in its abstract that it does not itself set out to build one. That is precisely the step I took in this post (after discussing both the expert evaluation and Porges’ response to it), and it is entirely my own: to give those constructs, and the subjective experiences they name, a voice, and to suggest an alternative account. An account grounded in my own understanding of social allostasis, active inference, constructed emotion and attachment science as a social neuroscientist. An account that aims at updating the neurophysiology without calling the constructs or the experiences themselves into question. Where anything in this post reads as a criticism of the expert evaluation, it is only of how far it chose to go, and never of its science.

For me as a researcher who regularly publishes academic papers and explains the latest findings and their implications to a broad audience, scientific accuracy and integrity are key.

In my accounts of the research procedures and observed results that I am describing, regardless of whether they are my own or somebody else’s, I always aim at being as factual and transparent as possible.

The best and most direct approach for ensuring scientific accuracy and integrity when writing is appropriate citation and referencing. Whenever a scientific statement is made, the full bibliographic information to the academic sources must be included.

There are several ways of doing so. A common way is to provide a link in the text that readers can click on, taking them directly to the website of the underlying study. Another way is to include a short citation - i.e., name of author(s) & year of publication - in the text and mention the full bibliographic information of the underlying study as a reference in a section at the end or in a footnote.

By providing appropriate citations and references, readers can verify the academic sources, i.e., not only see that they actually exist, but also that the underlying research really examined and found what is being claimed it did. They can also form their own and independent impression about the suggested interpretation of the scientific evidence.

Accordingly, when reading published academic papers and explanations of the latest findings and their implications for a broad audience, the same principles should be followed. In other words, readers should verify that this information also has scientific accuracy and integrity, i.e., that it provides appropriate citations and references.

Unfortunately, particularly when it comes to explanations of the latest scientific findings and their implications for a broad audience, I have recently observed a strong increase in ‘hallucinated citations’ and ‘phantom references’ most likely linked to the widespread use of general-purpose chatbots (like ChatGPT, Gemini, etc.) to generate content.

What are hallucinated citations and phantom references and how are they generated by AI?

‘Hallucinated citations’ and ‘phantom references’ are citations and references invented by Large Language Models (LLMs) that look real but don’t exist. They’re a creative combination of author names, paper titles, journal names and publication years that mimic genuine publications but in reality are completely fictional.

Most current LMMs are optimized to be agreeable and fluent, not to be fact-checking engines. When users ask for ‘scientific evidence’, these models often treat it as a creative writing prompt using ‘scientific-sounding’ components rather than a strict database query. They prioritise ‘looking right’ over ‘being right’.

What’s more, for non-expert content creators and readers alike, ‘hallucinated citations’ and ‘phantom references’ can easily look genuine and credible. Because LLMs create them based on real publications, they usually contain the names of real academics who previously published in the respective research areas, plausibly sounding paper titles and real academic journal names.

However, the specific combination of these elements is completely made up - authors who never published together are merged, paper titles are altered by either removing or adding words and real journal names are randomly picked and added to this cocktail.

To give an example, I can refer to an instance in which - as far as I can tell - a chatbot created a ‘hallucinated citation’ and ‘phantom reference’ of my own work. Back in 2024, an academic colleague made me aware of a published paper that comprised a citation and reference to one of my studies that read as follows:

Vrtička et al. (2019) - Vrtička, P., Bondolfi, G., and Sander, D. (2019). The neural substrates of social emotion regulation: A fMRI study on emotion reappraisal and suppression in individuals with and without social anxiety disorder. Frontiers in Human Neuroscience, 13.

At first glance, this reference looked familiar, even to me. However, upon closer inspection, I realised that it was flawed in several of its elements.

First, although I did publish with Bondolfi and Sander before, I never did so in a paper with only these three authors. Second, I did also publish a paper before that had ‘Neural substrates of social emotion regulation: A fMRI study’ in its title, but never combined with ‘on emotion reappraisal and suppression in individuals with and without social anxiety disorder’. And third, while I did publish before in the journal Frontiers in Human Neuroscience, I never did so in Volume 13 in 2019 (only in Volume 6 in 2012).

The two closest real references to my own work that I did publish, are:

• Vrticka et al. (2013) - Vrticka, P., Simioni, S., Fornari, E., Schluep, M., Vuilleumier, P., & Sander, D. (2013). Neural substrates of social emotion regulation: A fMRI study on imitation and expressive suppression to dynamic facial signals. Frontiers in Psychology, 4, Article 95. https://doi.org/10.3389/fpsyg.2013.00095

• Vrtička et al. (2012) - Vrtička, P., Bondolfi, G., Sander, D., & Vuilleumier, P. (2012). The Neural Substrates of Social Emotion Perception and Regulation are modulated by Adult Attachment Style. Social Neuroscience, 7(5):473-93. https://doi.org/10.1080/17470919.2011.647410

It seems highly unlikely to me that such a specific combination of elements emerged as part of a simple error or oversight. The initial citation and reference to the non-existent paper have since been removed as part of a correction and replaced by the Vrtička et al. (2012) reference listed above (second bullet point).

What can be done to restore scientific accuracy and integrity when citing and referencing?

I’m encountering ‘hallucinated citations’ and ‘phantom references’ more and more often these days. They’re basically popping up everywhere. And I personally feel that this is a real problem that systematically undermines scientific accuracy and integrity.

If a scientific claim is made, it must be verifiable. It’s associated citation and reference not only must exist but the research and results described in the cited and referenced work must closely reflect what is reported as factual information in the claim.

Well then, what can be done to avoid generating more ‘hallucinated citations’ and ‘phantom references’ and to more readily spot them in already existent content?

On the one hand, when creating content, I think that the first and foremost principle should be to not use general-purpose chatbots (like ChatGPT, Gemini, etc.) as a primary source for scientific evidence, especially in a generative way. These AI tools can be sensitively used for explaining complex concepts, summarizing scientific sources (if full texts are uploaded to specific AI tools in which context is limited to the sources provided) as well as drafting and brainstorming.

However, when using them as a primary source for scientific evidence in a generative way, the the risk of obtaining ‘hallucinated citations’ and ‘phantom references’ is just too strong of a liability. Especially because this liability is usually paired with AI’s aim to be conversational and ‘smooth’ and therefore to provide ‘polished’, oversimplified and/or factually incorrect accounts.

On the other hand, to promote scientific accuracy and transparency when consuming content, full citations and references should always be actively looked for. If they are absent or appear incomplete or inconsistent, more information should be requested from content creators. And more generally, any, and especially strong scientific claims should be regarded with great caution and always independently verified.

The tricky bit when doing so is that the devil lies in the details. Very often, ‘hallucinated citations’ and ‘phantom references’ look just fine at a first glance. As content consumers, close attention and sustained awareness are required to spot the sometimes subtle inconsistencies. And when creating content, it nowadays takes a conscious effort to resist the temptation of quantity over quality. Creators should always remember that scientific accuracy and integrity (i.e., properly researching and reporting scientific sources) are both a necessity and an ethical mandate.

Leave a comment

Here on Substack, I am now revisiting some of my earlier attachment insights posts. I am adding more details and depth, and I am checking whether my initial conclusions still hold.

My first re-examination was about the confusion surrounding ‘adult insecure-disorganised attachment’ — Lost in Translation: Disorganised Attachment. Make sure to have a look if you haven’t read it yet.

This time, as part of my second re-examination, I am focussing on dads’ role in attachment theory and science.

For decades, attachment theory and science treated mothers as the sun around which the infant world revolved. This was especially true regarding the formation and maintenance of deep, formative child-parent bonds and its consequences on child developmental outcomes. However, things were about to change, and these changes were to stay for good.

A quick journey through time

What happened to the concept of dads as attachment figures from the very origins of attachment theory in the 1950s all the way to the 2000s? A very nice summary of many of the important questions and developments can be found in a summary paper written by Inge Bretherton in 2010 — Fathers in attachment theory and research: a review.

Phase 0: The early beginnings
Bowlby first tentatively formulated attachment theory in 1958. At the time, he proposed an evolutionary explanation for the infant–mother bond that set him apart from his psychoanalytic colleagues — who argued that a child loves their mother only because she provides food (oral gratification).

In doing so, Bowlby described attachment as an adaptive, biological mechanism ensuring that a vulnerable individual stayed close to a stronger and differentiated person for protection. And in his early view, this stronger and differentiated person was the mother (or ‘mothering person’).

While this early view didn’t completely exclude other attachment figures like fathers, its language and concepts overwhelmingly focussed on the maternal role.

Phase 1: The father is possible, but secondary
During the 1960s, attachment researchers for the first time investigated whether fathers could serve as attachment figures for their children, and if so, what their relative ‘rank’ was to mothers. This happened as part of Mary Ainsworth’s early naturalistic home-observation studies in Uganda as well as Schaffer and Emerson’s research in Scotland.

Findings from this set of studies confirmed that infants readily attached to fathers (and other family members), challenging the strict ‘mother-only’ stereotype. As a consequence, Bowlby slightly revised his initial view but still noted that, while there may be other attachment figures than the mother, they are much less important and lower in the attachment hierarchy.

Bowlby also argued that researchers failed to distinguish between fathers as true and principal attachment figures versus preferred playmates. He said that children seek playmates (including fathers) when they are in good spirits and know where their attachment figure can be found, but seek their attachment figure (the mother) when they are tired, hungry, ill, alarmed or uncertain of the figure’s whereabouts.

Phase 2: In search for the tip of attachment hierarchy
Phase 2 coincided with the development of the Strange Situation Procedure (SSP), the first observational research method to classify infants’ attachment within the laboratory, at the end of the 1960s.

At that time, most mothers weren’t employed outside the home and could thus more readily be asked to come to the lab with their child while fathers were at work. Therefore, infant-father interactions as part of the SSP were simply not assessed due to practical reasons. And once available, the selective infant-mother SSP findings led to a strong upsurge of infant–mother attachment studies during the 1970s. Infant-father attachment once more seemed to get overlooked.

It took the work of Michael Lamb and colleagues to shift the focus back to dads. They observed that infants approached both parents equally for contact in non-stress environments. Yet, when a stranger was introduced or stress heightened, infants “dramatically shifted” their attachment behaviours toward the mother. These findings confirmed that mothers retained prominence as principal figures under stress.

However, Lamb and colleagues also suggested that infant-father and infant-mother relationships may provide infants with different kinds of experiences, which in turn may lead to differential developmental consequences. But this idea was not followed up immediately. Instead, another line of investigation gained momentum.

Phase 3: The quality check and the independent bond
During the 1980s and early 1990s, phase 3 focussed on the comparative quality of infant-father and infant-mother attachments and the independent intergenerational transmission of attachment from mothers and fathers to infants.

Two developments as part of this phase are noteworthy.

First, a series of studies was conducted to specifically look at infant-father versus infant-mother attachment during the SSP. Findings demonstrated that attachment quality measured via the SSP was quite often distinct for each parent, with roughly 40% of infants showing non-concordant patterns — i.e., they were secure with one parent but insecure with the other parent. In other words, the SSP showed that quite often, the quality of the attachment bond an infant forms with their mum and dad is different.

Second, with the development of the Adult Attachment Interview (AAI), it became possible to not only classify parental attachment representations as such, but to study whether parental attachment representations correspond to infants’ attachment classifications, and thus to study intergenerational attachment transmission. The conducted studies confirmed that both maternal and paternal attachment representations independently predicted their child’s attachment classifications, especially in the case of attachment security. In other words, parents classified as ‘secure-autonomous’ in the AAI were more likely to have infants classified as secure in the SSP. And although the associations were weaker for fathers, all were “highly significant”.

Nonetheless, the high rate of non-concordant child SSP classifications with regards to mum versus dad raised complex questions about how these distinct patterns arise and how they affect infant development.

Phase 4: Fathers foster secure exploration
Finally, during the late 1990s and early 2000s, phase 4 stated to ask nuanced questions about whether infant-father and infant-mother attachment outcomes differed and whether standard assessments like the SSP were equally appropriate for both parents.

Longitudinal research demonstrated that early father-related measures predicted adolescent and young adult outcomes, sometimes differently from maternal predictors. For example, the ‘Bielefeld study’ found that fathers’ sensitive and challenging play was a powerful predictor of children’s later AAI security during adulthood.

These findings reinforced the view that mothers might primarily foster secure attachment while fathers uniquely foster secure exploration, necessitating a move toward a family systems approach that considers both the distinct and joint influences of infant attachment to both parents.

Current considerations

Building upon phase 4, modern attachment science keeps expanding and refining the notion that both mums and dads are important in raising children and setting them up for optimal development. Importantly, to derive such current evidence, researchers can nowadays pool data from many individual studies with hundreds or even thousands of children and parents as participants through meta-analyses. What emerges from these efforts is a slightly different picture, though, which emphasises similarities rather than differences.

Secure attachment to both parents boosts children’s healthy development
In two separate meta-analyses summarised here, researchers assessed whether children’s attachment to mother and father predicted children’s mental health and language competence. The results showed that children who simultaneously had secure attachment relationships with both their mum and dad were likely to experience fewer symptoms of anxiety and depression and to exhibit better language skills (than kids with one or no secure attachment relationships within their intact, two-parent families).

It’s also important to note what the researchers didn’t find. There was no hierarchy of attachment, i.e., a difference in the importance in terms of which parent a child developed a secure attachment with. Children with secure attachment only to mothers (but not to fathers) and children with secure attachment only to fathers (but not to mothers) were not statistically different in their mental health and language competence outcomes.

A similar picture emerged from two other meta-analyses looking at child-parent attachment and child behavioural problems (i.e., internalising and externalising behaviour) and child pro-sociality. Neither of these investigations found any systematic differences based on parent biological sex.

One important takeaway from these studies is the following: Mums and dads are equally important in raising children and setting them up for optimal development. In other words, it’s the number of secure attachment relationships a child develops within the family network that matters, not the specific biological sex of the adult with whom a secure relationship is developed.

Not everything is the same, but more similar than different
In two additional meta-analyses, researchers looked at how sensitively mums and dads behave towards their children.

Sensitive parents pay attention, interpret, and respond to children’s signals in an appropriate and timely manner, and take children’s perspective when considering children’s needs. Extensive previous meta-analytical work has shown that maternal sensitivity is robustly related to children’s secure attachment, language development and social functioning. But most studies as part of these earlier meta-analyses only included samples of mothers.

The results of the two current meta-analyses showed that mothers generally display slightly higher levels of parental sensitivity than fathers, but that this difference is small enough (or even absent in European samples) to have little practical importance. Crucially, this small parental sensitivity gap has become even smaller in more recent studies, suggesting parental roles have been converging over time.

The current meta-analyses also confirmed that a parent’s sensitivity level is positively connected to their partner’s, supporting the idea that sensitive dads tend to be partnered with sensitive mums, and vice versa.

This last finding has important implications. For example, it raises awareness for family support programs to stop focusing only on mothers and instead including both parents to create positive ‘spill over effects’ across the entire family unit. Such a system-level approach is vital because dads, like mums, are fully capable of sensitive parenting. Furthermore, more policies like generous parental leave, as seen in European countries where the parental sensitivity gap is smallest, should be implemented because they are known to be effective in supporting greater, more sensitive involvement from dads over time.

Summary and outlook

Over the last 70 years, attachment theory and science saw an important transition in how they consider dads. Initially being invisible, fathers’ role evolved from being a playmate and ‘secondary parent’ to a capable, sensitive and equal attachment figure fulfilling a vital role for child development.

The next logical and necessary step is to act upon the insights gained from this transition and the current scientific evidence to adjust policy and early intervention efforts. Crucially, this is not about prioritising dads over mums. On the contrary, it’s about strengthening children’s attachment networks as a whole, including dads alongside mums, with beneficial effects for everybody involved.

More research should also be done on both different and larger attachment networks comprising same-sex parents and non-parental caregivers like grandma and grandpa as well as parents’ attachment figures. Raising children takes an entire village, and no member of this extended community should be forgotten.

© 2025-2026 by Pascal Vrticka. All rights reserved.

Leave a comment

It’s already been a while since the COVID-19 pandemic. But the profound impacts it had on all our work and living arrangements can still be felt.

Amongst these impacts, one particularly stood out for dads. During the nationwide lockdowns, fathers ditched the office and came home in their millions. The UK Fatherhood Institute realised this was a unique ‘social experiment’ and conducted research into “lockdown fathers” in spring 2020 to investigate its consequences.

The study, which surveyed 2,045 UK fathers, found that many reported spending more time than usual on childcare and education during the lockdown. Dads also predominantly indicated that they emerged from this experience more confident as parents and in better relationships with their children.

My SoNeAt Lab’s research, published in the journals Child Development and Social Neuroscience, unveils just how deeply this confidence can be traced. My team found stronger brain-to-brain synchrony for those father-child dyads in which dads have more positive attitudes about their parenting abilities. And we also observed that dads who have more positive attitudes about their parenting abilities, and about fatherhood in general, show differences in their brain structure (compared to those who don’t).

These findings are timely. In many societies, dads’ involvement in parenting has dramatically increased. And as the UK Fatherhood Institute’s study revealed, this has a wide range of effects — including fathers gaining insights into their partners’ caring roles and growing more aware of the importance of dads spending time with their children. Many of the fathers in that study reported wanting to keep some of the changes, such as more flexible working hours and the opportunity to work more from home once the pandemic was over.

We shouldn’t forget, however, that separated fathers saw their children less during the lockdown, and only a minority of these dads reported to have left lockdown feeling more confident in their ability to help with their children’s schoolwork. This crucially demonstrates the importance of supporting dads by giving them more time and space to bond with their kids — regardless of their relationship status and family constellation.

What we did

The number of scientific studies on the role of fathers in child development has steadily grown over the last few years. Nonetheless, the vast majority of available research on parenting, attachment and caregiving has focused on mothers. This bias is partly down to mothers historically shouldering most of the childcare burden and being more available as research participants. To fill this gap, we wanted to specifically investigate the father-child relationship from a social neuroscience perspective.

In a first step, my team invited 66 dads and their kids, aged five to six, to the lab and asked them to solve puzzles. Fathers and children did so either together or independently while we scanned their brains with functional near-infrared spectroscopy (fNIRS). This is a method that records brain activity with the help of optical sensors that are attached to a head cap. We then tested whether dads’ and kids’ brains get “in sync” with each other during cooperative puzzle-solving. Such synchrony arises when there are aligned decreases and increases in activity in the same brain area at roughly the same time.

In a second step, my team assessed 50 (out of the above 66) fathers’ brain anatomy with magnetic resonance imaging (MRI). Our main area of interest was the hypothalamus. The hypothalamus is involved in various physiological processes such as hormone production and maintaining body temperature, but also plays a key role in pair-bonding, parenting and caregiving.

Finally, dads provided answers to two questionnaires relating to their caregiving beliefs. With the “role of the father questionnaire”, we assessed how strongly dads believed that they are capable of being a good father, should be sensitive to their children, and involved in child development. Another questionnaire probed how much dads indicated enjoying spending time with their kids.

What we found

Dads’ attitudes toward parenting have previously been shown to be a strong predictor for their involvement and capability. My SoNeAt Lab’s results underscore the importance of fathers’ caregiving beliefs on a neurobiological level.

On the one hand, our findings revealed that dads’ and their kids’ brains were more ‘in sync’ during collaborative puzzle solving when dads scored higher on the first questionnaire – how strongly they believed they were good and involved fathers.

On the other hand, we also observed that, in the same (50 out of 66) fathers, brain anatomy, and specifically the size of the hypothalamus, was associated with dads’ caregiving beliefs: the higher fathers scored on the first questionnaire, the larger was their hypothalamus volume. Furthermore, dads’ hypothalamus volume was also positively related to how much they reported to enjoy interacting with their kids.

My team’s research therefore sheds light on both brain activity in dads and their kids during social interaction as well as dads’ brain anatomy as such. However, there are a few issues to be kept in mind.

What needs to be keep in mind

First, when it comes to parent-child brain-to-brain synchrony, the general impression is that ‘more is better’. Because increased synchrony is usually linked with facilitated coordination and communication as well as higher relationship quality between children and parents. However, this may not always be the case: more synchrony may not always be better.

In particular, increased brain-to-brain synchrony could also be linked to heightened cognitive effort related to attention and emotion regulation. In other words, if other interaction elements lack in synchrony and/or synchrony is related to negative emotions and experiences, both kids’ and parents’ brains may need to work harder to sustain the interaction. As a consequence of such more difficult and less mutual interactions, increased synchrony may yield wear-and-tear effects, especially in the long-term. Therefore, context really matters.

Second, our results (both regarding synchrony and brain anatomy) revealed correlations, not causations. We therefore cannot say whether stronger caregiving beliefs in fathers leads to more father-child neural synchrony, or whether being more ‘in sync’ increases fathers’ caregiving beliefs. And we also cannot say whether hypothalamus volume determines how good men are at being dads, or whether wanting to be a good dad leads to increased hypothalamus volume. Additional research is needed to answer these questions.

Take home message

Dads are more involved in childcare than ever before. And, as my SoNeAt Lab’s research shows, their caregiving attitudes seem to be related to both brain-to-brain synchrony during interactions with their kids as well as their brain structure.

It therefore seems very relevant and justified to promote the importance of dads’ involvement in childcare in a broader societal context, giving them enough time and space to bond with their kids. Doing so will not only raise dads’ confidence and capabilities, but be rewarding for them, their children and partners. After all, dads – like mums – are biologically wired to be parents.

For more information about the social neuroscience of attachment and caregiving in fathers, please feel free to check out my dedicated website on ‘Caring Dads’.

Leave a comment

This article first appeared in The Conversation UK on 08 November 2021 but has been edited and updated.

My first attachment insights post was about disorganised attachment — what it is, what it isn’t, and particularly the confusion surrounding insecure-disorganised attachment in adults.

I’ve had quite a few discussions about disorganised attachment with other attachment researchers since. And I saw that, more recently, my post was picked up and scrutinised by others — for example, as part of an episode of The Psychology in Seattle Podcast.

These experiences drew my attention back to disorganised attachment. What really is and isn’t disorganised attachment, and why do we seem to be lost in translation when talking about it?

What is disorganised attachment?

To understand what disorganised attachment is, we need to go back in time. Back to the beginnings of attachment research when Mary Ainsworth and her colleagues first developed the Strange Situation Procedure (SSP). As part of the SSP, they especially looked at two things: how infants explore an unfamiliar environment and how they respond to their caregiver when distressed by two brief separations and subsequent reunions.

Ainsworth and colleagues’ observations gave rise to three basic infant attachment patterns or classifications: secure, insecure-avoidant and insecure-anxious. Importantly, all of these classifications refer to infants’ organised, workable attachment strategies. In the case of secure attachment, the strategy involves striking a balance between attachment and exploration. For insecure-avoidant attachment, the strategy consists in minimising attention to attachment-related information and instead shifting it toward exploration of the environment. And for insecure-anxious attachment, the strategy involves maximising the engagement in attachment behaviour (such as clinging to the caregiver) at the expense of play, or mixing whiny or angry behaviour with distress.

However, researchers quite quickly realised that not all infant responses during the SSP could be placed within the original three classifications. Therefore, Main and Salomon (1986, 1990) introduced disorganised attachment as a fourth classification — distinct from organised secure, insecure-avoidant and insecure-anxious attachment.

If seen at sufficient intensity and in the presence of the caregiver during the SSP, a disorganised attachment classification is based on the following observable infant behaviours:

Well then, after having discussed what disorganised (infant) attachment is, it should be rather straightforward to conclude what it isn’t.

What is not disorganised attachment?

A first clue was already provided by Main and Solomon (1990). They advise that the disorganised infant attachment classification (as part of the SSP) should not be used for children older than 20 months. This is because older children generally develop more sophisticated strategies for coping with caregiver behaviours and may therefore no longer show the disorganised attachment behaviours cited above.

But wait. Does this mean then that there is no more disorganised attachment in anybody older than 20 months at all? And what happens to older children, adolescents and particularly adults who, as infants, were classified as disorganised — does their disorganisation just cease to exist?

This is the point where things around disorganised attachment usually get most confusing.

You see, purely technically speaking, for somebody to be classified as disorganised — and especially so in adulthood —, there needs to be a valid and reliable measure that comprises disorganised attachment as a direct outcome. With a list of clear criteria what does and doesn’t constitute disorganisation. And there just isn’t any such measure in adulthood.

In adults, attachment is most commonly assessed with two procedures: semi-structured interviews and self-report questionnaires. Most prominent examples are the Adult Attachment Interview (AAI) and the Experiences in Close Relationships (ECR) questionnaire, respectively.

In the case of the AAI, trained coders rate the coherence with which adult interviewees speak about early interactions with their caregivers, rather than the content of what they say. This coding then yields the three organised attachment classifications of secure (autonomous), avoidant (dismissing) and anxious (preoccupied), and a fourth classification of unresolved (loss or other trauma). The unresolved classification is marked by (often momentary) disoriented, confused speech around experiences of loss of an attachment figure or other potentially traumatic events like maltreatment and/or neglect. Naturally, unresolved attachment coded from the coherence of adult interviewees’ statements as part of the AAI is quite different from disorganised behaviours infants display towards their caregiver during the SSP.

In the case of the ECR-R, adult participants self-report their feelings, thoughts and behaviours within romantic (and sometimes other close) relationships on two attachment dimensions, avoidance and anxiety. This way of assessing attachment yields an outcome of security if self-reported scores on both the avoidance and anxiety dimensions are low. And if self-reported scores on both the avoidance and anxiety dimensions are high, which reflects a co-existence of avoidant and anxious tendencies, the outcome is usually referred to as fearful(-avoidant) attachment. Once more, rather intuitively, such a way of assessing attachment with a self-report questionnaire results in a quite different outcome than the observation of disorganised behaviours infants display towards a caregiver during the SSP.

Based on the above, we can conclude that, purely technically speaking, there just is no such thing as an insecure-disorganised adult attachment style.

Such a notion is corroborated by the Society for Emotion and Attachment Studies (SEAS). On their website providing Explanations of Attachment Theoretical Concepts, the following is noted:

A list of additional references that explore discrepancies in various attachment discourses as well as a “transmission gap” between research and practice is furthermore summarised on my website.

Where do we go from here?

Having concluded that, purely technically speaking, there is no such thing as an insecure-disorganised adult attachment style, we could just end the discussion here. After all, why should we keep talking about something that cannot be directly and reliably measured? And of what use could such a concept be for anybody?

This is where my first attachment insights post ended. Mainly due to space constraints, as it was primarily written for Instagram, which has a limit of 2200 characters including spaces. However, I do appreciate that there is more to this story than what I mentioned at the time, and also beyond what we have explored so far in this post.

First, although the AAI and ECR-R are the most common tools to assess adult attachment, there are many others. And quite a few them actually do comprise a disorganised classification (for an overview, see here, here and here). That said, those tools that do comprise a disorganised classification have been rated as ‘inadequate’ due to issues associated with measure development, including failing to employ appropriate qualitative methods in their measure development or other pilot testing with the target population. It therefore seems that while there is a scientific interest in disorganised adult attachment as a concept, there remains the need for further psychometric testing of existing measures, or development of new instruments, grounded in sound methodology.

Second, even within the attachment researcher community, there is no absolute consensus on the construct of disorganised attachment. For example, some researchers refer to the unresolved AAI classification interchangeably as disorganised (U/d) because interviewees may become incoherent or dissociative when discussing past abuse or loss. And even when it comes to infant disorganised attachment, an alternative view based on Crittenden’s Dynamic-Maturational Model (DMM) of attachment and adaptation suggests that many behaviours that others would classify as “disorganised” could instead be seen as “organised” adaptations (for a summary and critical evaluation, see here).

Finally, we need to keep in mind that attachment theory and its concepts of organised and disorganised attachment are referred to and used by a wide variety of individuals outside the immediate context of attachment science, including — amongst others — neurobiologists, psychotherapists, child welfare practitioners, parents, and the general public (see e.g., here). When these individuals refer to and use concepts of attachment theory, they very often do so without directly measuring attachment (e.g., with the SSP, AAI or ECR-R). As a consequence, for them, a discussion primarily about technicalities of particular attachment measures and an absence of a direct, valid and reliable classification of disorganised adult attachment isn’t very useful. If it is widely accepted and “correct” to refer to disorganised attachment in infants, why should it all of a sudden be “wrong” to talk about disorganised attachment in adults?

To be honest, I currently don’t have a quick fix for these discrepancies.

For me personally, I do think that when we talk about attachment, it is important for us to be as scientifically accurate as possible. Is our description of attachment mainly based on behavioural observations, coherence coded from semi-structured interviews or self-reported thoughts, feelings and behaviours? And do we refer to concepts put forward by developmental psychologists that describe child-parent attachment relationships, or are we rather discussing concepts that emerged from social psychology and therefore describe close adult relationships?

At the same time, we need to be careful that our aim for as much scientific accuracy as possible doesn’t interfere with, or even preclude, the use of attachment concepts altogether. There are many small details that only attachment researchers can realistically be expected to be fully aware of. And even attachment researchers don’t always agree on everything.

I nonetheless still stand by my initial statement that there — at least for now — is no such thing as an insecure-disorganised adult attachment style. However, I am fully aware that the discussion about disorganised attachment is to be continued, both within the field of attachment as well as between attachment researchers and everybody else. And I am happy to participate in this discussion, especially if it helps to further advance attachment theory and science.

© 2025-2026 by Pascal Vrticka. All rights reserved.

Leave a comment

Attachment theory is almost everywhere. In magazines and books, in the news, on social media and in our conversations with each other.

Originally rooted in developmental psychology, the theory explains how we form and maintain close relationships in order to survive and thrive in the environment we are born into. It was quickly picked up not only by pop culture but also social psychology, psychotherapy, psychiatry as well as child welfare practice.

But some of the most important features of attachment theory are getting lost in translation. Misunderstandings are leading people to believe they have a “bad” attachment type that is wreaking havoc on their relationships.

Attachment theory is not a gauge of whether someone has the “wrong” or “right” attachment type. Its purpose is to help people understand what coping strategies they use when the people they are closest to are, or are perceived to be, unavailable or inconsistently responsive.

When psychologist John Bowlby developed attachment theory about 70 years ago, he had two goals in mind. He wanted to create language that can be used to scientifically formulate and test hypotheses. It worked: hundreds of studies have since been carried out exploring the many facets of attachment across human lifetime.

Bowlby also wanted to develop a language that appealed to the public: evocative and relatable. Attachment theory soon became the talk of the town and was picked up by other scientists, practitioners, politicians, lawyers and parents. It gave people a new way to understand infant-parent and adult relationships with friends, family and romantic partners. Unfortunately the meaning has been diluted.

What attachment theory means

The confusion starts with the term attachment itself. It is often understood among the public to mean a child’s love for their parent or the bond a parent forms with their child. People also use attachment to describe how adults relate to and interact with others.

In attachment theory attachment refers to a more selective type of relationship in which only a few people are sources of support. A parent does not become attached to their baby but acts as the infants’ caregiver and attachment figure.

Their caregiving may nonetheless be influenced by their own attachment type. As adults we only become attached to a limited number of people although our attachment type in close relationships also tends to reveal itself during our social interactions with friends, colleagues, acquaintances and even strangers.

People also often say children have a weak or strong attachment to their caregiver or that we are more or less attached to our romantic partner. But attachment theory focuses on how we differ in the quality of our attachment bonds in terms of security.

There is consistent evidence for better developmental outcomes in securely attached children and fewer mental health related problems among securely attached adults. However, all attachment types, whether secure or insecure, are meaningful because they are adaptations. There aren’t good and bad attachment types.

It makes sense to develop an avoidant attachment type – being more self-reliant and socially distant – if there are no responsive attachment figures around in times of need. Attachment security does not describe a good and confident psychological state which should be desired by everyone. It reflects someone’s perception of the availability and responsiveness of attachment figures.

Attachment insecurity is about compensation mechanisms (called secondary attachment strategies). These strategies help us cope with insecurity when we experience others as unavailable or inconsistently responsive. They are appropriate, often necessary and should not be labelled as bad or maladaptive.

Many people think of so-called disorganised child attachment, which may develop when the child’s caregivers become a source of distress, as a chaotic state. In popular culture disorganised attachment is thought to almost always arise because of adverse childhood experiences like abuse or neglect. But developmental psychologists know there are many possible causes of disorganised attachment and its links with child maltreatment are complicated at best.

How to tackle the problem

Scientific studies that take an engaging and unusual angle can be a great tool to share knowledge with the public. The purpose of psychology after all is to help us understand ourselves and how we relate to others. For example, a recent study investigated how the lyrics of our favourite songs may be related to our attachment type.

It found that we tend to return to the tunes about relationships that mirror our own experiences. And particularly so if we are avoidantly attached, which makes us more likely to prefer music with avoidant lyrics. The study also found that the lyrics of over 800 songs written between 1946 and 2015 overall became more avoidant and less secure, which could reflect a trend in society.

Attachment researchers are also working to tackle the widespread confusion. For example, the Society for Emotion and Attachment Studies (SEAS) has created an excellent free online guide that lists many attachment concepts, their common misconceptions and accessible definitions.

My team’s research is taking another approach. We are building language to talk about the social neuroscience of human attachment (SoNeAt). This involves studying attachment theory in action by combining psychological attachment methods like behavioural observation, interviews and questionnaires with neurobiological measures like heart rate, secretion of hormones (such as oxytocin or cortisol) and brain activation, structure and connectivity.

We hope our efforts can help bring more clarity to attachment language not only across different scientific domains but also when communicating attachment research to the public.

Understanding how infant-parent and adult attachment works can give us important tools for reflecting upon our relationships. This only works, however, if we agree on a common attachment language.

This post first appeared in The Conversation UK on November 30, 2022.

Leave a comment

It’s crucial for healthy child development that children can form secure attachment bonds with their parents. Decades of research identified one key ingredient for this process: the coordination of parents’ and children’s brains and behaviour during social interactions.

Humans connect with each other by synchronising in many ways. Called bio-behavioural synchrony, this involves imitation of gestures and the alignment of heartbeats and hormone secretion (like cortisol and oxytocin). Even brains can synchronise – with brain activity decreasing and increasing in the same areas at roughly the same time when we spend time with others.

My colleagues and I carried out research which showed that brain-to-brain synchrony between parent and child can be helpful for children’s attachment, and tends to rise when a parent and child play, talk or solve problems together. Recently, however, we started wondering whether more synchrony is always better. Our recent study, published in Developmental Science, suggests it can sometimes be a sign of relationship difficulties.

Is current parenting advice up to date?

A lot of current parenting advice recommends parents to be constantly “in sync” with their kids. It tells parents to be physically close and attuned to their children and to anticipate and immediately respond to their every need.

The advice is building upon attachment theory and research, which show that higher parental sensitivity and reflective functioning are beneficial for child development and secure attachment formation.

Yet, despite its good intentions, this advice misses several important details. For example, research revealed that for about 50-70% of the time, parents and children are not “in sync”. During these times, they may be doing separate activities, such as a child exploring something on their own or a parent working. They rather engage in a constant “social dance” comprising being attuned to each other, failing to do so and repairing this disconnect.

And it’s this flow of connection, disconnection and reconnection that offers children an ideal mixture of parental support and moderate, useful stress that helps growing children’s social brains.

Researchers also agree that there can be negative consequences to parents and children constantly being tuned in to each other. For example, it can increase stress on the relationship and raise the risk for insecure child attachment. That is especially true if it is associated with parents overstimulating their child or being too responsive to their child’s every need.

For parent-child synchrony, there thus appears to be an “optimal midrange”. Or, in other words, more synchrony may not necessarily be better.

Brain-to-brain synchrony and attachment
Within a large international team of investigators from across Europe, my colleagues Trinh Nguyen, Melanie Kungl, Stefanie Hoehl, Lars White and I set out to investigate how exactly parent-child bio-behavioural synchrony is linked to attachment.

We invited parent-child pairs – 140 parents and their 5-to-6-year-old kids – to our SoNeAt Lab where they solved tangram puzzles together.

We measured brain activity with functional near-infrared spectroscopy (fNIRS) “hyperscanning”, for which parents and children were asked to wear caps linked up with optical sensors. We also recorded videos of their interactions so we could assess how much behavioural synchrony they demonstrated – how attuned and attentive they were to each other. And finally, we assessed parents’ and children’s type of attachment – known as attachment representations.

We previously discovered increased neural synchrony in mother-child and father-child pairs during different tasks. In mother-child pairs, neural synchrony was linked to taking turns in solving puzzles or conversations. And in father-child pairs, synchrony during puzzling was linked to dads being confident about and enjoying their role as fathers. But does that mean higher parent-child neural synchrony is always a measure of a good relationship?

In our new study, we actually observed that mothers who had an insecure, anxious or avoidant attachment type showed more neural synchrony with their children. Interestingly, mothers’ attachment types were unrelated to how synced mothers and children were in terms of their behaviour. We also found increased neural but decreased behavioural synchrony in father-child pairs (compared to mother-child pairs) independent of attachment.

Our findings suggest that higher neural synchrony may be the result of putting increased cognitive effort into the parent-child interaction. If mothers’ attachment representations are insecure, it may be more difficult for mums and kids to coordinate and help each other during activities such as puzzle solving.

A similar explanation may apply to neural synchrony during father-child problem-solving. Dads are more familiar with active, rough-and-tumble play. Engaging in structured and cognitively demanding activities such as puzzles may therefore be more challenging and require more neural synchrony for father-child pairs.

Lessons to be learned
What do our new findings mean?

Most importantly, parents should not feel that they must be “in sync” with their kids all the time and at all costs. High parent-child attunement can also reflect interaction difficulties and can often add up to parental burnout, further negatively impacting the parent-child relationship.

It is of course helpful if parents are emotionally available, skilled in reading their children’s cues and promptly and sensitively respond to their needs. Especially when children are young. However, it suffices for parents to be “good enough” – to be available when children need them rather than “always on”.

Children can also benefit from freedom and independence emotionally, socially and cognitively, especially as they get older.

What really counts is that the parent-child relationship functions well overall. That children can develop trust in their parents and that any mismatches, which naturally occur all the time, are successfully repaired. That’s the true essence of attachment theory, which is often missed and misrepresented in parenting advice

To better navigate the challenging parenting path, parents need access to trustworthy and up to date sources of information. Together with the UK Charity Babygro, we therefore published the free-of-charge Babygro Book for Parents that provides them with evidence-based knowledge on parenting and child development.

It is our hope that our book can empower parents so that they feel reassured and confident in their own parenting choices and can optimally support their children to grow and thrive.

© 2025-2026 by Pascal Vrticka. All rights reserved.

Share

Leave a comment

This post first appeared in The Conversation UK on 21 May 2024.