These aren’t merely playlists inspired by the book. They contain the songs referenced in it.

Why create a soundtrack for a book? My immediate purpose is to help readers appreciate the book’s musical references. I don’t want readers to have to take my word for what a song conveys. They can listen to it themselves, before, during or after reading the relevant chapter.

There’s also a deeper reason, connected to one of the book’s main arguments. I maintain that art is often sold with a false bill of goods: the promise that experiencing it will somehow transform us. Art can be transformative, but we consume so much of it that we seldom engage with particular works deeply enough (even if we engage with it repeatedly) to benefit fully from them.

The playlists are intended to encourage readers to do just that. Listen to a song, read what I have to say about it, then listen again. Perhaps you’ll hear something you hadn’t noticed before. Perhaps you’ll disagree with my response to it. Either way, the music becomes something to think about, rather than merely something to consume. Of course, the playlists can also be enjoyed for their own sake.

About the book

Discontinuities: Love, Art, Mind is about responding to art, rather than creating it. Furthermore, it goes well beyond interpreting or critiquing art. It ranges across music, films, novels, short stories, painting and other art forms. Its central concern is how we can engage with art deeply, including in ways that might transform us.

The book applies and extends ideas from my two Cognitive Productivity books. It draws on cognitive science, philosophy and psychology, as well as on art itself. Music also informs the organization of the book. Each one of its ten parts have musical concepts as titles, such as Overture, Symphonic poem, Canon, Affective variations, Lament, Rondo and Syncopation decelerando.

Many chapters are adapted from articles I published on CogZest.com starting in 2012. Others were written specifically for this book. The book also includes contributions from other writers and artists.

What’s with the title?

The title, Discontinuities: Love, Art, Mind, deserves a little explanation.

Two of the meanings of Discontinuities are as follows. First, the book itself is discontinuous: its chapters can be read independently and in almost any order. You need not read it from cover to cover. More fundamentally, I argue that reality is not merely continuous. Discontinuities are pervasive in the physical world, in minds conceived as virtual machines, and in the world of human creations. This connects with Karl Popper’s three-world ontology, among other philosophical and scientific ideas explored in the book.

Love refers partly to two important themes: limerence (the obsessive phase of romantic love) and sapiosexuality (romantic attraction based on another person’s intellectual qualities). But it also points to affect more generally: emotions, moods, attitudes and motivations.

Art reflects the book’s concern with how we can engage deeply with works of art.

And Mind reflects the cognitive science perspective underlying much of the book. My approach is integrative and design-oriented.

The book’s other indexes

The soundtrack isn’t the only guide to the many works discussed in Discontinuities. An appendix contains tables of the musical works referenced, organized by part and chapter, indicating each work’s title and creator. Each title linked to a YouTube video of the song. I also plan to include similar tables of novels, short stories and films.

I would reproduce the tables here, but Substack doesn’t support tables. (Tsk! Tsk!)

Completion status

The book is approximately 98% complete. I’ve written 38 chapters, with only the final chapter still requiring substantial work. Another round of proofreading remains to be done.

Contributors

As this is an edited book, I’ve also had the pleasure of including work by several contributors. I feel honoured that the following people have agreed to contribute new works of theirs to this book, or have made past works available for inclusion:

• Dr. Al Sather

• Andrea Paterson

• Arash Kameli

• Arnold Johnston

• David Deutsch, who recently gave permission to reproduce an excerpt from his discussion of continuity and discontinuity in physics

• Dr. Claude Lamontagne

• Lam Wong

• Dr. Laura Nys

• Laura Redmond

• Pierre Bellefeuille

The book also includes two short stories by Henry James and Willa Cather which are in the public domain.

Getting the book

You can purchase Discontinuities: Love, Art, Mind as an ebook from Leanpub. Use this coupon code URL to get the book at half the current price: https://leanpub.com/discontinuities/c/Substack

And whether or not you buy the book, you’re welcome to enjoy the main soundtrack and its French-language companion.

Colophon

I wrote the original draft of this article entirely by myself, then put it through ChatGPT for minor editing. I reviewed and either approved the changes or edited the draft further.

How can I possibly remember so many abbreviations? And why use a text expansion utility for this purpose when Hookmark already provides bookmarking and information retrieval facilities? The answers involve a rather interesting property of the human mind, one that I discussed in a 2015 CogZest article on TextExpander snippet conventions.

Using a text expansion utility as a link manager

Suppose I want to send someone a link to my recent Substack article about Clean My Links, a Hookmark feature that removes tracking information and other junk from web URLs. The URL is: https://luccogzest.substack.com/p/the-hidden-tracking-junk-in-web-links.

I could visit Substack, find the article, and copy its URL. Or I could search for it in Hookmark’s bookmarks. But I don’t need to do either of those things. Instead, I just type:

#ssclmls and the tab key.

This stands for “Substack Clean My Links” of course. If the link pointed to the Hookmark blog, I would have abbreviated as hmblcmls for “Hookmark, blog, Clean My Links”. A text expansion utility expands that abbreviation into the full URL. I can do this directly in an email, a document, or just about any other text field. There’s a method to this abbreviation. The # tells me that the snippet expands into a URL. The letters ss stand for Substack. The remaining letters refer to Clean My Links: cl for clean, m for my, and ls for links. Here’s another example:

#czblete followed by the tab key.

This expands into the URL of my aforementioned 2015 TextExpander article. In this case, czbl means CogZest blog. I use that prefix for CogZest blog articles. The remaining letters identify the particular article.

You can probably see where this is going. Rather than memorizing an arbitrary collection of abbreviations, I use conventions for constructing them. These conventions are especially useful when writing, because I can insert a link without interrupting my thinking to go looking for it.

Markdown links

I do a lot of my writing in Markdown. For this, I have another convention: my text expansion abbreviations that expand into Markdown links begin with [. For instance, given the URL snippet above, I could define the corresponding Markdown snippet as:

[ssclmls followed by the tab key.

which would expand into: [The Hidden Tracking Junk in Web Links](https://luccogzest.substack.com/p/the-hidden-tracking-junk-in-web-links)

Notice that the abbreviation begins with [, which is also the first character of a Markdown link. This makes the convention easy to remember.

So, in my personal shorthand system, # means that I want a URL, while [ means that I want a Markdown link. The rest of the abbreviation can remain the same. That’s important: once I’ve learned an abbreviation for a resource, I can reuse most of it for different purposes. This also makes it easier to give people links. Instead of having to re-search for the web page, or search through a bookmark manager (Hookmark has one), I just type in the abbreviation and expand it.

Developing a personal shorthand language

In my 2015 article, I explained that the key to using a text expansion utility extensively is to develop a system of conventions, or what one might call a grammar for abbreviations. This is an important cognitive point, reflecting my grounding in cognitive science. Our brains are not particularly good at memorizing long lists of arbitrary character sequences. But they are remarkably good at recognizing, learning and applying patterns. In fact, learning and using grammars is one of our most impressive cognitive abilities. (I discussed intelligence more generally here. Incidentally, I got that URL in one second using the abbreviation #ssint, standing for “Substack Intelligence”.)

By establishing conventions, I reduce the burden on my memory. When I need to construct a new snippet, I can often derive its abbreviation from conventions I already know. Similarly, when I need to use an existing snippet, I can often reconstruct its abbreviation rather than recall it verbatim. In other words, the conventions are generative. They allow me to produce and interpret many abbreviations from a relatively small number of rules.

This is much like learning shorthand for writing with pen and paper. Shorthand systems are not simply collections of arbitrary symbols one must memorize individually. They have rules for representing words, sounds, and common expressions. As one learns the system and practises using it, the abbreviations become increasingly natural.

Text expansion utilities provide a computerized form of shorthand, but one can go well beyond abbreviating words and sentences. One can use it to access information, enter bibliographical references, insert Markdown syntax, and much more.

Incidentally, I’m not suggesting that you adopt my particular conventions. Your abbreviations should reflect the types of information you work with, the language you use, and your own habits. What’s memorable to me might not be memorable to you. The important thing is to develop conventions that are systematic enough to be learnable, but sufficiently flexible to be useful. And, of course, you need to practise. It might be demanding at first, but like any other habit, it gets easier and faster with time. Besides having a good abbreviation system doesn’t help much if you never use it. This is another application of a central principle of cognitive productivity: one needs to invest in developing expertise with one’s tools. (Cal Newport’s digital minimalism goes too far in my opinion, a topic I will later write about.)

Opening web pages without searching for them

So far, I’ve emphasized using a text expansion utility to insert web links in writing. But I also use it to access web pages. Suppose I want to open one of my frequently referenced articles. I can simply:

1. Press Command-L in my browser to select the address bar.

2. Type the abbreviation followed by my expansion delimiter.

3. Press Return.

That’s it. The web page opens.

Of course, modern browsers have quite sophisticated address-bar completion facilities. I use those too. But sometimes they present several possible matches, requiring me to examine the results and select the right one. With a familiar shorthand abbreviation, I know exactly what I’m going to get. This is not merely a matter of saving a few seconds. Searching, visually scanning results, and deciding which item to select all require cognitive processing. When I’m deeply engaged in writing or thinking, I’d rather avoid those interruptions. Besides, repeatedly searching drains brain resources, contributing to fatigue.

That is one reason I have continued developing and using these conventions over the years.

These are benefits analogous to the advantages of using Hookmark. You can invoke Hookmark in the context of one resource (e.g., a PDF) and see links related to it (e.g., other PDFs, or a task list associated with it). No searching required beyond scanning the contents of Hookmark’s context window, which normally contains only a few key items.

Why use a text expansion utility when I have Hookmark?

Hookmark and text expansion utilities are quite different tools. Each enables some things that the other doesn’t.

Hookmark lets me copy links to information in many different apps, including web browsers, Finder, email clients, task managers, reference managers, and note-taking apps. It also lets me hook items together, creating bidirectional links between related resources. For example, I can hook a research paper to my notes about it, to an email discussion concerning it, and to the draft of an article in which I discuss its implications. Later, when I return to any one of these items, Hookmark can show me the related items. This is a tremendous help in managing complex information work.

Hookmark also automatically bookmarks information when I copy links to it or hook it to something. And it provides facilities for searching and navigating these bookmarks. I don’t need to assign memorable abbreviations to all those resources. Hookmark only bookmarks information you’ve already implicitly or explicitly indicated you want to re-access. So the Hookmark database is likely to include what you want to re-access, and only what you want to re-access.

A text expansion utility serves a somewhat different purpose. It allows me to give frequently used resources my own shorthand names.

Consider a situation in which I’m writing an article and want to refer to one of my earlier publications. If I remember its text expansion abbreviation, I can insert the link immediately. I don’t even need to open the publication or a bookmark manager. To be sure, Hookmark itself has a Copy Markdown Link command and provides searchable bookmarks. But a text expansion utility gives me another way to access particular links, one that is especially efficient when I’ve learned the relevant abbreviation. It avoids needing to type in a search query in one’s Bookmark manager.

Hookmark and text expansion utilities work well together. For instance, I can use a text expansion utility to produce a URL, copy it, and then use Hookmark’s Focus on Link in Clipboard command to work with that link in Hookmark. From there you can access information that was hooked to the link. From there I can for instance create a hook to the link or a .hookmark file for storage in the Finder. And of course Hookmark works not only with web URLs, it can link to files, emails, tasks, notes and more.

The more general point is that I don’t expect Hookmark, or any other app, to do everything related to accessing and managing information. Another example would be using an information manager such as DEVONthink or EagleFiler, both of which are compatible with Hookmark.

The meta-access problem

Readers of my academic papers, and my Cognitive Productivity books will recognize that this is another instance of what I call the meta-access problem. We possess, create and encounter vast amounts of potentially useful information. Yet when that information becomes relevant to our current activity, we often have difficulty finding it. We might remember having read an article, received an email, or written a note, but not remember where to find it. Sometimes we don’t even remember that we have the relevant information.

Hookmark addresses the meta-access problem by helping us create links between related information and retrieve items in context. A text expansion utility, with the conventions I’ve described, provides another means of access: a learned shorthand for frequently used resources.

These approaches are complementary. One exploits the relationships between information items. The other exploits our ability to learn and use patterns. Neither completely solves the meta-access problem. But both reduce the friction involved in accessing information, and both can help us make better use of what we know and what we’ve saved.

Upshot: develop your own conventions

If you use a text expansion utility, I recommend giving some thought to the conventions you use for your snippets. And if you regularly refer to particular web pages, consider storing their URLs and Markdown links in it. You don’t need to begin with hundreds of snippets. Start with the resources you use most frequently. Develop conventions that make sense to you. As your needs change and your collection grows, extend your system.

Remember that developing a useful shorthand is a skill. As I noted above, like other skills, it benefits from thought, practice, and occasional revision.

The broader lesson is one I’ve emphasized in my Cognitive Productivity: becoming more productive with technology often requires more than acquiring apps. We need to develop knowledge and expertise in using them, including ways of making them work together in a cognitively productive manner.

I was using text expansion to access web links long before we developed Hookmark. I still use it for that purpose today, notwithstanding Hookmark’s powerful bookmarking and linking facilities. There’s no contradiction here. The aim isn’t to use one app for everything. It’s to become more effective in our work by exploiting the complementary strengths of our tools.

Colophon

This article is based on my 2015, long-form, post on the subject. I used AI to help summarize and update the original article, adapting and extending it for a Substack audience. None of the ideas come from AI. The original post on CogZest.com has more examples than presented here, so it remains relevant.

Every day, we copy web links and paste them into emails, text messages, notes, documents, and social media posts. But have you noticed how ridiculously long some of these links have become?

You might want to share a link to a book on Amazon, a LinkedIn article, or a post on Facebook. Instead of a simple web address, you get a sprawling string of letters, numbers, and cryptic symbols. Much of that extra information has nothing to do with identifying the page you want to share.

It’s called URL cruft (per this Wikipedia article, and it’s more than an aesthetic nuisance. I published several web pages on the subject, and a 2-minute video on YouTube explaining this problem and how easy it is to fix it with Hookmark’s Clean My Links feature. Disclosure: Hookmark is an app I design at CogSci Apps to improve cognitive productivity.

What’s hiding in your web links?

Consider this link to my article on Ubiquitous Linking which I published on LinkedIn:

Original cruft-filled link:

https://www.linkedin.com/pulse/manifesto-ubiquitous-linking-luc-p-beaudoin/?trackingId=5%2FsgqsR%2FTs24OpaGFSnIXQ%3D%3D

Cleaned link:

https://www.linkedin.com/pulse/manifesto-ubiquitous-linking-luc-p-beaudoin/

Both lead to the same article. But the first contains a tracking identifier that serves no useful purpose for someone who simply wants to read it. The cruft is designed to serve the web site’s company’s interests, not yours. These URLs contain tracking and attribution information. When such links are shared, companies can use embedded parameters to infer marketing effectiveness, attribution paths, or user relationships.

Amazon and social media links

Amazon can be even worse. A link to a book can contain dozens of characters of advertising identifiers, referral information, search parameters, and other data. Yet the essential address of the book may be quite short.

Here is a typical link to a book page on Amazon:

https://www.amazon.ca/Enshittification-Everything-Suddenly-Worse-About/dp/0374619328/ref=sr_1_1?dib=eyJ2IjoiMSJ9._SjOlLplb7Z0Rvscl1ZihXsKv3j5PVjTYExbZMJmZuDtJLCIavbyJiKbtQJUy3Xn0bNemqglamuZWtNQzFB2_Q.x7nWZNgrA7o7I-w9roq4O53Ys4E1Wx9KWxrJHyQb-aY&dib_tag=se&gad_source=1&hvadid=788657704505&hvdev=c&hvexpln=0&hvlocphy=9001509&hvnetw=g&hvocijid=17902577840324087743–&hvqmt=b&hvrand=17902577840324087743&hvtargid=kwd-2078990547242&hydadcr=22430_13682881&keywords=enshittification&mcid=82e56b35149932d7aa07352f600aa68d&qid=1768014783&sr=8-1

This is not accidental. Many websites like Amazon use URL parameters to track advertising campaigns, measure referrals, and analyze how people reach their pages. Some parameters are useful or even essential, but many are unnecessary when sharing a link.

And every time we copy and distribute such a link, we may be propagating that tracking information.

Google, Facebook, X, and many other platforms add similar tracking cruft to URLs. For background, see Cruft – Wikipedia. For a broader critique of large-scale platform behavior, see Cory Doctorow’s Enshittification (2025) book (highly recommended reading).

Why should you care?

There are several reasons to prefer clean web links.

1. Privacy. Tracking parameters can communicate information about where a link came from or how it was obtained. Removing them limits the information passed along when you share a link. It won’t eliminate online tracking, but it’s a sensible privacy precaution.

2. Clarity. A clean URL is easier to read and recognize. You can often tell where it leads without wading through a jungle of identifiers.

3. Brevity Cruft filled links take up more space where you paste them. They call for more scrolling.

4. Durability. The same web page can have several different URLs, depending on how you reached it. A link copied from a search result may look different from one copied from a social media post. Removing unnecessary parameters helps establish a consistent address for the underlying resource. This doesn’t prevent websites from moving or disappearing, but it reduces one source of confusion.

5. Cognitive productivity. As I have argued in my books on cognitive productivity, information should be organized so that we can efficiently access and use it. Links are essential tools for navigating knowledge. Unnecessary complexity makes them harder to inspect, recognize, manage, and reuse.

In sum, a good link should identify its destination without carrying unnecessary baggage or violating your privacy.

Introducing Clean My Links

To address this problem, we at CogSci Apps introduced a feature called Clean My Links in version 7 of Hookmark, our linking utility for macOS. With Clean My Links enabled (by default), when you copy a link to a web page, Hookmark automatically removes known tracking parameters and other unnecessary information from web URLs when you copy them using Hookmark.

As a result, you don’t have to inspect the URL, figure out which parameters are safe to remove, or edit the link manually. Hookmark does the cleaning for you.

1. Importantly, this isn’t a conventional URL-shortening service. URL-shortening services keep a record of every link they shorten. It inherently tracks you, defeating one of the purposes of cleaning links. Hookmark is local to your computer only.

2. Hookmark’s Clean My Links feature doesn’t substitute a mysterious new domain or create a redirect through another company’s servers. The result remains a direct, private, link to the original website, minus the unnecessary cruft.

3. Most other cruft-removing software requires that you install a plugin in each browser that you use. Hookmark doesn’t. It works immediately across supported browsers through Hookmark itself. Plus, you might not trust the entity cleaning your links?

It’s free.

And here’s the best part: Hookmark’s Clean My Links is free in Hookmark Basic, even if you don’t purchase a Hookmark Standard or Hookmark Pro license.

See it in action

I’ve prepared a video demonstrating Clean My Links and how it improves the web links you copy and share.

You can also explore several before-and-after examples, including URLs from Amazon, LinkedIn, Facebook, X, and Goodreads.

How to use Clean My Links

If you’re using a Mac, the procedure is straightforward:

1. Download and install Hookmark.

2. Open a web page in your preferred browser.

3. Invoke Hookmark (with ⌃H)and use its Copy Link command (⌘C in Hookmark).

4. Paste the resulting clean link wherever you like.

Clean My Links is enabled by default. The feature uses conservative cleaning rules: if Hookmark can’t safely determine that part of a URL is unnecessary, it leaves that part alone.

Advanced users with Hookmark Pro can also define custom cleaning rules for particular websites.

You can read more about how Clean My Links works and why clean links matter.

Beyond clean links: connecting your information

Clean My Links is one feature of Hookmark. The app’s broader purpose is to help people create and navigate links between their information. For instance, you can use Hookmark to connect a web page to your notes about it, an email to a relevant document, or a list of tasks related to it (or anything else). These connections make related information easier to access, even when it resides in different applications.

For such connections to work reliably, information resources need stable identities. If the same web page appears under multiple superficially different URLs, it’s harder for software to recognize that they’re all referring to the same thing.

Cleaning URLs therefore contributes to Hookmark’s broader purpose: making information easier to connect, retrieve, and use.

A small improvement with cumulative benefits

Clean My Links might seem like a minor convenience. But consider how many links you copy, save, and share in a typical week, and how many years you may want some of those links to remain useful. I think of URL cleaning as a modest form of digital housekeeping. It’s a small improvement to an activity we perform constantly, often without thinking about it.

And because links are the connective tissue of our digital information, improving their quality can have benefits far beyond their appearance.

The web is cluttered enough already. There’s no reason our links should contribute to the mess.

Colophon

I used AI to help write this article based on several pages and a video I had previously published on the subject:

• Help page of Clean My Links feature

• Benefits of Clean My Links™

• Hookmark’s Clean My Links Settings tab

• Brief YouTube video: Tired of long, dirty, ugly web links that track you?

Despite the fact that some researchers posit a “sleep-state switch”, sleep onset is not an instantaneous switch from wakefulness to sleep. It is a complex, gradual transition involving changes in brain activity, physiology, consciousness, and information processing over several substages. As we approach sleep, our thinking often becomes less coherent, goal-directed thought diminishes, and fleeting, dreamlike imagery—known as hypnagogic imagery—may emerge.

But what controls this transition? And why do some people find it so difficult to fall asleep, even when they are exhausted?

An integrative design-oriented approach to sleep onset

In my lecture, I discussed somnolent information-processing (SIP) theory, which I have been developing for over a decade, most recently with my colleague Sheryl Guloy in a Cambridge University Press paper.

The SIP theory approaches sleep onset from an integrative design-oriented perspective. It is integrative in being interdisciplinary (involving psychology, AI and neuroscience) and intermodal as it posits connections among cognition, emotion, motivation and a distributed sleep onset control system. It asks how the brain’s various information-processing and neural mechanisms—including attention, motivation, emotion, executive control and their neural substrates—interact to regulate the transition to sleep.

A central concept is mental perturbance: a state in which pressing concerns, goals, or worries repeatedly capture our attention and interfere with our ability to control our thinking. At bedtime, such mental activity can be insomnolent—that is, it can interfere with falling asleep. It also claims that alarms —global responses to acute stressors, first posited by Hans Selye and expounded upon by Aaron Sloman and his H-CogAff theory— can also be insomnolent. These are processes I’ve discussed previously on my Substack (just follow the links above for more information on them).

However, the SIP theory makes a more interesting prediction: certain forms of mental activity, and their neural substrates — may positively facilitate sleep onset. I call these pro-somnolent processes. And these processes can be emulated by deliberate bedtime cognition.

The cognitive shuffle: putting theory into practice

This theoretical framework led me to develop the cognitive shuffle, a family of techniques designed to facilitate sleep onset.

One version, called serial diverse imagining, involves imagining a succession of diverse, unrelated objects, people, scenes, or activities. In the “2.0” version of this technique, I recommend (a) imagining these things in different locations (different rooms, buildings, etc), (b) imagining oneself physically interacting with them (e.g., for “piano”, imagine yourself playing the piano).

The technique is intended to work in two complementary ways. First, it engages limited neural information-processing resources, making it harder for worries and other insistent thoughts to dominate consciousness. Second, it emulates some characteristics of spontaneous sleep-onset mentation, potentially helping the brain progress toward sleep.

We call these proposed effects counter-insomnolent and pro-somnolent, respectively. They remain hypotheses requiring further empirical testing.

The cognitive shuffle illustrates a broader principle that has guided much of my career: theories of how the mind works can inspire practical techniques, which in turn generate new scientific questions and theory.

Try the app

We at CogSci Apps have developed a mobile app to help you engage in the cognitive shuffle while in bed wanting to fall asleep. It is called mySleepButton. The app can be useful when you are too tired to do the cognitive shuffle unaided (in its “DIY” form), yet remain unable to fall asleep. (Fatigue does not necessarily imply sleepiness.)

Watch the lecture

In this presentation, I explain the sleep-onset process, the cognitive mechanisms that may facilitate or inhibit it, and the theoretical reasoning behind the cognitive shuffle. I hope the lecture will interest sleep researchers, cognitive scientists, psychologists, and anyone curious about what happens in the mind as we fall asleep.

Watch my University of Ottawa lecture on YouTube.

Read the paper

Our upcoming paper in the Cambridge Handbook of Sleep Theories and Models provides more details on the SIP theory and cognitive shuffle technique:

Beaudoin, L.P. & Guloy, S. (in press). Towards a somnolent information-processing theory: Understanding the human sleep-onset control system from an integrative design-oriented perspective. In J. Dzierzewski, D. Kay & S. J. Aton (Eds), The Cambridge Handbook of Sleep Theories and Models. Cambridge University Press.

Receive training on sleep practices from Somnolence+

I am also co-founder of Somnolence+, which is a Canadian company that provides information and coaching to promote better sleep to individuals and organizations.

Imagine that you are revising a paper. Its Word document is open. Elsewhere are the outline, source PDFs, notes, an email from a reviewer, a figure, the submission instructions, and a task list. You know they are related, but they live in different apps. Finding each one can mean recalling where you put it, switching apps, inventing search terms, and then recovering your train of thought.

Search is useful. Yet the question you need to ask is often not “Where is the file containing these words?” It is “What belongs with the thing I am working on now?” I call the difficulty of answering that question the meta-access problem: the problem of accessing information because of its relationship to the information currently in the foreground.

My paper proposes hyper-relevant contextual information retrieval (HR-CIR) as a way to address it. “Hyper-relevant” means relevant to your immediate work context, not simply similar in subject matter. The central move is to use the current document, email, task, PDF, note, or other resource as a retrieval cue. Open its context, and the system can show both relationships you deliberately saved and potentially useful relationships it inferred. The resources stay in their own apps.

Start with what is in front of you

Most retrieval tools start by asking something of you. A search engine asks for a query. A bookmark manager asks you to browse or search a collection. HR-CIR starts with the foreground resource: the persistent digital item occupying your attention. It treats that item as a clue to the surrounding work.

Suppose you are reading a research paper as a PDF. Its context might offer your notes, the corresponding Bookends or Zotero record, an email discussion, a cited paper, and a task to follow up on a claim. Some of those connections may be explicit links you made. Others could be recommendations drawn from a DOI, citation relationships, metadata, or other evidence. You can follow one and return without losing sight of the paper that prompted the inquiry.

The same principle applies to writing. Returning to a manuscript after several weeks should give you a practical way back into its working context: outline, reference collection, figures, correspondence, and tasks. The manuscript becomes an access point to the project, not merely one file among many.

This is a form of hypertext across apps. It does not require you to move all your information into a single repository. A note can remain in your preferred notes app, a task in your task manager, and an email in your mail app. What matters is that each resource has a reliable address and that relationships between resources can be made navigable.

Why call this an extension of working memory

When we work on a complex problem, we cannot keep every relevant argument, source, objection, and unfinished task in mind at once. Psychologists K. Anders Ericsson and Walter Kintsch used the term long-term working memory to describe how experts use learned retrieval structures to access relevant knowledge rapidly while performing a task. My proposal is not that a software link gives someone expertise. It is that a link tied to the current task can play an analogous role as an external retrieval cue: it helps a person bring something useful back into active thought at the moment of need.

This matters for more than speed. An interruption to search for a source also asks you to decide where to look, remember a title or keyword, assess results, and reconstruct what you had been thinking. Making the right relationships accessible can reduce that burden and leave more attention for interpreting, comparing, criticizing, and writing.

In my Cognitive Productivity framework, this supports Strategic Surfing: the deliberate use of searching, browsing, skimming, scanning, and contextual retrieval in the service of deeper work. Light, purposeful retrieval is part of serious thinking. It becomes a problem when the mechanics of retrieval repeatedly displace the thinking they are meant to support.

What does consciousness have to do with it?

The title of my paper also invokes consciousness, and that deserves care. I draw on Merlin Donald’s account of consciousness as involving governance over more than the immediate moment. (See for example his excellent book, A Mind So Rare: The Evolution of Human Consciousness.) We maintain projects and lines of thought across interruptions and long stretches of time. Donald also describes how a computer display can serve as a temporary external working memory field as we look, write, and think.

A display by itself shows only what is open. Contextual retrieval can make the larger network around that item accessible again. When a paper brings its notes and sources into view, those external representations can re-enter the current episode of thought. When the links remain usable months later, they can help us resume a longer-running project. Whereas most theories of consciousness emphasize only sensory and working memory, Donald emphasizes intermediate and long-term governance. He emphasizes that external knowledge resources (external symbolic representations) can become part of a longer term conscious governance. This is possible due to what he calls the theoretic evolutionary transition in human consciousness.

However, I do not mean that software becomes conscious, or that it literally enlarges the brain’s memory capacity. My claim is about the effective reach of a person’s conscious thought and control: what relevant knowledge they can bring to bear, when they can bring it to bear, and how readily they can sustain a project over time. That is also a matter of epistemic agency, our capacity to direct our dealings with knowledge in pursuit of our goals. (See my Substack article on epistemic agency

From deliberate links to useful recommendations

The paper distinguishes two kinds of relationships. An explicit link records a connection you consider worth keeping. For example, you might link a reviewer’s email to your manuscript, or a note to the PDF that prompted it. In addition, contextual information retrieval software should be able to make contextual recommendations that you passively are presented with — i.e., it proposes a resource you did not link yourself: perhaps a cited paper, a page sharing a DOI, or a document identified through another app. Whereas web recommendations (e.g., which Google and Amazon show you) are designed to serve the interests of the company, hyper-relevant contextual recommendations serve your knowledge-building needs.

Both belong in the same view, but they should be distinguishable. A system ought to allow you to understand why it suggested something, let you dismiss irrelevant results, and keep you in charge of which relationships become durable parts of your work. A promising direction is for accepted recommendations to become explicit links when you decide they matter.

I designed Hookmark, the macOS app developed by my company, CogSci Apps, as an example of what this architecture can already do. Its context window shows links associated with the current item and has a RELATED section for recommendations. Hookmark is an illustration and a practical demonstration of feasibility; the paper’s broader proposal is a retrieval approach that could be implemented in other ways and on other platforms. I disclose my connection to the product in the paper.

This is a conceptual and design paper, not a controlled study showing that the approach improves memory or consciousness. Research still needs to test it in real reading and writing: not just how quickly people retrieve items, but how well they resume interrupted work, maintain continuity of thought, avoid distraction, and exercise control over their information environment.

The practical idea is simple: a resource you have already found should help you find your way back to the work around it. If our software gets better at restoring that context, it may help us use what we know more effectively and carry our thoughts further across both apps and time.

Both search and contextual information retrieval have their place. What’s stunning however is that by default, desktop and mobile operating systems do not provide contextual information retrieval. Hookmark and link-friendly software solve this problem

Twenty-four developers, professors and leaders co-signed the Manifesto for Ubiquitous Linking, which I wrote in 2020, advocates for software to be contextually aware, which means they must be link-friendly. The manifesto mimics the structure of the famous Agile Manifesto for software development. If you agree with this article, I invite you to sign and share the manifesto.

Read the full article on ResearchGate.

Colophon

ChatGPT co-authored this Substack article with me.

This does not replace the notion of intelligence as IQ (which is extremely useful), but complements it. IQ is not fully predictive of performance and achievement. The conception below could inform revisions of IQ measures.

Why a fourth article? My other posts (listed at the end of this one) set the context but they didn’t succinctly define intelligence. This time I cut to the chase.

Human-like Intelligence

To a first approximation, human-intelligence involves the capacity of an autonomous agent to generate and pursue multiple manifold sources of motivation in real-time with limited temporal and other resources. That is also how I have long defined autonomous agency. Human-like intelligence evinces the following mechanisms and abilities. This is not the full definition however. The following is part of my definition (qua specification) of human-like intelligence.

1. Reactive processing

• Perception, pattern recognition and perception-based action in the world

Ours is more than perceiving the spatiotemporal world, based on a modular system. In contrast AI perceptual systems are typically modular as far as I know (which makes it easier to design). Much mammalian perception, for instance in squirrels, involves a labyrinthine system with parts that have multiple relationships to other parts of the system, as described in Aaron Sloman (1989) “On designing a visual system”.

These sensory-motor mechanisms can

• react to the world in real-time

• generate alarms and motivators

• operate in parallel with management functions

In humans, these mechanisms are more or less under management control as described below. Since my thesis, I have proposed that executive functions are divided in management and meta-management layers.

2. Type-1 Management processing (deliberative executive functions)

(Not to be confused with the typical Type-1 (automatic) and Type-2 (controlled) distinction in cognitive psychology).

• use a rich internal generalized language or languages

• comprehend, reason, plan, prioritize, solve problems and innovate

• handle ambiguity

• improvise

These are also sources of learning (e.g., persisting fragments of plans for future use; decisions about priorities can be persisted for future use; etc.).

3. Type-2 management processing: decision-making and control (executive functions)

• inhibit action

• self-explain

• make decisions of whether, when, and how to do something (e.g., turn plans into possibly conditional decisions)

• enact/execute decisions and control action

4. Meta-management (executive functions)

• observe management processing

• deliberate about management processing, assess and criticize thinking

• decide, interrupt, schedule and control management processing (guide attention, etc.)

• evince cognitive flexibility

• self regulate in many other ways

5. Social and emotional competence

Human intelligence cannot be separated from the mechanisms that enable social and emotional competence with which it co-evolved. This involves mechanisms to

• use public language(s) to communicate.

• perceive and regulate moods and “emotions”

• engage in social signaling and use hierometers

• trade commitments and other resources

• cooperate, divide labor, negotiate and compete

• act (in the theatrical sense)

6. Development and learning

The mechanisms described above are subject to manifold learning and development. Some psychologists draw a sharp line between development and learning. However, from an information-processing perspective, much architectural development is possible throughout the lifespan. Other primates and some modern AI systems can learn in several ways. However as it stands, human learning is more manifold and complex. Humans can

• develop cognitive and motor skills (proceduralization)

• learn to recognize new perceptual and higher-order patterns

• retain information in various forms of memory (short-term memory, associative memory, semantic memory, episodic memory, long-term working memory, etc.)

• form new concepts, symbols and other representations, engaging in Piagetian accommodation (to a first approximation this is purely cognitive, but in a deeper analysis it includes motivational learning)

• build new public knowledge

• develop an understanding of knowledge (which involves developing relationships to knowledge, which overlaps with Piagetian accommodation)

• develop new motivators, new motivator generators and activators

• develop new priorities amongst motivators

• develop new motivator filters and suppressors

• extend their internal generalized languages and learn new public languages

• develop new models, narratives and analogies

Human learning happens on different time scales. their learning however is often remarkably rapid and sample-efficient (contrast The data black hole at the center of AI - by Dwarkesh Patel)

Is this a definition of intelligence?

“Defining” from an integrative design-oriented perspective is different from other forms of defining (such as genus-species definitions, or operationally as in IQ). Design-oriented definitions specify key mechanisms, their functions and organization.

With this in mind, the next time someone tells you AI is or is not intelligent, ask them which of the above mechanisms and abilities they are talking about.

Sketch of an architecture

Here is a sketch of an architecture for human-like intelligence reprised from my previous article. While it is an oversimplification, it does suggest more complexity than typical models of mind (e.g., Freud’s tripartite model of mind or dual process theories which are so common in cognitive psychology).

Relationship to consciousness

Like discussions of intelligence, discussions of consciousness tend to be binary (claiming a type of agent has consciousness or does not) or quantitative (claiming a type of agent has a certain amount of consciousness). In my view consciousness is like intelligence in that it is a polymorphic concept to be interpreted in terms of an information-processing architecture. There is a discontinuousspace of possible minds with different forms of consciousness based on their information processing architectures. I consider “p-consciousness” (“phenomenal consciousness”) to be largely irrelevant. To me it is obvious that different information processing architectures support different forms of p-consciousness.

See for example

• Merlin Donald’s A Mind So Rare: The Evolution of Human Consciousness

• Sloman (2010) Phenomenal and Access Consciousness and the “Hard” Problem: A View from the Designer Stance in The International Journal of Machine Consciousness who wrote:

The diversity of the phenomena related to the concept “consciousness” as ordinarily used makes it a polymorphic concept, partly analogous to concepts like “efficient”, “sensitive”, and “impediment” all of which need extra information to be provided before they can be applied to anything, and then the criteria of applicability differ. As a result there cannot be one explanation of consciousness, one set of neural associates of consciousness, one explanation for the evolution of consciousness, nor one machine model of consciousness. We need many of each. I present a way of making progress based on what McCarthy called “the designer stance”, using facts about running virtual machines, without which current computers obviously could not work. I suggest the same is true of biological minds, because biological evolution long ago “discovered” a need for something like virtual machinery for selfmonitoring and self-extending information processing systems, and produced far more sophisticated versions than human engineers have so far achieved.

• Sloman, A. (2020). Varieties of evolved forms of consciousness, including mathematical consciousness. Entropy, 22(66)

Many of the mechanisms entailed above are elaborated in my thesis and in Sloman (2008) The Cognition and Affect Project: Architectures, Architecture-Schemas, And The New Science of Mind..

About my previous articles on intelligence

1. Beyond “Is AI Intelligent?”: Intelligence as Architecture. This distinguishes four conceptions of intelligence. The first conception is binary (system is said to be intelligent or not intelligent). The second puts intelligence on a one-dimensional continuum. The third views intelligence as values in a multidimensional space. The fourth is architectural, which subsumes but extends the third. Each has its place. I focus on the fourth.

2. If Intelligence Isn’t Binary, What Is It? This article summarizes an architectural view of intelligence and characterizes human-like intelligence, but it did not actually define it. (I rewrote this one from scratch).

3. Beyond “Is AI Intelligent?”: Intelligence as Architecture This one took a deeper stab at defining intelligence in a comparative fashion, providing a couple of helpful tables one of which is pretty long.

What’s up next

• I owe SharpBrains an article on repetitive thought which I’ll publish here.

• I will update the Manifesto for an Integrative Design-oriented Approach to Understanding Humans as Autonomous Agents – CogZest here

• I will comment on the similarities and differences between Agnes Moor’s Goal Directed theory of “emotion” and my own. See for instance her July 2026 “Emotions as High-Impact Decisions: A Goal-Directed Theory”.I am a big fan of her work. We both agree that a goal-directed “eliminativist” view of emotion is required, and we both emphasize goal-directedness (should be obvious from the title of my 1994 thesis, Goal Processing in Autonomous Agents

• I will eventually write an article on consciousness in AI, humans and other primates

My new book Discontinuities: Love, Art, Mind is 95% done and on sale. I have two books in my pipeline but I’m undecided re what my next book should be.

• The Art and Science of Falling Asleep: Understanding Sleep Onset, Insomnolence, and Mental Perturbance

• Comparing Minds: Intelligence and Consciousness in Humans, Animals, and AI

Please let me know in the comments your thoughts about this article and what should come next.

In a previous article I expounded on the CUPA (caliber, utility, potency, appealingness) criteria for assessing assertive artifacts summarizing what I published in my two Cognitive Productivity books. In my previous article, I argued that creativity includes but requires more than Margaret A. Boden’s SIN criteria (surprise, importance, and novelty). In a new chapter ending the Creativity part of my book, Discontinuities: Love, Art, Mind, I combine and extend these two sets of ideas. I adapt the CUPA criteria to assess creative artifacts in general: conceptual and non-conceptual, assertive and non-assertive. The following is a draft of this new chapter. I hope to get feedback on the chapter so that I can improve it before the book is marked fully complete. Discontinuities is currently marked as 95% complete. The remain chapter is the eponymous and final chapter of the book, “Discontinuities”, which argues that the space of possible (artificial and natural) minds is discontinuous, echoing a theme of my recent articles on intelligence. I hope you will enjoy the following chapter, but let me know if you don’t 😉.

In a previous chapter, I argued that Margaret Boden’s SIN criteria—Surprise, Importance, and Novelty—and other theories of creativity in cognitive science I’ve encountered, while helpful, capture only part of what makes creative works memorable. They tell us something important about creative products themselves, but not enough about the relationship between those products and the minds that encounter them. The greatest works of art, literature, music, science, and even engineering do more than satisfy our curiosity or impress us with their originality. They seem crazy (bizarre and dubious), they may feel a bit funny, they not only capture our attention while we are experiencing them (as Brian Boyd aptly argued), they create a perturbance meaning they repeatedly come to mind. The greatest artifacts (assertive or artistic) reorganize our conceptual structures and reshape our motivation —i.e. require accommodation, — inspire new projects, and continue influencing our thinking years after our first encounter with them. Creativity, in other words, is not simply a property of artifacts. It is also a property of what those artifacts do to us: they have psychological effects on the audience.

That observation suggests a broader question. When we recommend a novel to a friend, praise a film, admire a painting, or become excited about a scientific theory we are not merely evaluating its creativity according to SIN criteria. We assess whether it is well crafted, whether it taught us something, whether it changed us, whether it helped us, whether we enjoyed it, and whether we think others ought to experience it.

Over the past twenty years, while working on Cognitive Productivity books and R&D (such as Hookmark), I found myself repeatedly evaluating books, scientific papers, software, lectures, and other knowledge resources using four broad questions. Although I originally developed the CUPA framework below to help knowledge workers make better decisions about what non-fiction to read, study and use, while working on this Discontinuities book I gradually realized that the framework applies just as naturally to works of art qua art. Indeed, one of the themes of this book has been that art itself can be a knowledge resource—not always propositional knowledge, to be sure — but often psychological, emotional, motivational, or existential knowledge. We therefore need a richer vocabulary for discussing its value; more precisely what I referred to as its helpfulness. The CUPA framework developed in Cognitive Productivity books involves four compound categories (not just dimensions) of assessment: Caliber, Utility, Potency, and Appeal. I recommend my Cognitive Productivity books for more details on the general framework. Here I expound its general applicability.

To be clear, each component of CUPA is compound: not merely a single dimension or set of dimension. Cognitive Productivity with macOS: 7 Principles for Getting Smarter with Knowledge makes that abundantly clear particularly for caliber which resists overly simplified ways of assessing scientific theories.

C: Caliber

The first set of criteria, caliber, concerns the quality of the artifact itself. These are the criteria with which traditional discussions of creativity are most concerned. (For assertive artificats, it subsumes but extends the CRAAP criteria often taught to undergraduates.) Is the work original? Is it insightful? Does it display technical mastery? Does it make an important contribution to its field? Does it exhibit elegance, subtlety, coherence, or depth?

Two of Boden’s SIN criteria, discussed above, fit quite naturally within the caliber category. Novelty is obviously an aspect of caliber and reflects whether the a work has achieved something genuinely original. Surprisingness is psychological and thus part of potency, discussed below. The importance of a scientific discovery or an artistic innovation likewise belongs here. In the opening chapter, however, I suggested that caliber extends beyond novelty. Some works become classics not because they were radically unprecedented, but because they achieve extraordinary levels of refinement, coherence and profundity. Shakespeare did not invent every dramatic device he employed, nor did Bach invent harmony. Greatness often consists not merely in novelty, but in realizing possibilities that others had only dimly perceived.

Caliber alone, however, cannot explain why certain works become personally transformative. Two books may be equally well written, yet one quietly returns to our thoughts for decades while the other fades from memory within a week. The difference, I suggest, often lies in what I call potency.

U: Utility

The third set of categories, utility, reminds us that readers, listeners, and viewers are not necessarily passive spectators. They are autonomous agents pursuing multiple goals, solving problems, forming relationships, grieving losses, raising children, designing software, conducting research, and attempting to live meaningful lives. A work may therefore be valuable because it helps someone pursue one or more of these projects. A novel about grief may help someone navigate bereavement. A philosophical essay may clarify a difficult decision. A scientific paper may provide the missing concept needed to solve a technical problem. A song may help regulate one’s emotional state during a difficult period. None of these forms of usefulness diminish the artistic value of the work. On the contrary, one of the recurring themes of Discontinuities has been that stories, music, and other forms of art can become tools for self-understanding and self-transformation when engaged with deliberately and thoughtfully.

Utility in my framework is not a sociological concept (though it can be in other frameworks). It is expected value relative to the individual. I mean personal utility. A treatise on quantum mechanics may possess extraordinary caliber while being almost useless to someone interested only in Renaissance painting. Likewise, a deeply personal memoir may profoundly help one reader while being irrelevant to another. Utility therefore asks not simply whether a work is useful, but useful for whom, and for what purpose.

P: Potency

Potency concerns the psychological consequences of engaging with an artifact. It is the degree of Piagetian accommodation (not mere assimilation) required to grasp the significance of the work. A potent work changes the way we think. Sometimes it forces us to accommodate new concepts rather than merely assimilating new information into existing conceptual structures. Sometimes it leaves us with productive uncertainty rather than tidy conclusions. Sometimes it motivates us to begin an entirely new project or to reconsider long-held assumptions.

Real accommodation requires intellectual effort. It is not for the cognitively miserly as Keith Stanovich discussed in What intelligence tests miss: The psychology of rational thought. It requires changes not merely in what I call comprehension, which is short term familiarity. It can take months, sometimes years, to grasp the full implications of a new theory, novel, or film. It can take months, sometimes years, to appreciate the full implications of a new theory, novel, or film.Charles Darwin’s theory of evolution, Richard Dawkins’s The Selfish Gene, Turing’s last paper, on morphogenesis (accessibly explained by Philip Ball for the The Royal Society), and Aaron Sloman’s The Computer Revolution in Philosophy, are four of the best examples I know.

One cannot do the work required to accommodate to every piece of information we encounter. However, the difference between a fluid expert and a crystallized one is that the fluid expert spends a portion of his time daily, or at least weekly, updating their mental models with potent new information. This is beautifully explained in Surpassing ourselves: An inquiry into the nature and implications of expertise by Carl Bereiter and Marlene Scardamalia. Many people who were experts fossilize. They remain inert, refusing to engage deeply with information that requires effort to accommodate, though they may superficially skim it.

This type of effort typically must come from what psychologist Robert White referred to as effectance, the motivation to improve oneself, which in turn is normally a matter of architecture-based motivation. That is inherent propensity to respond to information by effortful accommodation. (This is discussed in my substack article on Improving Concepts of Epistemic Agency.)

Humor, strangeness, emotional power, ambiguity, and even bewilderment can all contribute to potency—not because they are objectively valuable, but because they increase the likelihood that the work will continue occupying our minds.

Potency does not merely entail dry, cognitive change. It also includes affective change: new motivators, new motive generators, new preferences and new moral values may emerge. War and Peace has turned many readers into pacifists.

Throughout this book (and my career) I have emphasized the importance of mental perturbance: the tendency of mental content to disrupt, maintain and otherwise influence ongoing cognition. As I noted in the opening chapter of this part of the book, great works, like great love affairs and grief, are often potent precisely because they induce such perturbance. They refuse to let us remain intellectually or emotionally unchanged. By recurring repeatedly in one’s minds, one can deeply reshape oneself.

This is why I suggested in the opening chapter of this part of the book that several criteria often associated with creativity are better understood as indicators of potency than of creativity itself. A work that leaves us uncertain for weeks is if not creative at least noteworthy. Nor is a bizarre work, to use Eric Schwitzgebel concept, necessarily a better work. Yet both uncertainty and bizarreness may greatly increase the extent to which the artifact continues interacting with our cognitive architecture. From a design-oriented perspective, this continuing interaction is part of what makes the work potentially helpful.

A: Appeal (appealingness)

The final set of criteria concerns appealingness. Many academics are understandably suspicious of appeal because popularity is such an unreliable guide to quality. Yet appeal cannot simply be ignored. No artifact can influence anyone unless it first attracts attention. Attention, as we have seen repeatedly throughout this book, is the gateway to learning, emotional engagement, conceptual change, and self-transformation. An artifact may possess extraordinary caliber and enormous transformative potential, but if nobody is willing to engage with it, its influence will necessarily remain limited. Appeal concerns the capacity of a work to invite attention and sustain engagement long enough for its caliber, utility, and potency to exert their effects.

Seen in this light, Boden’s SIN framework finds a natural home within a broader evaluative framework. Novelty, surprise, and importance remain indispensable characteristics of many great creative works, but they largely describe one aspect of a work’s value—its caliber. Other characteristics discussed in the previous chapter, including conceptual accommodation, productive uncertainty, humor, emotional resonance, and persistent cognitive occupation, contribute more directly to potency. Utility reminds us that works interact with people pursuing real projects in the real world, while appeal recognizes the indispensable role of attention in every form of human learning and transformation.

The primacy of wholistic evaluation

These sets of criteria of overall helpfulness of an artifact must not be taken in isolation. They must be integrated with each other. Not every transformation (potency) is desirable. Reading Mein Kampf for instance turned many readers into lunatics. Moral standards, the first criterion, applies not only to the evaluation of the work, but the person that one would become in response to it. Some works must be considered as repugnant. They must trigger disgust. Some works are seductively appealing but counter-productive. Alternatively, the truth is not always beautiful. Darwin recognized that his theory while true entailed an ugliness in nature. The first two set of criteria — praiseworthiness and desirability — must normally dominate the second evaluation. One needs to adapt carefully (potency) to artifacts – not allowing the theory of evolution for instance to affect our assessment of the worth of other human beings.

The Cognitive Structure of “Emotions” Ortony, Clore & Collins (1988/2022)

Surprising as it may seem, the CUPA framework actually draws on a theory of emotions. I will take the shortcut of quoting from above:

According to Ortony, Clore and Collins, in their book, The Cognitive Structure of Emotions, there are three dispositional kinds of affect each yielding its respective type of emotional occurrence:

1. Standards or norms, which generate assessments of the praiseworthiness of agents’ actions, yielding attribution emotions.

2. Goals, which generate assessments of the desirability of events, yielding event-based emotions.

3. Tastes (also known as attitudes, and dispositional liking and disliking) that generate assessments of the appealingness of objects (episodic liking and disliking). These assessments are called attraction emotions. We can generalize these to include pain and pleasure mechanisms which generate, as you would expect, feelings of pain and pleasure.

You can thus see that the Caliber criteria of CUPA map onto Ortony’s praiseworthiness criterion. The “Utility” criteria map onto Ortony’s desirability criterion. The appealingness criteria are, well, Ortony’s appealingness criterion applied to responses to artifacts.

I acknowledge however, that the mapping between Ortony, Clore & Collins’s dimensions and the CUPA framework is not perfect. For example, praiseworthiness in their theory is usually applied to agents’ actions, whereas caliber in mine applies to artifacts. So the CUPA framework generalizes their framework.

Leveraging their framework is a fittingly strange loop (a recurring feature of this book), reflecting the caliber, utility, potency and appeal of The Cognitive Structure of Emotions theory, which is deservedly famous in affective sciences.

Conclusion

The CUPA criteria don’t replace the creativity criteria mentioned above. They include and extend them. Artists, scientists, engineers, teachers, and software designers all create things intended to interact with human minds. Creativity remains one hallmark of excellence, but it is not the final objective. The final objective is to create artifacts that matter: artifacts of high caliber, capable of helping people pursue their projects, potent enough to reshape their thinking, and appealing enough to earn their sustained attention.

Having said that, I expect some readers will be inclined to define creativity in terms of the CUPA criteria. A rose by any other name…

Here follows another chapter on creativity from my book Discontinuities: Love, Art, Mind on which I’ve been working on-and-off since… 2012! (See my First [2012] chapter.) Now, I’m delighted to finally complete a chapter that consolidates many of the psychological criteria underlying judgments of creativity that have occupied my mind for years, based on reading several very potent books on the subject.

Margaret Boden’s The Creative Mind: Myths and Mechanisms remains the most insightful book I’ve read on creativity. Rather than treating creativity as an inexplicable gift possessed by a fortunate few, Boden analyzes it as a phenomenon that can be understood scientifically. She argues that genuinely creative products possess three defining characteristics: they are surprising, important, and novel. I remember these three properties with the acronym SIN. It is an elegant characterization. Creative products are not merely different from what came before; they also matter, and they violate our expectations in interesting ways.

Boden is certainly not alone in emphasizing usefulness or importance as an essential component of creativity. Keith Simonton, for example, has long argued that creativity cannot be reduced to originality alone: ideas must also prove valuable or useful within their domain. See, for example, Keith Simonton’s Creativity in Science: Chance, Logic, Genius, and Zeitgeist. Throughout her book, Boden further argues that creativity is constrained rather than random: it arises through the exploration or transformation of conceptual spaces rather than through arbitrary originality.

Image courtesy of ChatGPT.

I believe, however, that SIN captures only one important aspect of what makes the greatest creative works memorable. Boden’s criteria characterize the creativity of products. I would like to complement this perspective by focusing on the relationship between the product and the cognitive architecture encountering it. Creativity is not simply a property of an idea, theory, painting, novel, or scientific discovery. It is also a property of what that artifact does to the mind that encounters it. The greatest creative works do more than surprise us. They reorganize our conceptual structures, challenge our assumptions, alter our motivations, recruit and sustain our attention, and continue occupying our thoughts long after the initial encounter. They can even leave us uncertain about what to believe or how to proceed, precisely because they stretch beyond our existing conceptual frameworks. In short, creativity is not merely about producing something new. It is also about producing something capable of changing another mind.

Thus, the first additional property I would add is potency. I discussed this at length in my two Cognitive Productivity books and elsewhere in this book. (Just search for “potency” in this book.) By potency I mean the degree of Piagetian accommodation required before the observer can genuinely understand the work. Some new ideas can be assimilated (in Piaget’s sense) into existing conceptual structures with little effort. We simply add another fact to what we already know. But highly creative works resist assimilation. They require accommodation. We must reorganize our existing mental models before the work itself becomes intelligible. Einstein’s theory of relativity, Darwin’s theory of evolution by natural selection, cubism, atonal music, Joyce’s Ulysses, and Gödel’s incompleteness theorem are all potent in this sense. Their creativity lies not merely in their novelty but in the conceptual restructuring they demand of the audience. Potency therefore measures something different from novelty: it measures how profoundly an artifact transforms the conceptual organization of the observer.

A second characteristic of many creative works is what philosopher Eric Schwitzgebel calls craziness. In a fascinating discussion on the Rationally Speaking podcast, Schwitzgebel carefully distinguishes between three related dimensions of ideas: dubiety, bizarreness, and craziness. A claim is dubious when the available evidence does not epistemically justify believing it. A claim is bizarre when it is highly contrary to common sense. Finally, a claim is crazy when it is both bizarre and dubious—that is, when it is highly contrary to common sense and we are not yet epistemically compelled to accept it. This distinction is extremely useful because it reminds us that apparent craziness is not the same as falsehood. Many of history’s greatest scientific and artistic innovations initially appeared crazy precisely because they challenged deeply entrenched conceptual frameworks before sufficient evidence had accumulated. Continental drift, heliocentrism, the theory of relativity, the theory of evolution, and germ theory all looked crazy long before they became accepted science.

Closely related is what might be called the “That’s funny…” phenomenon. In Inside Jokes: Using Humor to Reverse-Engineer the Mind, Matthew Hurley, Daniel Dennett, and Reginald Adams argue that humor often arises when the mind suddenly abandons one interpretation in favour of another after detecting an inconsistency in its current mental space. Drawing on Gilles Fauconnier’s theory of mental spaces, they describe a mental space as a temporary conceptual structure constructed during comprehension that organizes currently relevant concepts, percepts, assumptions, and inferences. Mental spaces are continually constructed, revised, and sometimes abandoned as new information arrives. Humor arises when an inconsistency is detected within the current mental space, prompting the rapid construction of a superior interpretation. Scientific creativity frequently begins in much the same way. Isaac Asimov famously observed that the most exciting phrase in science is not “Eureka!” but “That’s funny….” Whether or not Asimov actually coined the quotation, it captures something profound about discovery. Creative breakthroughs often begin with anomalies—observations that stubbornly refuse to fit our existing understanding. Before there is explanation, there is puzzlement. Before there is theory, there is the nagging feeling that something is wrong with our current mental space. The inconsistency need not be between two conscious beliefs. Rather, new information often conflicts with assumptions or expectations that remain outside conscious awareness. Humans appear to be exquisitely designed to detect such inconsistencies, and the pleasurable experience of humor arises when a richer mental space suddenly replaces the old one. (I highly recommend Inside Jokes.)

Brian Boyd approaches creativity from a complementary perspective. In On the Origin of Stories: Evolution, Cognition, and Fiction, discussed elsewhere in this book, discussed elsewhere in this book, he argues that propensity to create artistic work evolved under selective pressures to capture, direct, and reward attention. Stories, paintings, music, and other works of art compete for our limited cognitive resources, and those that better engage our attention are more likely to be appreciated, remembered, and transmitted culturally.

I would extend Boyd’s account one step further using an architectural theory of mind based on Sloman’s H-CogAff but which I significantly expanded. The greatest creative works do not merely capture our attention while we are experiencing them. They continue occupying our attention afterwards. They generate what Sloman and I have called perturbance: an information processing state in which unresolved motivators repeatedly compete for attention because they have acquired high insistence. We replay scenes from a novel days later. We continue turning over a philosophical argument while driving home. We awaken during the night thinking about a scientific puzzle or an unresolved piece of music. The artifact has become motivationally active within our cognitive architecture. It repeatedly interrupts ongoing cognition because unresolved concerns continue recruiting management and meta-management processes. In this sense, great creativity is not merely attention-grabbing; it is perturbance-inducing. It generates insistent motivators that remain cognitively active until they are integrated, resolved, or abandoned. See the very accessible paper I coauthored with Sylwia Hyniewska and Monica Pudlo.

This perspective also helps explain why immediate popularity is often a poor measure of creativity. Boyd is surely right that creative works must ultimately earn attention if they are to survive culturally. But the most potent works are frequently misunderstood when they first appear because they require more accommodation than contemporary audiences are prepared to undertake. Einstein’s 1905 papers did not immediately revolutionize physics. Van Gogh sold almost none of his paintings during his lifetime. Joyce bewildered many early readers. Their creativity did not consist in immediate popularity but in their remarkable capacity to reorganize the minds of those willing to engage with them deeply enough.

Keith Simonton argues that chance plays a fundamental—but often misunderstood—role in scientific creativity. Creative breakthroughs rarely emerge from straightforward logical deduction alone. Instead, scientists generate many tentative combinations of ideas, analogies, methods, and observations, only a few of which prove fruitful. Chance enters because no one can reliably predict in advance which combinations will lead to major discoveries. Yet this is not blind randomness: logic, expertise, and historical context guide the generation, evaluation, and refinement of ideas. Creativity thus depends on a disciplined search in which serendipitous combinations are recognized, developed, and ultimately integrated into scientific knowledge. I believe this also applies in other domains of human creativity.

Angus Fletcher (author of Wonderworks) offers a complementary perspective by arguing that creativity is closely tied to uncertainty rather than originality alone. Because it is notoriously difficult for experts to judge whether a new idea is genuinely creative, Fletcher proposes asking a different question: How uncertain are you that this solution will work? A highly creative idea is one that experts cannot confidently classify as either correct or incorrect because they have not encountered anything quite like it before. If they are uncertain, the idea likely contains genuine novelty; if they are also unwilling to dismiss it, it has creative potential. Rather than treating uncertainty as a flaw, Fletcher suggests that it is often a hallmark of innovation and a signal that exploration should continue rather than cease. This perspective aligns well with Simonton’s emphasis on chance: creativity often involves venturing into regions where outcomes cannot be predicted in advance. This idea is not in his book. Instead, Fletcher discusses this idea in The Uniting Power of Story (YouTube, Episode 205)., where he also connects uncertainty to the distinctive power of stories to help us navigate an unpredictable world.

Ingenuity

Another characteristic of many great creative works is what might be called ingenuity. We recognize that some creative artifacts could not have been produced by an ordinary person. They bear unmistakable evidence of extraordinary intelligence, expertise, imagination, or technical mastery. The audience is left wondering, How could anyone have thought of that? or How could anyone possibly have created that? This reaction differs from mere surprise. We are not simply surprised by the product; we are astonished by the remarkable cognitive achievement that must have been required to produce it.

This is especially evident in music. The compositions of J. S. Bach, the late string quartets of Beethoven, the orchestral works of Stravinsky, or the improvisations of Charlie Parker all convey an extraordinary degree of ingenuity. Even listeners with little formal musical training often recognize that these works embody a level of craftsmanship far beyond ordinary competence. Part of their creative power lies in the audience’s appreciation of the exceptional cognitive abilities required to conceive and execute them.

A striking contemporary example is the Quebec experimental music collective Angine de Poitrine. Their performances are simultaneously bewildering, captivating, technically astonishing, and emotionally compelling. Watching both the performances and the many reaction videos posted online, one repeatedly encounters comments expressing not merely surprise but amazement at the performers’ apparent abilities. Viewers frequently remark that the music seems almost impossible to reproduce, that the performers must possess extraordinary musical intelligence, and that the work repeatedly draws them back for further listening despite—or perhaps because of—its initial difficulty. The live performance at the Electric Ballroom in London provides an excellent illustration. The work does not simply command attention during performance; it often continues recruiting thought afterwards as listeners attempt to understand how such music could have been conceived and performed.

Through their costumes, stage presence, and deliberately uncanny performance style, Angine de Poitrine also satisfy the funny criterion. Their visual presentation continually encourages the audience to reconstruct its mental spaces in response to unexpected juxtapositions, making the performance psychologically as inventive as the music itself.

Conclusion

Boden’s SIN framework remains an outstanding foundation for understanding creativity. I would simply augment it with a more explicitly architectural perspective. The greatest creative works are not merely surprising, important, and novel. They are also potent enough to require accommodation, crazy enough to challenge prevailing assumptions before the evidence becomes compelling, funny enough to begin with anomalies that provoke curiosity, and perturbing enough to generate insistent motivators that continue to occupy the mind long after the initial encounter. It may leave us uncertain about the utility of the idea. Chance co-determines the significance of creations. Truly creative work also are also perceived as the result of tremendous ingenuity. Creativity, on this view, is not merely about producing something new. It is about producing something that changes the organization of another mind—and keeps changing it until a new equilibrium is reached.

David has won four Juno Awards, and has been nominated for four others. His Skating Rink is the theme song for Hockey Day in Canada. His songs are simple, precise, elegant, powerful and exquisitely well written.

The following is a new chapter from my book Discontinuities: Love, Art, Mind discussing the interview. It is drawn from the part of this book that deals with creativity:

David Francey on Creativity

One of the pleasures of writing this book has been the opportunity to revisit conversations that have stayed with me for years. In July 2013, I interviewed Canadian singer-songwriter David Francey while he was at the Mission Folk Music Festival. At the time, my intention was simply to publish the interview. Looking back more than a decade later, however, I realize that a synopsis of our conversation belongs in this book because it illuminates a question that has occupied me for much of my career: What actually happens in the mind of someone who creates?

David and I approach that question from opposite directions. I came to it as a cognitive scientist interested in expertise, learning, emotion, and motivation. David came to it as someone who had spent decades writing songs—first while working construction, and later as one of Canada’s most celebrated folk musicians. What struck me throughout our conversation was not simply the quality of his songwriting but the clarity with which he understood his own creative process. Although he repeatedly insisted that he “just does it,” his descriptions reveal principles that resonate strongly with what psychologists have learned about creativity.

One theme emerged almost immediately. David does not begin by looking for songs. Songs begin by finding him.

When we discussed Things They Do, he described hearing a CBC radio report about two fourteen-year-old girls who had died by suicide by stepping in front of a train. The story stayed with him while he worked throughout the day. As he continued roofing, another memory resurfaced: a close friend who had struggled with severe depression and had eventually taken his own life. The two experiences became intertwined until he felt compelled to write.

Listening to David describe this process, I was reminded of the theory of perturbance, which has pervaded my thinking about mental life. According to that theory, certain thoughts and motivators become insistent, meaning they repeatedly interrupt our thinking, sometimes because they remain unresolved. Love, grief, guilt, fear, and anger all have this quality. They refuse to stay politely in the background. David never used this language, but he described exactly this phenomenon. Certain experiences simply would not leave him alone until they had found expression in song.

Perhaps the most revealing part of our conversation came when I asked why he writes at all. His answer had nothing to do with audiences, albums, or even music. He explained that writing helped him understand what was happening inside himself. If something troubled him, he would brood over it for days. His wife would eventually tell him that she hoped he finished the song soon because he had become unbearable to live with. Then he added something that immediately caught my attention: once the song was written, he felt better. Writing was not merely the production of art; it was a way of making sense of experience.

Psychologists today speak of expressive writing and narrative construction as ways of organizing emotional experience. David arrived at essentially the same conclusion through lived experience. His songs were not simply expressions of emotion. They were tools for thinking.

Another aspect of his creative process that impressed me was his extraordinary patience. He rejected the idea that every creative problem should be solved immediately. Sometimes a song emerged almost effortlessly, but often he would write a promising first line and then wait. If the next word took two years to arrive, then so be it. He had no interest in forcing the process. Instead, he trusted that his mind continued working on the problem even when he was consciously occupied with something else. To explain this, he compared songwriting to doing a crossword puzzle. You can struggle unsuccessfully with several clues before going to bed, only to discover that the answers seem obvious the next evening. Something has happened in the meantime, even though you were not consciously thinking about the puzzle. His account is one of the clearest descriptions of creative incubation I have encountered from an artist.

David also made an observation that particularly interested me because it connects creativity with the physical environment. Many of his songs were composed while he was in motion—driving to work, travelling on trains, or spending long days roofing houses. Roofing eventually became so familiar that much of it required little deliberate thought, leaving his mind free to work on melodies and lyrics. He even suggested that movement itself helped ideas emerge. Looking through the windshield, seeing one image after another unfold, somehow primed the creative process. That observation resonated with research showing that changing environments and engaging in physical movement often promote creative thinking, although David reached that conclusion simply by paying attention to his own experience.

One idea surfaced repeatedly throughout our interview, regardless of the question I asked. Whether we were discussing songwriting, editing, or influences, David kept returning to the same principle: never overwrite.

He spoke admiringly of Joni Mitchell because, in his view, there is not a wasted word anywhere in her songs. He teaches aspiring songwriters the same lesson. If a line is beautiful but does not belong in the song, save it for another one. The temptation to include every good idea is one of the greatest dangers facing writers. Good songs, like good theories, achieve their power not through abundance but through economy.

That observation particularly appealed to me because it extends well beyond music. Scientists seek parsimonious theories that explain as much as possible with as little as necessary. Engineers strive for elegant designs that eliminate unnecessary complexity. Good writers remove words that do not contribute to meaning. Creativity often depends less on adding than on removing.

Looking back on our conversation today, I think that is what impressed me most about David Francey. His account of creativity contains very little mystique. Inspiration certainly matters, but it is only the beginning. The real work consists of paying close attention to experience, allowing ideas time to mature, refining them patiently, and resisting the temptation to say more than needs to be said. Those principles are as applicable to science and engineering as they are to songwriting. They also illustrate one of the central themes of this book: art provides one of our richest windows into the architecture of the human mind.

In the interview, David provides us with insights into the creation of several specific songs including:

• So Say We All

• Lucky Man

• Things They Do

• Mill Towns

• Hammers from Far End Of Summer

• Hard Steel Mill

Many productive minds have had to overcome adversity. The interview uncovers a surprising antagonist in David’s early life, someone who tried to stunt his creative development. David’s is a story of triumph!

Debates about artificial intelligence often ask whether AI is—or is not—intelligent. Other approaches treat intelligence as a single quantity or as a profile of several measures of abilities. Each approach can be useful, but none adequately explains how intelligent behaviour is produced. None of them sufficiently captures what it means to be intelligent.

This article presents a fourth conception: intelligence as a property of information-processing architectures. Intelligent behaviour arises from organized systems of mechanisms, representations, memories, motivational processes, communication pathways, and layers of control. These architectural properties include reactive and deliberative processing, management and meta-management, working memory, long-term working memory, alarms, self-reminding, learning, explanation, social information processing, and the integration of external cognitive resources.

This article is a condensed synthesis of two longer essays: “Why You Can’t Say AI Is—or Is Not—Intelligent” and “If Intelligence Isn’t Binary, What Is It?”. Readers who want the fuller argument, its intellectual history, additional examples, and a more detailed comparison of humans, other animals, and AI should consult those articles.

Contents

1. The binary conception

2. The scalar conception

3. The multidimensional conception

4. The architectural conception

5. From abilities to architectures

6. The architecture of human-like intelligence

7. Learning as architectural change

8. Humans, other animals, and contemporary AI

9. Intelligence reconsidered

10. From intelligence to consciousness

Introduction

Debates about artificial intelligence often revolve around a deceptively simple question:

Is AI intelligent?

Some people answer yes, pointing to the ability of contemporary AI systems to write, reason, translate, program, explain difficult ideas, and solve problems that once seemed to require human intelligence. Others answer no, arguing that these systems merely manipulate patterns, predict tokens, or simulate capacities they do not genuinely possess.

The disagreement partly reflects different assumptions about what intelligence is. Some people treat intelligence as something a system either possesses or lacks. Others treat it as a quantity: a person, animal, or machine can be more or less intelligent. Still others regard intelligence as a collection of capacities, such as language, planning, memory, creativity, social understanding, and learning.

Each of these conceptions captures something useful. None, however, provides a sufficiently deep explanation of intelligent behaviour.

This article proposes a fourth conception. Intelligence is more fundamentally a property of information-processing architectures: organized systems of mechanisms, representations, memories, communication pathways, evaluative processes, and layers of control. Intelligent behaviour arises not from a single faculty, nor merely from a collection of abilities, but from the organization and interaction of these architectural components.

From this perspective, the most illuminating question is not simply whether a system is intelligent. It is:

What kind of information-processing architecture does the system possess, and how does that architecture produce, regulate, and develop intelligent behaviour?

This approach does not require us to abandon IQ tests, benchmarks, multidimensional profiles, or practical classifications. Rather, it places them within a broader explanatory framework. Performance provides evidence about intelligence; architecture helps explain that performance.

A common definition of intelligence

Here’s a definition of human intelligence which can be measured, taken from Wikipedia:

Human intelligence is the intellectual capability of humans, which is marked by complex cognitive feats and high levels of motivation and self-awareness. Using their intelligence, humans are able to learn, form concepts, understand, and apply logic and reason. Human intelligence is also thought to encompass their capacities to recognize patterns, plan, innovate, solve problems, make decisions, retain information, and use language to communicate.

Now let’s look at four conceptions of intelligence which more or less fit this mold.

Four ways of thinking about intelligence

1. The binary conception

The binary conception treats intelligence as an all-or-none property. A creature or system either is intelligent or it is not.

Binary judgments can be useful when a practical decision requires a threshold. We may need to decide whether a person has sufficient cognitive capacity to give informed consent, whether someone can drive safely, or whether a robot can navigate independently in a specified environment.

But binary classifications suppress enormous differences among the systems placed on either side of the threshold. Calling both a crow and a human intelligent tells us little about the very different ways their minds are organized. Saying that a calculator is not intelligent tells us little about its remarkable competence in symbolic arithmetic. Saying that LLM-based AI is or is not intelligent tells us little about how it really compares to human intelligence.

A binary judgment may tell us whether a system meets a chosen criterion. It does not provide a satisfactory general theory of intelligence.

2. The scalar conception

The scalar conception replaces the yes-or-no question with a quantitative one:

How intelligent is this person or system?

Psychometric intelligence research made this question empirically tractable. Standardized tests allow performances to be compared with those of other people, and overall scores can summarize evidence from several cognitive tasks.

Psychometric work on IQ is scientifically important. IQ is neither a meaningless number nor simply an obsolete attempt to rank people. Intelligence testing is one of psychology’s most extensively developed measurement traditions, and intelligence-test performance predicts outcomes that matter.

Nevertheless, a summary score necessarily compresses information. Two people with similar overall IQ scores may have different profiles of verbal comprehension, visual-spatial ability, fluid reasoning, working memory, and processing speed. Even people with similar profiles may employ different strategies, possess different knowledge, and respond differently to novelty, uncertainty, or distraction.

The limitations become still clearer when scalar measures are extended beyond the human populations for which they were designed. Assigning an IQ-like score to a crow, an octopus, a great ape, or a large language model may produce an arresting comparison, but the number can conceal profound differences in what these systems can do and how their abilities are produced.

A scale may compare performance on a defined set of tasks without showing that the systems possess intelligence of the same kind.

3. The multidimensional conception

The multidimensional conception asks:

In what respects is this person or system intelligent?

Instead of representing intelligence with one number, it describes a profile of quantitative measures of capacities. These might include language, memory, planning, spatial reasoning, creativity, learning, social understanding, metacognition, and sample efficiency.

This is a substantial improvement when comparing very different kinds of systems. A crow may excel at tool use and causal reasoning while lacking human language. An octopus may possess extraordinary sensorimotor adaptability through an organization of control radically different from that of vertebrates. A large language model may produce fluent explanations and write software while lacking continuous bodily perception, autonomous motivation, and an enduring personal history.

It would be misleading simply to declare one of these systems more intelligent than another. A multidimensional profile preserves important differences that a single score compresses.

Yet multidimensional approaches usually remain descriptions of how much of each measured ability a system displays. Two systems may obtain similar planning scores while relying on entirely different processes. One may explicitly represent alternative futures, evaluate them against competing concerns, schedule intermediate actions, and revise its plan when circumstances change. Another may generate an equally successful answer by exploiting learned statistical regularities.

At the level of observable performance, the systems may look similar. At the level of underlying organization, they may be radically different.

4. The architectural conception

The architectural conception changes the main questions:

What information-processing architecture gives rise to the system’s intelligent behaviour? How do two agents or classes of agents differ in their underlying autonomous agency?

An information-processing architecture is not simply a list of abilities. It is the organized arrangement of mechanisms through which a system perceives, remembers, evaluates, learns, reasons, plans, acts, and regulates itself over time.

Architectures can differ in at least four broad ways:

1. They may contain different layers of processing, including reactive, deliberative, managerial, and meta-managerial processes.

2. They may possess different mechanisms within those layers, including attention control, working memory, planning, scheduling, conflict resolution, self-reminding, strategy selection, and ambiguity detection.

3. They may differ in their communication pathways. An alarm signal might influence only immediate action, interrupt or otherwise influence ongoing executive processes, or reach meta-management and initiate a broader reassessment.

4. They may use different representations and algorithms even when they accomplish similar tasks. They may depend on symbolic structures, vector representations, production rules, neural networks, episodic simulations, narratives, other mechanisms, or combinations of these.

Some architectural differences are quantitative. Two systems with working memory may differ in capacity, duration, precision, or reliability. Two planning systems may differ in how many alternatives they can represent or how far ahead they can search.

Other differences are qualitative. An architecture that can schedule deliberation possesses a mechanism that is absent from a system that deliberates only when externally prompted. A system with meta-management does not merely have more reflection than one without it; it possesses an additional level of organization through which some of its own management processes can be evaluated and redirected.

The introduction of a new mechanism, layer, representation, or communication pathway may make entirely new forms of intelligent activity possible. These changes create genuine discontinuities in the space of possible minds. The differences among humans, various animals, and artificial systems are therefore not always differences of degree. They may be differences in kind, arising from the presence, absence, or organization of architectural mechanisms. I explore the broader significance of such qualitative transitions in Discontinuities: Love, Art, Mind.

Table 1. Comparing the four conceptions

The four conceptions are not mutually exclusive. Binary classifications support practical decisions. Scalar measures summarize performance. Multidimensional profiles reveal relative strengths and weaknesses. An architectural account can incorporate all three while asking a deeper question: what underlying organization produces the classifications, scores, and profiles?

5. From abilities to architectures

The architectural approach belongs to an established tradition in artificial intelligence and cognitive science. Aaron Slomanargued that these disciplines should investigate the space of possible minds: the vast range of information-processing architectures that could exist in animals, humans, machines, and systems unlike anything yet encountered.

This perspective shifts attention away from searching for a single defining property of intelligence. Different minds may contain different mechanisms, organize them into different layers, use different representations, and connect their components in different ways.

The same architecture can also operate differently from moment to moment. In an emergency, an alarm may interrupt ongoing thought, redirect attention, increase the priority of an urgent concern, and initiate rapid action. While writing an essay, the same person may recruit semantic and episodic memory, analogy, planning, ambiguity detection, narrative construction, and extended deliberation. During rumination, persistent motivators, narrowed attention, repeated retrieval, and failures of disengagement may dominate.

The broad architecture remains relatively stable, but the configuration of mechanisms currently engaged changes.

This dynamic organization is one reason familiar divisions such as Daniel Kahneman’s distinction between fast and slow thinking, although valuable, are insufficient as general accounts of intelligence. Human information processing involves more than two systems. It includes many mechanisms operating at different levels, interacting through complex pathways, and becoming active in different combinations.

An architectural account also encourages an integrative, design-oriented approach. Rather than studying memory, reasoning, motivation, affect, and executive control as if they belonged to independent systems, this approach asks how all of them must operate together within a complete autonomous agent.

By autonomous agency, I mean the capacity of a system to generate, evaluate, prioritize, pursue, suspend, revise, and abandon its own motivators while responding adaptively to changes in its environment and internal state, given limited resources (time, working memory capacity, other agents, etc.).

Explaining such agency requires more than explaining cognition narrowly understood. It requires understanding how perception, motivation, memory, learning, affect, deliberation, management, and action are coordinated over time.

[Insert Figure: Human-like Autonomous Agency Architecture]

6. The architecture of human-like intelligence

No definitive architecture of human-like intelligence yet exists. Nevertheless, we can identify several broad classes of mechanisms that a satisfactory account will need to explain.

6.1 Reactive processing, management, and meta-management

Reactive mechanisms connect situations relatively directly with responses. They allow a system to recognize familiar conditions and respond rapidly without constructing and comparing elaborate alternatives. This includes motive generators that operate asynchronously to executive processes (management and meta-management).

Management processes perform deliberative activities. The can evaluate motivators and situations and make decisions about whether, when and how to act. Deliberative processing supports the representation of possibilities. It allows an agent to consider actions that have not yet occurred, predict consequences, compare alternatives, and construct plans.

Meta-management adds another level. It includes but goes beyond what is sometimes called reflective processing. It allows some management processes themselves to become objects of evaluation. A person may notice not only that a problem remains unsolved, but that they are approaching it unproductively. They may recognize that they are fixated on one representation, avoiding an uncomfortable possibility, repeatedly checking the same evidence, or terminating inquiry too quickly.

Management and meta-management overlap with what psychologists often call executive functions. But executive functioning should not be treated as one more dimension beside memory or language. It refers to mechanisms that help organize and regulate the use of many other capacities.

6.2 Motivation, evaluation, and control

Intelligent agents do not merely process information; they must determine what matters.

Human motivation is not adequately described by goals alone. People are influenced by projects they want to complete, norms they believe should be respected, and preferences—also known as attitudes—concerning what they like or dislike. Motives are goals that have intensity (behavioral propensity) and insistence (attentional propensity), and valence. There are different states of motives: wishes, wants and intentions. Motivators compete for attention, planning, and action.

Architectural mechanisms must evaluate motivators according to such properties as importance, urgency, opportunity, cost, and compatibility with other commitments. Motivators don’t merely differ in “expected utility”, which is a quantitative measure, but in many qualitative ways. Some become sufficiently insistent to interrupt or otherwise influence ongoing executive processes. Others are deferred, suppressed, forgotten, or abandoned. Hence human-like information processing architecture contains filters and suppressors of motivators.

Human-like autonomous agency depends partly on how these processes are governed. A highly capable person may reason poorly if attention is repeatedly captured by irrelevant concerns, if important motivators fail to become insistent, or if short-term impulses continually displace long-term projects.

Intelligence concerns what an architecture is capable of doing. Rationality also concerns whether, when, and why those capacities are appropriately deployed. See What intelligence tests miss: The psychology of rational thought.

6.3 Alarms, attention, and interruption

Alarm mechanisms are a particularly important architectural class. They detect situations requiring rapid reassessment and can redirect processing before slower deliberation has run its course.

Architectures may differ in where alarm signals travel. An alarm might trigger an immediate reaction, influence management, or reach meta-management and cause an agent to reconsider a broader pattern of activity.

Alarms are not themselves emotions. Fear, anxiety, anger, grief, and other affective phenomena may emerge from interactions among alarm mechanisms, motivator processing, attention, memory, bodily changes, management, and action.

Mental perturbance, for example, can occur when insistent motivators repeatedly capture attention and disrupt other activities. The perturbance is an emergent condition of the architecture, not a mechanism located in one dedicated emotional module.

6.4 Memory, self-reminding, and governance across time

Human-like agency requires more than the storage of information. It depends on bringing the right information to mind at the right time.

Working memory makes some information temporarily available for current processing. Long-term memory preserves semantic knowledge, experiences, procedures, and associations. Long-term working memory, a concept developed by K. Anders Ericsson and Walter Kintsch, allows experts to use learned retrieval structures to gain rapid access to information stored in long-term memory while a task unfolds. It’s long-term memory retrieval from which approaches the speed of retrieval from working memory.

A chess master, scientist, musician, physician, or experienced writer does not maintain everything relevant to a complex activity in working memory simultaneously. Instead, organized knowledge and retrieval cues allow relevant information to become available as it is needed. Long-term working memory is therefore not simply a larger working-memory container. It is a learned architecture of access.

Self-reminding is also essential. An autonomous agent must recover deferred intentions, notice emerging opportunities, remember commitments, and resume interrupted projects. A person who forms excellent plans but never retrieves them when they become relevant will not behave intelligently over extended periods.

These mechanisms help connect the present to the future. They allow people to govern activity not only from moment to moment but across days, years, and sometimes an entire life.

In A Mind So Rare: The Evolution of Human Consciousness, Merlin Donald distinguishes levels of conscious integration. Immediate sensory awareness and short-term control are supplemented in humans by forms of intermediate- and long-term consciousness. These allow intentions, narratives, commitments, and projects to organize behaviour across periods far exceeding the duration of immediate awareness.

Intermediate- and long-term consciousness depend on long-term memory, long-term working memory, prospective memory, self-reminding, contextual retrieval, and external symbolic resources. An intention formed today may need to influence activity tomorrow, next year, or decades later. Human-like autonomous agency depends on bridging the discontinuities that occur whenever attention moves elsewhere or consciousness is interrupted.

A language model’s context window bears a limited functional resemblance to working memory, but the two should not be identified. Human working memory is embedded in continuing perception, action, motivation, autobiographical memory, and self-directed agency.

Human intelligence thus involves different forms of memory and different time-spans of consciousness.

6.5 Social and cultural architecture

Human intelligence is not merely the intelligence of an isolated biological organism.

Humans represent relationships, obligations, commitments, reputations, norms, institutions, and shared projects. These representations support cooperation, division of labour, shared intentions, teaching, cumulative culture, and organizations that continue beyond the lives of their individual members.

Social signalling is an important part of this architecture. Much human behaviour communicates information about competence, trustworthiness, loyalty, status, identity, intentions, and group membership. Kevin Simler and Robin Hanson’s The Elephant in the Brain emphasizes that many activities have hidden social functions, including signalling desirable characteristics to others. Will Storr’s The Status Game examines the pervasive role of status seeking, recognition, and social comparison in human life.

Whether or not one accepts every evolutionary claim in these books, both highlight the extensive information-processing requirements of social life. Human agents must represent reputations, alliances, obligations, trust, competence, social expectations, and relative status. They must also interpret the signals transmitted by others and regulate the signals they themselves produce.

A major resource of human motivation is commitments. Michel Aubé argued that commitments—to people, groups, organizations, institutions, and projects—constitute a distinctive class of motivational resource. Commitments become objects of perception, evaluation, planning, protection, and emotional regulation. His account is developed in “A Commitment Theory of Emotions” and “Beyond Needs: Emotions and the Commitments Requirement.”. See also my blog post on On The Relationship-building Proclivities of Human Nature.

This helps explain why threats to relationships, obligations, identities, and institutions can become so motivationally insistent. Human agents do not merely pursue individual rewards or satisfy biological needs. They organize much of their lives around commitments that persist across circumstances and coordinate their behaviour with that of others. Similarly, the prospect of securing and protecting commitments can generate strong motivation.

Human beings are also immersed in external symbolic systems: language, writing, mathematics, diagrams, books, laws, scientific theories, digital tools, and cultural practices. As Merlin Donald emphasized, mature human cognition develops within such symbolic cultures.

External resources do not merely store information that an otherwise complete mind occasionally consults. They become functionally integrated with memory, attention, planning, and management. A notebook can preserve an intention. A diagram can reorganize a problem. A calendar can implement prospective memory. A search system can extend retrieval. An AI system can support explanation, criticism, and the exploration of alternatives. Human minds are extended into their environments.

Human intelligence is therefore simultaneously biological, social, cultural, and technological.

7. Learning as architectural change

Intelligence is often defined partly as an ability to learn. But saying that a system learns does not tell us what changes, which mechanisms produce the change, or whether the system can influence its own development.

From an architectural perspective, learning is not one process that transfers information into long-term memory. It is a heterogeneous family of processes capable of modifying many parts of an architecture.

Learning may change factual knowledge, concepts, procedures, habits, attentional dispositions, retrieval pathways, motivational priorities, evaluative standards, or management strategies. It may establish new connections between mechanisms or alter when existing processes are recruited.

A central difference among systems concerns sample efficiency: how much experience they require to acquire a useful capability. Humans sometimes learn from remarkably few examples because new information can be interpreted through extensive prior knowledge, causal understanding, language, analogy, social instruction, and existing conceptual structures.

The enormous difference between the amount of data used to train contemporary AI and the amount from which individual humans learn has been described as “the sample-efficiency black hole”. Although the comparison must be made carefully, it points to a major architectural question: why can humans sometimes extract durable, transferable understanding from so little direct experience? Human learning is often rapid and transcends data. (E.g., Kindergarten children’s sensitivity to geometry in maps:

Children spontaneously extracted and used relationships of both distance and angle in the maps, without prior demonstration, instruction, or feedback, but they failed to use the sense information that distinguishes an array from its mirror image.

Contemporary AI systems often require enormous training datasets, although their trained representations may then support rapid adaptation and impressive generalization. “Learning” therefore names very different architectural processes in humans and machines.

The most consequential learning is sometimes recursive. An agent may improve not only what it knows, but how it learns, practises, retrieves information, manages attention, evaluates evidence, and corrects errors.

7.1 Productive practice

Productive practice requires activities that target appropriate components of a developing capability, generate informative feedback, and produce changes that transfer beyond the practised instance.

Practice does not improve performance merely through repetition. A person can repeat ineffective methods, automate errors, or become highly skilled at a narrowly practised task without acquiring a more general capability.

A person can also learn to design better practice. They may become more skilled at diagnosing weaknesses, choosing exercises, arranging feedback, spacing learning, retrieving knowledge, and modifying their environment.

The object of learning can therefore be another learning or management process. Through productive practice, architectural development can become partly self-directed.

7.2 Explanation, criticism, and error correction

One especially important form of human learning involves constructing and improving explanations.

An explanation does more than summarize observations or reproduce a pattern. It proposes an account of why something happens, how a mechanism operates, or what underlying structure produces an observed regularity.

Good explanations support transfer because they help distinguish relevant from irrelevant variation. They allow an agent to reason about new cases, anticipate consequences, and intervene more intelligently.

Explanation construction recruits many mechanisms: prior knowledge, analogy, causal representation, working memory, long-term memory, imagination, language, and narrative. Management processes may formulate questions, retrieve information, compare candidate accounts, and determine whether an explanation is adequate for the current purpose. Meta-management may detect that the problem has been framed poorly or that an attractive answer is being accepted too quickly.

As Karl Popper pointed out, explanations improve through criticism. Criticism involves searching for inconsistency, conflicting evidence, hidden assumptions, counterexamples, explanatory gaps, and more successful alternatives. It depends not only on logical competence but on motivation: an inconsistency produces no intellectual progress unless it is noticed and treated as important.

Error detection is not identical to error correction. A system may receive feedback that an answer is wrong without possessing the representations needed to diagnose the failure. Correction may require revising a belief, replacing an explanatory model, restructuring a representation, abandoning a goal, changing a practice method, or altering the standards by which future proposals are evaluated.

These activities form a recursive cycle:

Experience and problems → explanatory conjectures → criticism → detection and correction of error → architectural change → improved capacity for further explanation and criticism.

The cycle does not merely add information to memory. It can change what the architecture subsequently notices, retrieves, questions, explains, and corrects.

Predictions and observations are indispensable within this process, but they should not be treated as the ultimate products of intelligence. Their deepest epistemic value often lies in how they contribute to the creation, criticism, and improvement of explanations.

7.3 Meta-effectiveness

I use the term meta-effectiveness for the capacity and disposition to become increasingly effective. Meta-effectiveness includes learning how to learn, but it is broader. An agent may improve how it allocates attention, retrieves and applies knowledge, manages projects, regulates interruptions, evaluates priorities, constructs explanations, detects errors, uses external resources, or modifies its own habits.

Meta-effectiveness is not a separate faculty. It emerges from the interaction of management, meta-management, self-reminding, motivation, learning, productive practice, and long-term governance.

Two agents may possess similar current abilities but very different developmental potential. One may diagnose its limitations, seek criticism, organize productive practice, and improve the mechanisms underlying later performance. Another may perform equally well in familiar situations but lack the capacity to reorganize itself when conditions change.

An architectural theory must therefore ask not only what a system can currently do, but what it can learn to do, which parts of it are modifiable, and whether it can participate deliberately in its own development.

Intelligence is partly a property of current organization and partly a property of architectural plasticity—and of how that plasticity is governed.

8. Humans, other animals, and contemporary AI

The architectural conception provides a better way to compare humans, other animals, and artificial systems.

Instead of asking which is most intelligent, we can ask:

• Which mechanisms are present?

• How are they organized and connected?

• What representations do they employ?

• How are motivators generated and regulated?

• How does learning change the system?

• Can it participate in directing its own development?

• How is it integrated with social and external cognitive systems?

• And several other questions.

Contemporary AI exceeds humans in numerous tasks, including calculation, search, pattern recognition, linguistic production, and the manipulation of large quantities of information. Great apes possess perceptual, practical, social, and motivational abilities that differ substantially from both human cognition and present AI.

Humans are distinctive not because they perform best at every task, but because they integrate a particular collection of architectural properties: language, autobiographical and semantic memory, narrative, planning, management, meta-management, architecture-based motivation, social commitments, productive practice, external cognitive scaffolding, cumulative culture, and sample-efficient learning and lifelong development.

The most consequential differences concern combinations of properties rather than isolated abilities.

Language supports explanation. Explanation reorganizes memory. Memory supports planning. Planning organizes practice. Practice changes capabilities. Meta-management evaluates those changes. External tools preserve and extend the results. Social institutions distribute criticism and knowledge across individuals and generations.

Human intelligence cannot be understood by simply adding up separate abilities. Its distinctive character depends partly on the organization and recursive interaction of architectural properties.

Contemporary AI can generate sophisticated explanations, criticize proposals, identify inconsistencies, and revise its outputs. These capacities should not be dismissed merely because they often occur within prompted interactions.

The architectural differences concern persistence, integration, autonomy, motivation, and developmental continuity. Humans can use explanation within enduring projects of education, inquiry, identity formation, professional development, and self-regulation. Explanations may alter not only an immediate answer but a person’s lasting knowledge, habits, standards, commitments, and methods.

Similarly, AI can produce narratives, but its narratives are generally not grounded in an enduring autobiographical history, persistent identity, and long-term projects of its own. AI systems may retrieve information through context windows, model parameters, databases, and external tools, but they do not ordinarily develop personally organized long-term working memory through a continuous life of self-directed activity.

These are not permanent boundaries. Artificial systems may increasingly acquire persistent memory, autonomous projects, continual learning, richer motivational organization, self-monitoring, narrative continuity, and the capacity to modify their own methods.

The architectural conception does not depend on declaring that AI either is or is not intelligent. It gives us a vocabulary for describing which properties a system possesses, how they are organized, how strongly they are developed, and how they change.

Table 2 below provides a much more detailed way of comparing the intelligence of humans, great apes, and contemporary AI. Rather than scalar or multidimensional comparisons these comparisons in terms of mechanism of information processing.

Table 2. Comparing agent by processing capacity

9. Intelligence reconsidered

The central claim of this article is that intelligence is best understood as a property of information-processing architectures.

Binary classifications, scalar measures, and multidimensional profiles remain useful. They answer different questions and summarize important aspects of intelligent performance. But they do not by themselves explain how intelligent performance is produced.

An architectural conception directs attention toward mechanisms, representations, communication pathways, memory systems, motivational organization, learning processes, and developmental capacities.

It also helps us understand why two systems can perform similarly while relying on different architectures—or perform differently despite possessing many of the same mechanisms. Behaviour depends not only on architectural capacity but on knowledge, experience, current motivation, opportunity, and which mechanisms happen to be engaged.

Recasting intelligence in architectural terms does not diminish the importance of psychometrics, comparative cognition, neuroscience, education, or AI benchmarking. These disciplines provide evidence from which architectural properties can be inferred. Their findings become components of a deeper explanatory enterprise.

The architectural view also changes how we think about developing intelligence. Becoming more intelligent is not only a matter of accumulating information or improving performance on selected tasks. It can involve improving attention, memory retrieval, planning, motivation, criticism, error correction, management, meta-management, and the design of one’s cognitive environment.

Such development is individual, social, and technological. People improve through interaction with teachers, collaborators, critics, institutions, books, diagrams, software, and AI systems. These resources can become functionally integrated with internal processes of memory, motivation, and control.

Intelligence is therefore not simply something an isolated organism possesses. It is something that can be developed, scaffolded, socially distributed, and technologically extended.

The question with which we began—Is AI intelligent?—invites a binary answer to a question that requires a much richer analysis.

A better set of questions is:

What kind of information-processing architecture does this system possess? What forms of intelligent activity does that architecture support? How are its mechanisms organized and regulated? What can it learn, and can it participate in directing its own development? How does this information processing architecture compare with specific others?

Performance gives us evidence with which to investigate these questions. But performance is not the final object of explanation. Behaviour provides evidence. Architecture provides the explanation.

If intelligence is not binary, neither is it merely scalar or multidimensional. More fundamentally, intelligence is a property of information-processing architectures.

10. From intelligence to consciousness

The same principles apply when we ask whether AI, other animals, or other forms of autonomous agency are conscious. The question “Is this system conscious?” treats consciousness as a binary property and risks suppressing important differences among forms of awareness and control. Asking only how conscious a system is replaces the binary judgment with a scale but may still conceal qualitative differences. A multidimensional account can distinguish sensory awareness, bodily awareness, self-awareness, social awareness, reflective awareness, and intermediate- or long-term consciousness, but it remains primarily descriptive.

An architectural conception asks how these forms of consciousness are produced: which perceptual, attentional, mnemonic, motivational, managerial, meta-managerial, alarm, narrative, and self-representational mechanisms are present; how they interact; and over what timescales they can govern activity. Humans, other animals, and artificial systems may exhibit both continuities and discontinuities in these respects. We should therefore investigate not only whether or to what extent a system is conscious, but what kinds of consciousness its architecture makes possible and how those forms of consciousness contribute to autonomous agency.

If you want to find out more and have my definition of intelligence see a long exerp If Intelligence Isn't Binary, What Is It?

Most people think of links as things that open files, web pages, emails, or notes. But what if you could create a link that finds information instead?

That’s exactly what Hookmark’s Search Links do.

Rather than pointing to a single item, a Search Link points to a search. Click it, and Hookmark instantly retrieves whatever currently matches your search—whether that’s one file or twenty. As your information grows and changes, the search results evolve with it.

This opens up some surprisingly powerful workflows. You can create reusable searches for projects, research topics, product names, customer IDs, or any other text that matters to you. Instead of remembering where something lives—or even whether it still lives there—you simply activate the Search Link.

It’s one of Hookmark’s powerful yet least-known capabilities.

I’ve written a short illustrated article that explains how Search Links work and shows several practical examples. If you’re interested in spending less time searching and more time thinking, I think you’ll enjoy it.

👉 Read the article: https://hookproductivity.com/blog/2026/07/find-anything-later-with-hookmark-search-links/

Find out more

Read more about Hookmark here.

Read my Cognitive Productivity with macOS: 7 Principles for Getting Smarter with Knowledge for more information. It’s a very practical book with over 60 short videos. The seven principles are

I. Principles of self-governance

• Principle 1: Lead yourself with knowledge

• Principle 2: Manage your cognitive life mindfully

II. Principles of productive information-processing

• Principle 3: Assess analytically

• Principle 4. Surf strategically

• Principle 5. Delve deeply

III. Principles of mastery

• Principle 6. Practice productively

• Principle 7. Apply knowledge

Rejecting the binary question, however, leaves a more interesting one: If intelligence isn’t binary, then what is it? My answer is that intelligence is best understood not as a label, not as a score, and not even primarily as a profile of abilities. More fundamentally, intelligence is a property of information-processing architectures.

Figure generated by AI based on my specification

Psychology has traditionally described intelligence in two familiar ways. The first is scalar. Intelligence tests ask how intelligent someone is, producing measures such as IQ. The second is multidimensional. Rather than reducing intelligence to one number, many researchers distinguish language, memory, reasoning, creativity, social understanding, executive function, and many other abilities. These approaches have taught us a great deal and remain extremely valuable. But they remain largely descriptive.

Suppose two systems perform equally well on a planning task. One constructs explicit plans, monitors progress, interrupts itself when circumstances change, and deliberately revises its strategy. The other arrives at similar answers using entirely different internal mechanisms. From the standpoint of behaviour, they may look alike. From the standpoint of explanation, they may be profoundly different. Performance provides evidence about intelligence. It is not intelligence itself.

To explain intelligence we need to ask what kind of information-processing organization gives rise to intelligent behaviour. The architectural perspective has a long history in artificial intelligence and cognitive science. I think Aaron Sloman was the first to emphasize the importance of information-processing architectures in AI. Search for “architecture” in his still relevant 1978 book, The Computer Revolution in Philosophy: Philosophy, Science and Models of Mind, and you will find many hits. Allen Newell later argued that cognitive science should seek unified theories capable of explaining memory, reasoning, learning, problem solving, and language within one integrated system rather than as isolated modules.

The present essay belongs to that tradition.

An information-processing architecture is much more than a list of abilities. It includes the mechanisms through which an agent generates goals, allocates attention, stores and retrieves information, learns, plans, evaluates alternatives, interrupts itself, governs its own thinking, and adapts over time. It also includes the communication pathways linking these mechanisms together. Two systems can exhibit similar behaviour while possessing very different architectures, just as birds and airplanes both fly despite relying on radically different mechanisms.

Human intelligence, from this perspective, is not explained by any single faculty. It emerges from the interaction of many architectural mechanisms.

Some of these mechanisms operate reactively, responding quickly to opportunities and dangers. Others support deliberation, planning, scheduling, and decision making. Higher-level management processes coordinate ongoing activities, while meta-management processes monitor and regulate the management processes themselves. Human intelligence also depends on systems for generating and evaluating motivations, detecting important changes through alarm mechanisms, governing behaviour across long periods of time, constructing explanations, learning from criticism, and continually reorganizing itself through experience.

These mechanisms are not independent. They continually interact. Memory supports planning. Planning guides learning. Learning changes motivation. Motivation directs attention. Meta-management evaluates whether current strategies are working. The intelligence of the whole arises from the organization of the parts rather than from any one component.

Evolution provides another perspective on this organization. Building on Merlin Donald’s work, I suggest that uniquely human intelligence emerged through successive architectural innovations rather than through one major breakthrough. Memesis enabled the following development of mythic consciousness. Language enabled narrative thought and shared culture. Writing and other external symbol systems transformed memory and reasoning by allowing knowledge to accumulate outside the brain. Today, books, diagrams, search systems, note-taking tools, contextual information retrieval, and AI continue extending human cognitive architectures. Mature human intelligence is therefore partly biological, partly social, and partly technological.

Learning itself also looks different from an architectural perspective. Rather than viewing learning simply as storing information or developing procedural knowledge, we can think of it as changing the architecture. Learning may alter knowledge, habits, motivational systems, evaluative processes, strategies, executive control, or even the mechanisms through which future learning occurs. One can develop new motive generators and mechanisms for generating alarms. Humans are especially remarkable because they can deliberately participate in this process. Through explanation, criticism, productive practice, and reflection, they can redesign aspects of their own information-processing architecture.

This leads naturally to what I have elsewhere called meta-effectiveness: the capacity and disposition to become increasingly effective. Perhaps the highest expression of intelligence is not simply solving today’s problems but improving the architecture that will solve tomorrow’s.

The architectural perspective also provides a richer framework for comparing humans, other animals, and artificial intelligence. Instead of asking whether one system is more intelligent than another, we can ask which architectural mechanisms they possess, how those mechanisms are organized, how they interact, and how they develop over time. Humans, great apes, and contemporary AI share important continuities while also exhibiting significant architectural differences. These comparisons become far more informative than attempting to place every mind on a single intelligence scale.

None of this diminishes the value of IQ tests, psychometric research, or AI benchmarks. On the contrary, they become even more useful because they provide evidence from which we can infer underlying architectural properties. The mistake is treating those measurements as intelligence itself rather than as clues about the organization that produces intelligent behaviour.

I therefore propose adding a fourth conception of intelligence to the familiar binary, scalar, and multidimensional conceptions. Intelligence should also be understood architecturally.

Viewed this way, the central scientific challenge changes. Instead of asking whether a system is intelligent, we ask what kind of information-processing architecture it possesses, how that architecture developed, how it learns, how it governs itself, and how it supports the remarkable variety of behaviours we associate with intelligence.

Behaviour provides the evidence for intelligence. Architecture provides the explanation.

I have a new project to which my expertise and interests in integrative design-oriented R&D I am applying. With Eva Hudlicka as first author, I’m co-authoring a journal article on using AI for psychotherapy. Many people are quite negative about this use of AI. We are more balanced. One of the advantages of ChatGPT is its expertise in all major models of psychotherapy: ACT, CBT, DBT, EFT, and so forth. One of its weaknesses is its verbosity. But one can tell AI to keep it short. The journal article will include a table, generated with the help of AI, comparing the effectiveness of AI and human therapists. On a number of important dimensions, AI already compares surprisingly well with human therapists—though, of course, the various dimensions need to be weighted. Some dimensions are much more important than others.

In a future post, as a precusor to the journal article, I’ll explain how I use AI for psychoeducation and psychotherapy. These days I’m learning about emotion-focused therapy and finding it useful. I’m reading several books and articles about it and interacting with ChatGPT. I will write about this too.

The AI and psychotherapy project mentioned above reflects my broader interest in integrative models — i.e., that include cognition, alarm systems, motive generators, motive filters, executive functions (management and meta-management) and other mechanisms. You can’t do psychotherapy without considering the entire person’s information processing architecture. Rather than asking whether humans or AI are better, I’m interested in how thoughtfully designed combinations of both can produce better outcomes than either alone. That is, I don’t see AI as a replacement for the therapist but an augmentation. I also believe psychotherapists need to use AI for their professional development and to help deal with their patients. For privacy reasons that requires using local, not cloud-based, AI, of course.

Eva Hudlicka is a unique person. She is one of the most well published researchers on AI and emotions, meaning she is both an emotion researcher and an AI expert. She’s also a therapist. I feel honored that she, a close friend of mine, invited me onto her paper.

Also, as a conflict of interest disclosure: I often refer to my products in my writing — they are after all the biggest part of my professional life. I have three books: Cognitive Productivity: Using Knowledge to Become Profoundly Effective, Cognitive Productivity with macOS: 7 Principles for Getting Smarter with Knowledge and Discontinuities: Love, Art, Mind. And I am CEO of CogSci Apps Corp. which makes mySleepButton and Hookmark. That last post mentioned all of those products except mySleepButton.

Discontinuities

My latest book, Discontinuities: Love, Art, Mind, is 90% complete but already for sale as an ebook on Leanpub. It is currently on sale, 50% off, for people who use this coupon url https://leanpub.com/discontinuities/c/Substack. Discontinuities collects essays that I have written from 2012 onwards and also contains much new material. (E.g., this 2012 article: A Zestful Response to Philippe Falardeau’s Monsieur Lazhar and this 2014 article: Meta-painting & Science of the Human Mind: An Epistolary Response to Lam Wong’s 21 Elements – CogZest, both long form responses to art.) This book, which is very dear to my heart, applies Cognitive Productivity ideas to learning from stories and other forms of art. It also deals with other themes that are central to my work. Here is the current publisher’s blurb —the publisher being my sole proprietorship, CogZest:

In this the Knowledge Age, we gorge on fragmented factual and artistic information. But are we experiencing commensurate personal growth? In Cognitive Productivity Beaudoin discussed how to develop ourselves and solve problems with the knowledge we “consume”. But what about self-development from art created by others?

Discontinuities illustrates the challenge, possibility and necessity of striving for integrated understanding of the world and of ourselves through the art of others. This book presents variations on the theme of learning about ourselves and each other by responding deeply to art in all its forms—story, visual art, music, etc. Through essays, epistles, stories, visual art, puzzles, and more, this book begins to unfurl a new framework for more deeply being transformed by the art we experience. This includes bibliotherapy and related concepts.

With contributions from National Teaching Award recipient Prof. Claude Lamontagne, visual artist and Buddhist scholar Lam Wong, sculptor and geneticist, Dr. Al Sather, historian, Dr. Laura Nys, photographer, Andrea Paterson, and painter and psychotherapist Arash Kameli.

While unique in form and content, this book is inspired by Geoffrey Miller’s The Mating Mind, Dorothy Tennov’s Love and Limerence, Warren McCulloch’s Embodiments of Mind, Marvin Minsky’s The Society of Mind, Aaron Sloman’s notion of discontinuities in the space of possible minds, other works in affective/cognitive science, novels by Paul Auster, Milan Kundera and others, Edward Elgar’s Enigma Variations, Lizst’s Piano Concerto No. 2, Willa Cather’s The Troll Garden, E.C. Escher, songs of Jacques Brel, and who knows what else?

One thing that strikes me as odd is that the concept of alarms is rarely mooted in empirical psychology. Yet the first stage of Hans Selye’s stress response, in his seminal 1936 paper which introduced the concept of stress, is an alarm reaction. You’d think modern theories of stress and anxiety would be full of alarms. But they’re not.

Aaron Sloman proposed human-like autonomous agents have computational alarm systems. I think he was spot on. But he never related his ideas to empirical psychology. I’m trying to fix that.

Here’s an excerpt from : Beaudoin, L.P. & Guloy, S. (in press). Towards a somnolent information-processing theory: Understanding the human sleep-onset control system from an integrative design-oriented perspective. In J. Dzierzewski, D. Kay & S. J. Aton (Eds), The Cambridge Handbook of Sleep Theories and Models. Cambridge University Press. I wrote the chapter in 2023 but it’s only going to be published in 2026.

Hans Selye described the first stage of the stress response as involving an “alarm reaction” (Selye, 1973; Selye, 1936). Sloman & Croucher (1981); Sloman (2003); Sloman et al. (2005) introduced the concept of alarms in AI research on autonomous agents and emotion, an idea that was incorporated into the Global Workspace theory of consciousness (Baars & Franklin, 2009). In the H-CogAff computational architecture (Sloman, 2003; Sloman et al., 2005), alarms can be trigged by perception and by deliberative processes; and they can directly and globally affect two H-CogAff levels: the reactive and deliberative levels. The reactive effects of alarms in H-CogAff can include arousal and behaviors or readiness for behavior (freezing, fleeing, etc.). Like motives, the effects of alarms on executive processes are assumed to be subject to attention filtering. This leads to the possibility of alarm-based perturbance, where reflective processes would have difficulty controlling deliberative processes due to the alarm.

The term “alarm” does not figure prominently in psychology, even in stress literature. In addition to in H-CogAff papers, it appears informally in some research on pain (Moreno et al., 2015; Plaghki et al., 2010; Cervero, 2012; Eisenberger & Lieberman, 2004) and on emotion (Oatley, 1992). The concept of alarm is admittedly highly speculative and vague. Still, attempting to understand emotion, stress and ad hoc arousal in terms of computational alarms seems a promising avenue for both empirical and integrative design-oriented research in psychology.

I wish I had been able to go into more detail about alarms, but I was at the maximum word count limit for the chapter. (I hate word count limits!)

Here’s what I had to say about alarms in a 2020 paper on mental perturbance:

Computational Alarm Systems Further addressing requirements of autonomous agency, and further accounting for psychological phenomena (such as aspects of ‘emotional’ and stress reactions), H-CogAff assumes mechanisms for generating and processing alarms. Alarms are control signals that have global effects in the architecture. At a physiological level, alarms can activate the sympathetic nervous system (Buck, 2014). We assume they can parameterize executive processes, such as leading to more vigilance, changing the level of abstraction of thinking, or make deliberation more or less careful. They may have other effects on management processes and “action readiness” that more precise formulations of the theory may specify.

The H-CogAff architecture distinguishes between 1) alarms triggered by perceptual information, such as an angry glare or the unexpected appearance of the object of one’s infatuation; and 2) alarms triggered by noticing significant issues in executive layer content e.g., suddenly realizing a plan of action may have a disastrous side-effect (Sloman, 2003; Sloman, Chrisley & Scheutz, 2005).

Selye originally described stress as an alarm reaction (1936)—an idea that before the current paper had not been linked to computational alarms. Alarms have also (briefly) been posited in theories of consciousness (Baars & Franklin, 2009), emotion (Oatley, 1992) and pain (Eisenberger & Lieberman, 2004). We believe the IDO conception of alarms, modernizing Selye’s concept (1936), is worthy of future IDO and empirical research.

I was pleased to find that two recent CBT books, meant for lay readers but written by clinical psychologists, discussed alarms in some detail. They are:

• The intrusive thoughts toolkit: quick relief for obsessive, unwanted, or disturbing thoughts and

• Rewire Your Anxious Brain: How to Use the Neuroscience of Fear to End Anxiety, Panic, and Worry

They attribute alarms to the amygdala, which makes sense.

I will write more about alarms in a future post on “emotion”. I write “emotion” in scare quotes because Agnes Moors and I take an eliminativist view of emotion, suggesting they are best viewed as descriptive but not explanatory constructs. Agnes published two papers in this month’s issue of Emotion Review: a target article, to which several authors replied, and her response. I have a lot of time for Agnes Moors and her Goal-Directed Theory — that should be no surprise because, as I shall argue, it is very compatible with my own theory. My thesis after all was on Goal Processing in Autonomous Agents. I will argue that there may be no emotion mechanisms, but there are alarm mechanisms which resemble emotions. In fact Aaron Sloman referred to alarms as primary emotions, e.g., in Sloman (2003) How many separately evolved emotional beasties live within us

Abstract

Abstract

The concept of epistemic agency has become increasingly influential in educational psychology, the learning sciences, and AI-in-education research. Existing conceptions emphasize important epistemic practices, including knowledge building, collaborative inquiry, evidence evaluation, self-regulated learning, and responsible engagement with generative artificial intelligence. This article argues that these perspectives, while making substantial contributions, generally describe epistemic practices more fully than they explain the nature of the autonomous agent who performs them. I propose that epistemic agency should be understood as one manifestation of a broader phenomenon: Cognitive Productivity, defined as the effective use of knowledge resources to solve problems, develop products, and improve ourselves. Drawing on an integrative design-oriented perspective and the H-CogAff model of autonomous agency, I argue that epistemic agency presupposes a theory of autonomous agents organized around motivators (projects and goals, norms, preferences, and needs), management and meta-management processes, and architecture-based effectance—the motivation to improve oneself that emerges from an autonomous cognitive architecture. Building on this foundation, I introduce meta-effectiveness, the capacity to improve one’s own methods of becoming effective, and argue that education should aim not merely to build knowledge but to transform learners by developing their knowledge, skills, habits, motivators, motive generators, and other forms of mindware. The article reviews major conceptions of epistemic agency, relates them to research on self-regulated learning, and presents the seven principles of Cognitive Productivity as a theoretical and practical framework for developing epistemic agency across formal education, professional knowledge work, and lifelong learning. The result is a broader conception of epistemic agency that integrates educational psychology, self-regulated learning, autonomous-agent research, and theories of knowledge work within a common theory of Cognitive Productivity, providing a foundation for future research on learning, AI, professional knowledge work, and lifelong human development.

Preamble

This article is an initial attempt to improve and generalize the concept of epistemic agency by bringing it into contact with integrative design-oriented approach to autonomous agency and my work on Cognitive Productivity. It is intended as a precursor to a peer-reviewed conceptual article. I hope readers will provide feedback in the comments (or email me), which may help shape the forthcoming paper.

This article goes into too much detail for many readers, including in particular the Surf Strategically principle. It’s partly that it consolidates over 30 years of my work You can skip parts of the article if you prefer. I hope you’ll at least tune into the intro and conclusion and read here and there. You can also ask AI to summarize it for you. I trust you will apply your epistemic agency 😊.

Contents

• Introduction

• Conceptions of epistemic agency in the educational literature

• Epistemic agency presupposes autonomous agency

• Learning should transform the agent

• Effectance as architecture-based motivation

• New concepts for learning and epistemic agency

• Meta-effectiveness: improving the ability to improve onself

• Cognitive Productivity

• The seven principles as a framework for epistemic agency

  • Principle 1. Lead Yourself with Knowledge

  • Principle 2. Manage Your Cognitive Life

  • Principle 3. Assess Analytically

  • Principle 4. Surf Strategically

    • The meta-access problem

    • Contextual retrieval and consciousness

    • Surfing as epistemic control

  • Principle 5. Delve Deeply

  • Principle 6. Practice Productively

  • Principle 7. Apply Knowledge

• Individual and shared epistemic agency

• Relation to research on self-regulated

• Human-Al relations reconsidered

• A proposed general conception of

• Conclusion

• Colophon

Introduction

The term epistemic agency is gaining traction in educational psychology, the learning sciences, science education, educational technology, and research on generative AI. The phrase is compelling. It suggests that learners should not merely receive information or comply with instructions; they should take responsibility for knowing. They should formulate questions, assess claims, seek and interpret evidence, contribute ideas, regulate inquiry, and participate in the advancement of knowledge.

These are important aims. Yet the concept remains theoretically unsettled, as argued in a recent Ph.D. thesis by Diana Meek. Different researchers use epistemic agency to refer to overlapping but significantly different phenomena. It may mean taking responsibility for one’s own learning, contributing to a community’s knowledge, directing collaborative inquiry, participating in disciplinary practices, evaluating evidence, or maintaining human judgment when interacting with artificial intelligence. The literature has produced many valuable descriptions of epistemic practices but has not yet supplied a sufficiently general model of the agent who performs them.

I will argue that epistemic agency presupposes autonomous agency. More specifically, it needs an integrative design-orientedmodel of the learner and knowledge worker as autonomous agents. Such an account should explain how agents agents generate and manage motivators, how they select and process knowledge resources, how learning transforms them, and how they improve the very mechanisms through which they learn and act.

My interest includes not only students. It extends to professional knowledge workers and lifelong learners — people who continually use knowledge resources to understand situations, make decisions, solve problems, develop products, improve themselves, and help others.

Nor is AI the defining context of epistemic agency. Generative AI has made the issue urgent because it can produce plausible answers, explanations, arguments, and designs without ensuring that its users understand or can justify them. But AI is one class of knowledge resource among many. Books, articles, conversations, teachers, colleagues, notes, diagrams, databases, search engines, reference managers, software tools, and one’s own prior writings also mediate cognition. A theory that explains agency only in relation to AI is therefore not a general theory of epistemic agency.

The broader framework I propose is developed in my books Cognitive Productivity: Using Knowledge to Become Profoundly Effective and Cognitive Productivity with macOS: 7 Principles for Getting Smarter with Knowledge. The first develops a theory of meta-effectiveness, meaning the knowledge-based development of competence as further defined below. The second expresses that theory through seven practical principles which it illustrates extensively (including over 60 videos of how to use software for them). The principles, elaborated below are:

1. Lead yourself with knowledge

2. Manage your cognitive life mindfully

3. Assess analytically

4. Surf strategically

5. Delve deeply

6. Practice productively and

7. Apply knowledge.

These are not merely tips for managing information on a Mac. Taken together, they constitute a theoretical and practical framework for individual epistemic agency. The books’ structures themselves reflect this progression: self-governance, productive information processing, and mastery culminating in application.

Conceptions of epistemic agency in the educational literature

The educational idea of epistemic agency is rooted in the knowledge building tradition initiated by Carl Bereiter and Marlene Scardamalia – most elaborately in Bereiter’s 2002 masterpiece, Education and Mind in the Knowledge Age. Their work shifted attention from students merely learning to students assuming collective cognitive responsibility for advancing ideas, and generalized it to professional knowledge workers. Students should not simply complete activities selected by teachers. They should help identify problems, propose explanations, improve theories, make use of authoritative sources constructively, and contribute to knowledge that matters to their community.

This was a major conceptual advance. It located agency not only in choosing how to study but in participating in knowledge creation in communities of knowledge builders. It also challenged the assumption that the teacher must retain nearly all cognitive authority. Students could become legitimate contributors to a community’s knowledge rather than remaining consumers of curricular content.

• Crina Damşa and colleagues developed this tradition in their seminal 2010 article, “Shared epistemic agency: an empirical study of an emergent construct.” They define shared epistemic agency as a group’s capacity to undertake deliberate, sustained, knowledge-driven collaboration aimed at producing shared knowledge objects. The knowledge object may be an instructional design, report, model, or other tangible product in which the group’s developing ideas become materialized. Their analysis distinguishes an epistemic dimension—searching for information, sharing and structuring ideas, generating new ideas, and developing the knowledge object—from a regulative dimension concerned with goals, planning, coordination, monitoring, negotiation, and management of the collaborative process.

  • The Damşa et al. paper is especially valuable because it does not treat agency as an invisible trait inferred from successful performance. They seek observable indicators in the group’s activity. Shared epistemic agency emerges when members establish a common understanding, propose and compare alternatives, integrate contributions, regulate their work, and translate emerging knowledge into their common object. It is not automatically present whenever people are assigned to a group. It is recursive, gradual, and dependent on patterns of interaction.

• Lai and Campbell’s 2018 study of a secondary-school art-history class places epistemic agency within a knowledge building community supported by Bereiter’s & Scardamalia’s Knowledge Forum software. They define it as the active process of choosing when, what, and where one learns, how one knows, and how one participates in creating knowledge within a community. Students exercised agency by identifying gaps in their knowledge, sharing information, developing ideas, building on the contributions of others, and creating communal knowledge. Their account also highlights familiar knowledge building principles: collective responsibility, improvable ideas, idea diversity, constructive use of authoritative sources, and assessment embedded within the knowledge-building process.

• The concept of epistemic agency has also become prominent in science education, where it is often framed as participation in the practices through which scientific knowledge is proposed, evaluated, communicated, and legitimized. Vasconcelos and colleagues describe students as epistemic agents when they conduct self-directed inquiry, generate propositions, formulate and test hypotheses, evaluate evidence against theories, produce explanations, refine inquiry practices, and connect their knowledge to the standards and activities of a scientific community. Their study of preservice teachers’ robotics-enhanced lessons is especially instructive because the technology itself did not guarantee agency. The robots were often used to transmit content or stage activities rather than to support student-driven inquiry. A technologically rich lesson can remain epistemically poor.

• Nam-Hwa Kang’s perspective on physics education similarly characterizes epistemic agents as students who assume accountability for their learning, set goals, regulate their processes, and share cognitive authority with instructors. This resembles the concept of self-regulated learning, discussed below. Students must learn not only the content of physics but its standards, methods, and forms of reasoning. They should regard themselves as members of a community who bear some responsibility for what they and their peers believe. Digital technologies can support this transition, but only when they are used to enable authentic inquiry rather than simply to modernize information delivery.

• Recent work on generative AI extends these concerns. Wu and colleagues (2025) define epistemic agency as the capacity of individuals or groups to shape the knowledge and practices of a community. They propose a human–AI form of shared epistemic agency while insisting that the human must remain its fundamental initiator and driver. Users should actively acquire, question, verify, interpret, and modify AI-generated content rather than receive it passively. They also connect epistemic agency to epistemic stances: absolutist, multiplist, and evaluativist orientations toward knowledge and justification. Evaluativist learners, who recognize that claims can be compared using evidence and domain-relevant standards, are better positioned to benefit from AI without surrendering judgment to it.

• The 2025 editorial by Yan and colleagues situates epistemic agency alongside cognition, metacognition, self-regulation, and the socio-emotional consequences of sustained AI use. It warns that educational attention to efficiency and performance can obscure deeper questions about superficial processing, overreliance, epistemic vigilance, and the changing relation between learners and knowledge. Their synthesis points toward an education in which learners do not merely use AI effectively but critically and ethically co-author an AI-mediated future.

• Rachel Horst’s funded Insight Development Grant research proposal, developed with colleagues at UBC, advances a particularly explicit sociotechnical conception. In that proposal, epistemic agency is the situated exercise of judgment, ethical deliberation, verification, and responsible refusal in AI-mediated knowledge work. It concerns deciding when AI outputs require checking, how deeply they should be interrogated, and when delegating a cognitive task would itself be epistemically inappropriate. Crucially, this is not treated as a fixed disposition residing wholly inside an individual. It is enacted through interactions among learners, instructors, assignments, institutional policies, pedagogical protocols, and course-embedded AI systems.

Across these accounts, epistemic agency is associated with ownership, responsibility, inquiry, evaluation, justification, knowledge creation, regulation, participation in disciplinary practices, and increasingly the maintenance of human judgment in interactions with AI. These are substantial contributions. Collectively, however, the literature tends to define epistemic agency in terms of participation in epistemic practices with a particular emphasis on evaluative processes (a more general conception of which is conveyed by my “Assess Analytically” principle and its CUPA evaluative schema, summarized below.) It tells us something about what epistemically agentic learners and groups do. It says less about the general architecture of an autonomous agent capable of doing these things. My concern is partly with the information-processing architecture that makes those practices possible.

That is the opening for an integrative design-oriented account of human or human-like autonomous agents.

Epistemic agency presupposes autonomous agency

Agency is not merely activity. A thermostat acts causally upon its environment, but we do not normally treat it as an autonomous agent. Autonomous agency involves internally generated and managed motivators. An autonomous agent must be able to generate or acquire motivators (bottom-up and top-down), detect relevant opportunities and problems, deliberate among alternatives, regulate competing demands, learn from consequences, and alter its future behaviour.

Autonomous agency was the subject of my 1994 Ph.D. thesis, Goal Processing in Autonomous Agents, which I believe remains relevant today, and can help ground the concept of epistemic agency.

Epistemic agency is a specialization of this more general capacity. It concerns the agent’s relations to knowledge: how knowledge resources are selected, interpreted, assessed, mastered, transformed, and applied in the service of what the agent values.

Existing educational accounts often presuppose these capacities. A student is expected to “take responsibility,” “set inquiry goals,” “evaluate evidence,” “decide when verification is warranted,” or “refuse delegation.” But each of those descriptions raises important largely unanswered questions about the agent’s information-processing architecture. How are candidate goals or concerns generated? What makes one of them salient? How does it gain priority over competitors? What enables an agent to interrupt an ongoing activity, reflect upon it, and change course? How does a concern with evidence become sufficiently entrenched that it influences future thinking without requiring a teacher’s immediate prompt? How do principles learned declaratively become operative dispositions and habits? What kinds of processes manage (deliberate over) the agent’s motivators? How are management or deliberative processes themselves regulated through what I call meta-management and others call reflective processes? (Management and meta-management processes are executive processes.)

An integrative design-oriented approach asks such questions directly. Instead of treating cognition, motivation, emotion, attention, learning, and executive control as separate topics, it asks what interacting mechanisms an autonomous information processing architecture requires. Alan Newell’s seminal 1990 book, Unified Theories of Cognition popularized information processing architectures, however unlike my work with Aaron Sloman, who was my Ph.D. thesis advisor, it only dealt with “dry” cognition, not motivation and affect.

• The approach is design-oriented because it takes the design stance elaborated by David Marr, Daniel Dennett, Aaron Sloman of the university of Birmingham and others. (On this subject, I recommend Aaron Sloman’s Prospects for ai as the general science of intelligence). That is to say that one way to expose the incompleteness of a verbal theory is to ask what would be required to design an agent that exhibited the competence in question.

• The approach is integrative because a plausible autonomous agent cannot be built from an isolated theory of memory, motivation, reasoning, emotion, or learning. These functions and others constrain and partly constitute one another.

The H-CogAff account belongs to this tradition. It distinguishes reactive processes, deliberative or management processes, and reflective or meta-management processes, while also incorporating motivator generators, attention and interrupt mechanisms, working memory, long-term memory, and forms of affective control. This is a general architecture-schema rather than a model only of students—or even only of humans. It can guide theories of natural and artificial autonomous agents.

H-CogAff is also fundamentally goal-directed. In that respect it is compatible with Agnes Moors’s Goal-Directed Theory, in which much emotional and non-emotional behaviour is explained through interacting goal-directed cycles rather than special-purpose emotion mechanisms. Moors’s theory is focused on humans and the phenomena commonly called emotions; ours ( H-CogAff, etc.) is more general in scope and explicitly integrative design-oriented, encompassing human and artificial agents. A fuller comparison belongs in a separate article. For present purposes, the relevant point is that epistemic agency cannot be detached from the goal-directed organization of the agent as a whole. Moors herself notes that epistemic goals, including the goal to obtain information, can support many other goals and can acquire intrinsic value.

The H-CogAff architecture is updated, in integrative design-oriented terms, in my 2020 paper with Sylwia Hyniewska and Monica Pudlo: Mental perturbance: An integrative design-oriented concept for understanding repetitive thought emotions and related phenomena involving a loss of control of executive. (See also part 2 of my first Cognitive Productivity book.)

Learning should transform the agent

The literature on epistemic agency rightly emphasizes knowledge building. Bereiter is explicit that knowledge building is not learning. It deals with Popper’s third world of objective artifacts. But knowledge building is not the only aim of education. Learning itself is also an aim, and true learning should transform the agent.

An integrative design-oriented account does not identify learning solely with acquiring declarative propositions or procedural knowledge. Learning may produce objective knowledge in a community, but within the learner it changes what David Perkins (in Outsmarting IQ: The Emerging Science of Learnable Intelligence and Keith Stanovich (in What Intelligence Tests Miss: The Psychology of Rational Thought) called mindware. It is the “mental stuff” of knowledge, now in the sense of Popper’s World 2 (as opposed to World 3): _ skills, procedures, monitors, habits, motivators, motive generators, and other dispositions that affect future cognition and action. In other words, it transforms the agent’s information processing architecture by adding new structures and mechanisms, and modifying them.

This aligns with Marvin Minsky’s seminal 1985 book, The Society of Mind. In this integrative design-oriented account, Minsky emphasized that learning could involve adding new If-Do-Then rules, change low-level connections, make new subgoals for goals, choosing better search techniques, changing high-level descriptions and narratives, making suppressors and censors (to prevent errors), linking older fragments of knowledge, making new kinds of analogies, making new models, virtual worlds, and other types of mental representations. Making sense of this requires an information processing architecture well beyond the simple distinctions between sensory memory, short-term and long-term memory.

Chapter 2 of Cognitive Productivity distinguishes several contributors to effectiveness, including mastery of objective knowledge, development of implicit understanding, skills, norms, attitudes, propensities, habits, and other dispositions. The book then situates these changes within adult mental development and the pursuit of greater meta-effectiveness.

Suppose a student learns a norm against manipulating evidence. The outcome is not adequate if the student can merely recite the norm on an exam. The learning becomes consequential when the norm acquires motivational force (which Sloman and I refer to as intensity and insistence) —when it helps detect questionable conduct, generates concern, interrupts expedient behaviour, and influences decisions even in the absence of external surveillance.

Similarly, learning about the importance of considering alternative explanations should not terminate in a declarative belief that alternatives matter. It should help develop a monitor that notices when an attractive explanation is being accepted too quickly and generates a motive to search for competitors. In the terminology of my autonomous-agent work, learning may create or refine motive generators: mechanisms that detect relevant conditions and generate motivators in response.

The same is true of preferences. Education can cultivate a preference for depth over superficial fluency, for well-structured explanations over jargon, for evidence over tribal reassurance, and for cognitively potent resources over seductive but unhelpful content. It can also change habits: whether a person checks a source, records an insight, revisits an important concept, tests their recall, compares theories, or applies a principle outside the context in which it was taught.

Learning should therefore generate not only knowledge but new motivators, motive generators, and habits. A learner who has acquired a durable concern for explanatory adequacy, a habit of checking assumptions, and a preference for intellectually honest disagreement has not merely added content to memory. The learner has become a different kind of agent.

The meta-effectiveness component of epistemic agency involves become better at using knowledge resources (AI or static) for such self-transformation. One can become more effective at becoming more effective. Effectiveness is the core measure of agency.

Effectance as architecture-based motivation

This account raises a question that epistemic agency research rarely addresses directly: why should an agent be motivated to improve itself?

Robert White’s influential 1959 paper introduced effectance as motivation for competence. The proposal helped challenge theories that reduced motivation to the satisfaction of primary biological drives.

My use of the concept is both indebted to White and theoretically revisionary. I recast effectance using Aaron Sloman’s revolutionary concept of architecture-based motivation, in this case architecture-based motivation to improve oneself. This is not merely a terminological update. It is an architectural extension, meaning it interprets the motivation of effectance in terms of information processing architecture.

An autonomous architecture must continue functioning in environments that are variable, partly unknown, and populated by opportunities and threats the agent has not encountered before. Fixed competences cannot be sufficient. An architecture capable of detecting limitations, monitoring outcomes, representing possible improvements, and generating motivators to learn will be more adaptive than one restricted to a static repertoire.

Effectance, on this account, must not be represented as a single standing goal such as “become more competent” from which more specific self-improvement goals are derived (through means-end reasoning). A child playing with lego is simply motivated to play, he is not deliberately then motivated to become more effective. However, this architecture-based, intrinsic, motivation to play as a consequence helps the child become more effective. Effectance can thus emerge from the architecture’s organization. Deficiencies, prediction failures, blocked projects, admired exemplars, unexplained phenomena, and encounters with potent knowledge resources can generate motives the pursuit of which incidentally, but importantly, help the agent to develop. Architecture-based motivation was instilled by evolution, not by means-ends analysis. The architecture supplies a basis for self-improvement across many domains. Curiosity, an aspect of effectance, can similarly be interpreted in terms of architecture-based motivation.

This provides a motivational foundation for epistemic agency. An agent does not only seek answers to current questions. It can implicitly pursue motives that as a side-effect make them become a better questioner, reasoner, learner, designer, collaborator, or user of knowledge. The agent can seek resources that alter their future capacities, not merely resources that solve the immediate problem.

Having said that, architecture-based motivation for effectance is not sufficient. Learners need to be deliberate about their learning. That is, after all, why I published Cognitive Productivity books.

New concepts for learning and epistemic agency

This may be a good point to pause for a moment and acknowledge that this paper introduces much terminology , which while not historically original, is new to most of my readers. We are not just introducing new words, but new concepts. My argument is that epistemic agency cannot be understood without these improvements beyond rudimentary cognitive psychology and folk psychology. Next, I continue to introduce concepts new to most readers.

Meta-effectiveness: improving the ability to improve

The concept of effectance leads to meta-effectiveness, the central construct of my Cognitive Productivity books. I define meta-effectiveness as the skills, dispositions, and underlying information-processing mechanisms—mindware—that enable and drive people to improve themselves. The motivation component of meta-effectiveness is effectance as described above (as architecture-based motivation).

Effectiveness concerns achieving valuable outcomes. It is the core concept of Albert Bandura’s concept of perceived self-efficacy, which itself is explicitly part of his theory of agency. (See his contribution to the canon of psychology, his book Self-Efficacy: The Exercise of Control .) Meta-effectiveness concerns improving one’s capacity to achieve valuable outcomes in the future. It is therefore recursive. A person may learn a fact, acquire a skill, or solve a problem; a meta-effective person also improves the processes by which facts are selected and retained, skills are developed, and problems are represented and solved.

This is not reducible to metacognition, although metacognition is part of it. Monitoring whether one understands a passage is metacognitive. Creating a better method for selecting what to read, assessing its potential value, delving into it, extracting knowledge gems, practising them, developing relevant habits, and applying them across contexts is a broader meta-effectiveness project.

Nor is meta-effectiveness reducible to self-regulated learning, as usually studied and discussed below. Self-regulated learning often focuses on learners’ planning, monitoring, and strategy use within educational tasks. Meta-effectiveness extends across adult development and professional life. It includes choosing cognitively potent tools, redesigning workflows, developing standards and habits, resisting illusions of comprehension and future recall, building external cognitive systems, and learning how to make knowledge usable when circumstances require it.

Epistemic agency without meta-effectiveness can become static. An agent may possess some effective epistemic practices while lacking the disposition and mechanisms required to revise them. A researcher may evaluate evidence carefully but manage sources poorly. A student may know that retrieval practice is effective yet never develop a routine for using it. A professional may be skilled in a familiar domain but avoid knowledge that would challenge established methods.

A general theory of epistemic agency should therefore include not only the exercise of epistemic competence but the implicitly sought and deliberate development of that competence.

Cognitive Productivity

Another concept is required…

Cognitive productivity is the effective use of knowledge resources to solve problems, develop products, and improve ourselves. It links knowledge to valued outcomes.

The term productivity can evoke quantity, speed, or economic output. That is not my intended meaning. For cognitive productivity is not just about efficiency, though efficiency is part of it. Cognitive productivity concerns productiveness, the production of objective knowledge (knowledge building), the solution of problems, self-improvement, and effectance. The concept of cognitive productivity is indicated by the subtitle of my first book: Using Knowledge to Become Profoundly Effective.

A person may consume enormous quantities of information while becoming no more capable of understanding a problem, improving a relationship, developing a theory, making a decision, or changing a harmful habit. Such a person is informationally busy but cognitively unproductive.

Cognitive Productivity provides a broader setting for epistemic agency than has previously been recognized because it encompasses the full trajectory from self-governance to action. Evaluation, covered by the Assess Analytically principle 3 below, is essential but insufficient. A well-assessed idea that cannot be remembered or accessed when relevant will not guide behaviour. A deeply understood theory that remains inert will not improve practice. A useful norm that never becomes a motivator or habit will not regulate conduct.

This is why the seven principles are best interpreted as a process model of epistemic agency.

The seven principles as a framework for epistemic agency

Here is a summary of the seven principles of Cognitive Productivity which I proposed in Cognitive Productivity with macOS: 7 Principles for Getting Smarter with Knowledge. I elaborate on them here because they are effectively principles required for full epistemic agency. Whereas this book illustrated the principles with macOS and iPhone technology, the ideas in that book, of course, apply to other platforms. (The book is not just theoretical, it contains very specific guidance, including over 60 videos. The more theoretical chapters are contained in Part 2 of my first book, Cognitive Productivity: Using Knowledge to Become Profoundly Effective.)

Principle 1. Lead Yourself with Knowledge

The first principle in Cognitive Productivity with macOS is Lead Yourself with Knowledge. This is not a motivational slogan attached to a book about software. It is an autonomous-agency claim.

To lead oneself is to govern one’s activities through an evolving understanding of what matters. Knowledge is not simply accumulated. It is recruited to interpret situations, choose directions, revise commitments, avoid foreseeable mistakes, and coordinate action across time. Epistemic agency therefore begins before formal evidence assessment. It begins when an agent aims to use knowledge to shape its own conduct, and does so.

To lead oneself entails pursuing a valued direction. This calls for a theoretical framework. The account of value I use draws from Ortony, Clore, and Collins’s 1988 cognitive theory of emotion updated in the 2022 revision of their highly influential book The Cognitive Structure of Emotions. It distinguishes three broad kinds of value. Projects and goals determine the desirability or usefulness of events. Norms provide standards for evaluating actions, including whether an action is praiseworthy or blameworthy.Preferences, or attitudes, determine attraction and aversion toward objects, people, ideas, experiences, and situations. They vary in appealingness. I refer to all information processing forms of value — projects and goals, norms, preferences, and needs — collectively as motivators. See Chapter 3 of my thesis..

This taxonomy matters because autonomous agency cannot be reduced to reflexive behavior. A researcher may be guided by a norm of intellectual honesty, a long-term project of understanding a problem, and a preference for elegant explanations. A physician may pursue the goal of identifying a treatment, act under professional norms governing consent and evidence, and develop preferences for methods that reduce avoidable burden on patients. A student may want a high grade while also accepting norms against plagiarism and gradually acquiring a genuine preference for deep understanding over superficial completion.

Motivators of the same and/or different types may cooperate or conflict. Knowledge helps agents understand those conflicts, predict their implications, and decide which commitments should govern action. It can also alter the motivators themselves. A person may learn that a cherished project is harmful, that a norm was poorly justified, or that a preference was manufactured by repeated exposure rather than reflective endorsement. Leading oneself with knowledge thus includes using knowledge to revise the motivational organization through which future knowledge will be processed.

This is one reason typical definitions of epistemic agency are too narrow when they centre primarily on knowledge claims. Agency requires not merely asking, “Is this claim true?” but also, “Why does this matter?”, “Which project does it serve?”, “Which norms govern my use of it?”, “What preferences are influencing my judgment?”, and “What kind of person or agent will I become by repeatedly acting this way?”

AI makes these questions visible in new, more dynamic and perhaps more urgent ways, but does not create them. A student deciding whether to delegate an essay to an AI system is negotiating projects, norms, and preferences. The project may be to finish the assignment, learn the subject, earn a credential, or avoid embarrassment. Norms may concern academic integrity, responsibility, disclosure, and fairness. Preferences may include convenience, intellectual challenge, originality, or admiration for technical novelty. Responsible refusal, emphasized in Horst’s proposal, is therefore not simply an epistemic judgment about AI reliability. It is an agentic resolution of competing motivators. (The section below on Assessing Analytically deals with a subset of epistemic agency dealing with evaluating knowledge resources – from AI or other).

To be clear: autonomous agency, including epistemic agency, requires self-leadership, which is what this principle is about.

Principle 2. Manage Your Cognitive Life Mindfully

The second principle concerns stewardship of cognitive conditions. Projects, knowledge resources, attention, time, tools, and interruptions must be managed in ways that reflect the agent’s values rather than the contingencies of the moment.

This principle is divided into 3 concerns:

• Managing Your goals and projects (e.g., with task management software such as OmniFocus), goal management being foundational to all forms of agency, including epistemic agency.

• Manage Your knowledge resources (e.g., with Finder, tagging software and information managers such as DEVONthink,

• Manage your attention (your information processing time) with time tracking software like Timing.app and my free mySelfQuantifier spreadsheet — attention and time being the most precious resources in knowledge work.

An agent may possess good intentions and valuable knowledge yet remain ineffective because attention is fragmented, commitments are poorly managed, relevant resources are inaccessible, or immediate demands repeatedly displace more important projects. Epistemic agency requires control not only over conclusions but over the conditions under which conclusions are reached.

Managing a cognitive life mindfully includes designing external systems and internal policies that make sustained, responsible knowledge work more likely. Epistemic agency without adequate cognitive management, as detailed in my books, is unproductive.

Principle 3. Assess Analytically

Much of the current literature on epistemic agency concentrates on (a small subset) of the territory of Principle 3. It focuses on evaluating evidence, assessing sources, verifying AI outputs, and determining whether claims are warranted. My principle Assess Analytically encompasses these matters but is considerably broader, and based on cognitive science.

It asks whether a knowledge resource is helpful and analyzes that helpfulness through the “CUPA” framework for assessing information resources.. This goes far beyond the traditional “CRAAP” test. The dimensions of CUPA are caliber, utility, potency, and appeal. Caliber concerns epistemic quality: accuracy, justification, argumentation, evidence, coherence, the adequacy of explanatory theories, etc. Utility concerns relevance to the agent’s projects and other motivators (discussed in the previous principle). Potency concerns the resource’s capacity to contribute to mental development—to generate concepts, skills, motivators, motive generators, habits, and other mindware. Appeal concerns the agent’s affective response and the need to interpret rather than blindly trust that response. Appeal is a potential red herring: not all truths are beautiful; not every proposition that is beautiful is true. Not every seductively appealing bit of information is high caliber: often alluring information, including from AI, is misleading.

You may have noticed that the forms of value of “CUPA” map onto the forms of value of Principle 1 (norms, goals and preferences), but are epistemically oriented —thus highly relevant to epistemic agency.

My book’s CUPA assessment framework includes general epistemic criteria, argument assessment, conceptual analysis, and explicit criteria for evaluating explanatory theories. It therefore goes beyond asking whether a source is credible. It asks whether an explanation has sufficient scope, depth, coherence, precision, plausibility, and power; whether it advances understanding; and whether rival accounts have been treated fairly. The fuller CUP’A framework is summarized here.

The CUPA framework has direct implications for AI literacy. Checking whether an AI has fabricated a citation is necessary but elementary. A sophisticated epistemic agent must assess the caliber, utility, potency and appeal of the AI output.

Principle 4. Surf Strategically

I examine this principle in unusual depth because it illustrates how a simple type of software (contextual information retrieval softwares) can apply cognitive science to support epistemic agency. Having said that, contextual information retrieval concerns only a fraction of what is required to “surf strategically”, and discussed in that principle. Every Cognitive Productivity principle is similarly grounded in cognitive science.

The fourth principle, Surf Strategically, addresses forms of interaction with knowledge resources that may appear superficial but are often indispensable. Knowledge workers do not—and should not—process every potentially relevant book, paper, message, webpage, dataset, or AI response in depth. Before deciding what deserves sustained attention, they must survey a field, inspect documents, skim passages, scan tables of contents, examine references, compare search results, classify resources, create links, file materials, and move rapidly among related items. These activities are shallow in the descriptive sense that they do not ordinarily produce deep comprehension. They are not intellectually trivial.

Surfing is a distinct phase of cognitive work with its own purposes and standards. It allows an agent to identify promising knowledge resources, estimate their potential CUPA (caliber, utility, potency, and appeal) before a more detailed assessment can be made, locate relevant passages, detect relationships among resources, and decide where scarce attention should be invested. Strategic surfing therefore requires judgment. One must know why one is surveying an informational space, what signals of value to look for, when a resource merits deeper processing, and when continued searching is unlikely to justify its cognitive cost. A person who delves deeply into everything is not more epistemically responsible than a person who skims intelligently. Indiscriminate depth can be as unproductive as indiscriminate superficiality.

This distinction matters because words such as skimming and browsing often carry a pejorative connotation. They can suggest impatience, distraction, or intellectual laziness. Yet sophisticated research depends on rapid, selective interactions with large bodies of information. A scholar reviewing a new literature may first inspect hundreds of titles and abstracts, follow citation trails, scan diagrams, search within documents, and compare competing terminologies before choosing a small number of works for careful study. A lawyer may rapidly inspect cases to determine which warrant full analysis. A physician may survey guidelines and abstracts before examining the most pertinent evidence. In each instance, superficially processed information guides the allocation of deeper cognitive effort.

Surfing (which in the Cognitive Productivity framework is not just “surfing the net”) also includes apparently mundane actions such as naming, tagging, filing, bookmarking, linking, and recording why a resource may matter. These actions are easily dismissed as clerical information management rather than epistemic activity. That would be a mistake. They shape the agent’s future access to knowledge and therefore influence what can later enter reasoning, deliberation, learning, and application. In an environment of informational abundance, deciding what should remain accessible—and constructing reliable routes back to it—is part of epistemic agency.

Surfing strategically can, and should, be supported by technology, such as using search engines. Such technology is reviewed in my book. In what follows we focus on a particular neglected aspect of strategic surfing: contextual information retrieval which is supported by our Hookmark app. (COI disclosure: I am CEO of CogSci Apps Corp. which develops the app.) But first we must consider the deep problem solved by contextual information retrieval.

The meta-access problem

This leads to what I call the meta-access problem. Epistemically agentic knowledge workers do not merely need access to individual resources. They need rapid access to the other resources that are relevant to whatever they are currently doing. While drafting a paper, for example, a researcher may need its project plan, outline, source PDFs, notes, correspondence, figures, task list, conceptual diagrams, bibliographic records, and relevant webpages. These resources are commonly scattered across the file system, the web and multiple applications. The challenge is therefore not simply to find a document in isolation. It is to access a changing constellation of contextually relevant resources without disrupting the cognitive work that made them relevant.

Traditional search remains essential, but it places a series of ancillary demands on cognition. The user must suspend the superordinate task, formulate a query, remember identifying details, invoke a search interface, inspect results, distinguish the intended item from plausible alternatives, open it, and then reconstruct the cognitive context that was interrupted. Each step consumes time and mental resources. The problem is especially severe because working memory is severely limited and the activation of information in working memory declines rapidly. When the agent knows that a relevant resource exists but cannot readily recall where it is stored, what it was called, or which application contains it, their productivity suffers — not just their efficiency but their productiveness.

I coined the expression contextual information retrieval to describe access in which the current informational context helps determine what resources are presented or made immediately reachable. Unlike ordinary (random-access) retrieval, which begins primarily with a query, contextual information retrieval starts from the item or activity already occupying the foreground of cognition. In computer science, this is known as content addressable memory rather than a random access memory. The question becomes not merely “Where is the information?” but “What information is relevant here (to this document), and how can I reach it with minimal cognitive disruption?” I discuss the concept and the meta-access problem more fully in “Contextual information retrieval: The missing link in knowledge work”.

This is where Hookmark becomes theoretically relevant rather than merely convenient. Hookmark enables knowledge workers (students or professionals) to create and use robust links among resources in different applications. A draft can be linked bidirectionally to its outline, source documents, notes, relevant email, task list, diagrams, bibliographic record, and other resources. When Hookmark is invoked in the context of the draft, it presents links associated with that context; when invoked on one of the linked resources, it can provide a route back. A 2023 paper of mine illustrates this with both a PDF surrounded by related notes, tasks, diagrams, emails, ebooks, and practice materials, and a Word draft connected to the resources required to develop it.

Hookmark is therefore not simply a bookmark manager (though it is also the world’s first contextual bookmark manager). It implements a form of user-controlled associative (non-random) memory. The user deliberately or implicitly records relationships among conceptual artifacts, creating multiple retrieval routes that remain available when the context changes. This complements random-access search. Search is indispensable when the retrieval target is unknown or when the user needs to explore a large corpus. Contextual retrieval is especially valuable when relationships among resources have already been recognized and should remain available for future cognition. (Hookmark also contains a growing recommendation system where, soon with AI, the software guesses what information the user might need next.)

Hookmark’s ability to create deep links to particular segment of a PDF or Word document takes this capability even further. A global link to a fifty-page PDF is useful; a link to the exact passage under consideration is even more useful. Deep links reduce the need to reopen a document, reconstruct the original location, search within it, and visually scan for the relevant material. Similar principles apply to locations in ebooks, videos and webpages and other structured resources. These forms of access help preserve the agent’s orientation toward the superordinate project rather than forcing repeated detours through the mechanics of retrieval.

Contextual retrieval and consciousness

It should be clear that every principle of the meta-effectiveness framework grounds epistemic agency and is informed by cognitive science. Here I discuss how contextual information retrieval relates to research on consciousness and long-term working memory (a key cognitive feature of expertise).

The significance of contextual information retrieval becomes clearer when considered in relation to Merlin Donald’s account of consciousness in A Mind So Rare: The Evolution of Human Consciousness (2001). Donald rejects the idea that consciousness can be understood solely as a momentary biological phenomenon confined to a brief working-memory interval. Human consciousness depends on the integration of information across multiple neural systems, other people, and external symbolic resources. It operates not only over the immediate present but across what Donald describes as intermediate and long-term temporal ranges.

External symbolic technologies are central to this expanded form of consciousness. Writing, diagrams, books, computer displays, and other conceptual artifacts allow information to be stabilized outside the brain and repeatedly reintroduced into conscious processing. As Donald later observed, a computer screen can function as a temporary external working-memory field: the user enters an interactive loop with the display while thinking, writing, and creating. Contextual information retrieval tightens this loop by reducing the friction involved in bringing relevant external representations back into consciousness.

I have therefore argued that software such as Hookmark can extend intermediate and long-term consciousness. This does not mean that the software is itself conscious or that a link becomes part of the biological mind. The claim is functional. Contextual links enable an agent to reconstitute, across hours, days, months, or years, configurations of information that had previously supported a cognitive project. They help restore not just a file but a working intellectual context.

Consider a researcher returning to a manuscript after several weeks. Without contextual retrieval, the manuscript may be available while the surrounding cognitive system has largely disappeared. The researcher must remember or rediscover which email contained an important objection, which PDF supported a claim, where the outline was stored, which diagram represented the argument, and what tasks remained unresolved. Hookmark can make these resources immediately available from the manuscript itself. In doing so, it helps the researcher reconstruct an extended episode of knowledge work and resume it with less loss of orientation.

This is an extension of intermediate consciousness because it supports integration over the minutes and hours required to work across multiple artifacts during an active session. It is an extension of long-term consciousness because deliberately constructed relationships can persist and become available much later, after the corresponding contents have faded from biological working memory and can no longer be reliably recalled unaided.

The point is not simply that retrieval becomes faster (as is characteristic of long-term working memory). Speed matters because every additional retrieval step competes with the contents and control states that must be maintained during demanding knowledge work. But contextual retrieval also affects what kinds of cognitive projects are feasible. If it is difficult to reconnect a source with separate notes, diagrams, tasks, correspondence, and practice materials, knowledge workers may avoid producing those artifacts in the first place. They may keep inadequate notes, forgo conceptual diagrams, or abandon useful connections because they anticipate difficulty retrieving them later. The meta-access problem can therefore constrain knowledge building and learning before any explicit search failure occurs.

By lowering this barrier, contextual information retrieval can support a richer cognitive ecology. It becomes more practical to take persistent notes in the app of one’s choice, link them to the source, associate a paper with the projects it informs, connect a claim to evidence and criticism, and attach productive-practice challenges to the knowledge they are meant to develop. The result is not merely a more orderly information system. It is an environment in which conceptual artifacts can more readily participate in ongoing cognition.

Contextual information retrieval software calls for all software to be link-friendly. It needs to be possible to form links not only to web pages, but emails, task lists, PDFs, diagrams, and more. I have led an effort involving more than 20 cognitive scientists and software developers to ensure software is link-friendly. That is the Manifesto for Ubiquitous Linking. It is structurally modeled on the famous and influential Agile Manifesto for software development. Its rationale section details its grounding in cognitive science in highly accessible terms.

Surfing as epistemic control

Surfing strategically should therefore be understood as a cognitive science-informed form of epistemic control. It governs the transitions among informational breadth, selective access, and depth. It includes deciding where to look, what to sample, what to preserve, what to connect, what to retrieve, and what deserves delving. The epistemically agentic researcher does not treat all superficial processing as suspect. Rather, the researcher distinguishes superficiality caused by distraction or avoidance from deliberate shallow processing that serves a larger cognitive strategy.

This principle also qualifies familiar concerns about digital reading. Digital environments can certainly encourage fragmented attention and compulsive switching. But switching among resources is not inherently a cognitive failure. Complex knowledge work often requires the integration of heterogeneous materials. The relevant question is whether those transitions are controlled by the agent’s projects, motivators and priorities or by the attention-capturing (superficially appealing) designs of applications, feeds, and platforms. Contextual information retrieval supports the former by allowing the agent to construct and preserve personally meaningful routes through an informational environment.

For epistemic agency, then, access is not a secondary convenience that comes after inquiry, evaluation, and knowledge construction. It is one of their enabling conditions. Agents can assess, integrate, practise, and apply only the knowledge resources they can bring into an appropriate cognitive context. A general theory of epistemic agency must therefore encompass not merely the appraisal of accessible claims but the design and management of access itself.

Principle 5. Delve Deeply

Delving is my general term for deeply processing a knowledge resource, whether one reads a paper, listens to a lecture, studies a diagram, watches a video, explores software, participates in a demanding conversation, or interacts with AI. English has many medium-specific verbs (reading, watching, etc.) but lacked a satisfactory generic term for this essential cognitive activity.

Delving involves much more than surfing. It includes identifying the structure of a resource, clarifying concepts, reconstructing arguments, connecting the material with prior knowledge, generating questions, recording insights, determining what deserves further mastery or application, etc.

This principle corrects an important weakness in discussions of access to information. Easy access does not imply understanding. Generative AI can compress, summarize, and explain, but it can also allow users to move quickly past the cognitive work through which concepts become integrated into mindware. Epistemic agency includes deciding when convenience is appropriate and when deep engagement is indispensable.

My Cognitive Productivity with macOS book demonstrates with many videos, in detail, how to delve deeply into material, such as with the most power PDF reader, Skim which has amazing inline annotation features, and by creating meta-docs. Meta-docs are documents about knowledge resource, such as textual notes and diagrammatic notes. One can use Hookmark to bidirectionally link a central resource (PDF, web page, video, AI chat, etc.) to one’s meta-docs. (In fact, one can simultaneously create a meta-doc, in the app of one’s choice, and hook it [meaning bidirectionally link, it] in a single command, called Hook to New. This illustrates the interrelatedness of the seven principles, in this case surfing and delving.)

Principle 6. Practice Productively

Productive practice is deliberate practice and retrieval-based learning directed toward the development of mindware. It is not limited to memorizing declarative facts. It can be used to develop concepts, skills, norms, attitudes, motivators, motive generators, monitors, propensities, and habits.

This is where the transformative conception of learning becomes practical. Suppose a learner encounters a knowledge gem: an idea with the potential to improve judgment or action. Simply highlighting it or agreeing with it is unlikely to produce durable change. The learner may need to retrieve it repeatedly, discriminate the conditions in which it applies, practise generating it from meaningful cues, use it in examples, and reflect upon failures to apply it.

Productive practice is therefore designed not just for rote learning but for architectural change. It can cultivate a norm by rehearsing its implications, a motive generator by practising detection of relevant situations, or a habit by repeatedly coupling a cue to a suitable response. It can help turn declarative knowledge into a propensity to act.

Productive practice taps into a deep heuristic the brain uses to determine what to learn, which I call the heuristic relevance-signaling hypothesis. During sleep, the brain decides to learn information (e.g., to make easy to retrieve) information which the cortex has tried to retrieve. Retrieval practice is a cue for what to learn. This idea comes from J.R. Anderson in his seminal 1990 book, The Adaptive Character of Thought. I elaborate this in chapter 7 of my first Cognitive Productivity book.

Most good students know they need to practice, often with flashcards. But knowledge workers tend to fall short. Productive practice is especially important for professional knowledge workers, who often continue reading and attending conferences long after they have stopped practising. They may possess sophisticated explicit knowledge while failing to transform the monitors, habits, and dispositions that govern work under pressure.

The concept and specification of practicing productively explicitly draws on literature from cognitive science literature in different areas: test-enhanced learning, memory testing effects, retrieval practice, spaced learning, distributed practice. The literature is expounded on in chapters 7 of my first book Cognitive Productivity: Using Knowledge to Become Profoundly Effective. Because my second book is more applied. Having said that chapters 13 and 14 of my first book describe the applied concept.

My Cognitive Productivity books explain in great depth, with many videos, how the world’s most powerful flashcard app, Anki, can be used for productive practice. Remnote is also a contender productive practice app, though not described yet in my books.

Principle 7. Apply Knowledge

The final principle of Cognitive Productivity with macOS is Apply Knowledge. Its position is deliberate.

Application is sometimes treated as a relatively low level in taxonomies of cognition. Bloom’s original and revised taxonomies have often encouraged educators to regard remembering and application as lower than evaluation and creation. After all, they figure in low levels of the taxonomies. My framework rejects the implication that application is merely a routine operation.

Productive application of knowledge calls for the crown jewel of learning, transfer of learning beyond the original context in which the information was acquired. Transfer of learning is discussed in cognitive science-based fashion in detail in Cognitive Productivity: Using Knowledge to Become Profoundly Effective. I recommend the interested reader search for “transfer” in that book.

Application is the culmination of cognitive productivity. Knowledge is cognitively productive (not inert) when it helps solve problems, develop products, improve theories, make decisions, transform practices, help others, and improve the agent. Applyingsophisticated knowledge to a novel, ambiguous, value-laden situation can require every other capacity in the framework: self-leadership, mindful management, analytical assessment, strategic search, deep understanding, and productive mastery.

As Mark McDaniel remarked “The primary gauge of expertise is transfer performance. […] How to train to promote transfer is the most fundamental challenge of education.” Transfer of learning is the ability to use knowledge in contexts different from when it was learned. See amongst many other sources on this subject, Haskell’s (2000). Transfer of learning: Cognition and instructionbook.

The ultimate test of much World 2 knowledge (as mindware) is whether it can inform intelligent action. That does not imply a crude instrumentalism in which every idea must generate an immediate practical payoff. The application may be to improve an explanation, refine a preference, acquire a norm, reinterpret an experience, understand another person, or prepare for future learning. But cognition does not reach its culmination merely when an artifact has been created or a proposition has been accepted. It culminates when knowledge becomes available to the autonomous agent in ways that can make a difference.

Individual and shared epistemic agency

The account developed here is intentionally centred on the individual autonomous agent. That is a choice of theoretical level, not a denial of social cognition.

The knowledge building tradition and Damşa et al.’s theory of shared epistemic agency illuminate collective processes that my framework does not attempt to explain fully. Groups can generate structures, practices, norms, and products that cannot be reduced to the isolated behaviour of their members. Shared epistemic agency is therefore a legitimate and important research topic in its own right.

Nevertheless, collective agency depends partly on the internal capabilities of individual participating agents. Groups need members who can assess resources, represent goals, detect knowledge gaps, manage attention, understand others’ contributions, revise commitments, and apply ideas. Even a well-designed knowledge-building environment cannot substitute completely for undeveloped individual epistemic capability.

The relation should therefore be treated as complementary. Knowledge building explains how communities take collective responsibility for advancing ideas as public, “third world” objects. Cognitive Productivity explains how individual agents use knowledge resources to lead themselves, develop, and contribute effectively. A later, more comprehensive theory could examine how individual and collective epistemic agency recursively shape one another. The present article deliberately does not cast that wider net.

Relation to research on self-regulated learning

Research on self-regulated learning has, in my view, made some of the most important contributions to understanding epistemic agency, even though its researchers rarely use that term. Among the most influential researchers in this tradition are Philip Winne of Simon Fraser University where I am an adjunct professor of Education, and Allyson Hadwin. Although Winne does not to my knowledge use the phrase epistemic agency, his work arguably provides one of its richest information-processing accounts. Their excellent chapter, Studying as Self-Regulated Learning, models studying as four recursively connected phases: defining the task, setting goals and planning, enacting tactics and strategies, and adapting future learning. Their COPES framework—conditions, operations, products, evaluations, and standards—shows how learners actively interpret tasks, monitor discrepancies, exercise control, and revise the cognitive structures that guide subsequent learning. They expounded on COPES in their influential paper, Self-Regulated Learning Viewed from Models of Information Processing. Although they do not use the phrase epistemic agency, their learner is unmistakably an autonomous, self-directing epistemic agent rather than a passive recipient of instruction.

Winne’s equally important chapter, Self-Regulated Learning Viewed from Models of Information Processing, develops this account in greater architectural detail. His SMART processes—searching, monitoring, assembling, rehearsing, and translating—together with explicit models of tactics, strategies, motivation, monitoring, and feedback, provides one of the most sophisticated information-processing theories of learner agency in the educational psychology literature. He explicitly interprets information-processing control theories as theories of agency: learners represent goals, compare alternatives, monitor progress, and regulate their own cognition.

Winne’s work and H-CogAff share an important common foundation: both model humans as autonomous, goal-directed agents who actively regulate their own cognition through monitoring, feedback, strategy selection, and adaptation. The Cognitive Productivity framework builds directly on this agentic tradition. Where it differs is primarily in scope. Winne and Hadwin focus on how students regulate learning and studying. Cognitive Productivity asks how autonomous agents regulate their entire relationship with knowledge across formal education, informal and lifelong learning, professional knowledge work, research, design, decision-making, and self-improvement. It also places self-regulated learning within a broader integrative design-oriented architecture that explicitly incorporates motivators, motive generators, habits, architecture-based effectance, meta-management, and meta-effectiveness.

Thus, self-regulated learning is not an alternative to epistemic agency but one of its most important foundations, while Cognitive Productivity seeks to provide a broader theory of epistemic agency itself.

Human–AI relations reconsidered

I believe the foregoing account changes how we should think about AI and epistemic agency. (Incidentally, production of this account is an instance of knowledge building).

The core issue is not simply whether AI possesses agency or whether responsibility should be “shared” between humans and systems. A more immediate question is how interaction with AI affects the architecture and development of the human agent. Does it improve or weaken analytical assessment? Does it create habits of verification or passive acceptance? Does it stimulate deeper questions or truncate inquiry? Does it help users develop concepts, motivators, and motive generators, or merely provide satisfactory-looking outputs? Does it increase meta-effectiveness, or does it conceal deficiencies that the learner might otherwise detect?

However, more important than how AI currently does affect the agent is how agents could use AI for cognitive productivity: to build knowledge, solve problems and improve themselves. This I believe is the objective of Rachel Horst’s recent IDG grant-funded research project at the University of British Columbia.

AI can support every principle of Cognitive Productivity. It can help articulate motivators, plan projects, inspect cognitive routines, assess arguments, survey literatures, clarify difficult texts, generate practice challenges, and explore applications. It can also undermine every principle by displacing self-leadership, fragmenting attention, flattering weak reasoning, encouraging shallow processing, simulating mastery, and completing tasks that users needed to practise.

The appropriate stance is therefore neither rejection nor unconditional partnership. AI should be assessed as a knowledge resource and cognitive tool in relation to its caliber, utility, potency, and appeal. Its role should be selected according to the projects, norms, preferences, capabilities, and developmental needs of the agent. Sometimes extensive use will be justified. Normally, verification will be required. Sometimes, as Horst’s proposal emphasizes, refusal to delegate will be epistemically necessary.

But refusal is only one element of a general theory. A person may refuse AI assistance and remain a poor epistemic agent. Conversely, a person may use AI extensively while exercising sophisticated self-leadership, assessment, learning, and application. The relevant question is not simply whether AI was used but what kind of cognitive process and personal development its use supported.

A proposed general conception of epistemic agency

The existing literature has illuminated important aspects of epistemic agency. Knowledge building has shown how learners and communities can take responsibility for advancing ideas. Research on shared epistemic agency has explained how groups regulate sustained work on common knowledge objects. Science-education research has connected agency to authentic disciplinary practices. Assessment research has foregrounded judgment and responsibility. AI-related work has made verification, epistemic vigilance, and refusal newly urgent.

In sum, my proposal is not to replace these contributions but to provide them with a broader, deeper integrative design-orientedtheoretical foundation supported by practical recommendations for using knowledge and technology in a cognitively productive manner in pursuit of valued motivators, while developing their own capabilities.

On this account:

Epistemic agency is the capacity of an autonomous agent to direct and improve its interactions with knowledge resources in the service of its projects and goals, norms (standards), and attitudes (preferences). This includes managing its processing of knowledge resources: surfing strategically (selecting, skimming, retrieving, etc), assessing analytically, delving deeply, practicing productively, and applying the knowledge (solving problems, building new knowledge and improve themselves). The agent’s self-improvement ideally involves developing the motivators, motive generators, habits, and other mindware through which future knowing and acting occur.

Cognitive Productivity names the effective exercise of this capacity. Meta-effectiveness names the agent’s capacity to recursively improve it. Architecture-based effectance helps explain why agents are motivated, sometimes implicitly sometimes explicitly, to do so. The seven principles provide an organized practical framework through which epistemic agency can be developed and expressed.

Conclusion

Epistemic agency has become an important idea because contemporary learners and knowledge workers confront an unprecedented abundance of information, increasingly powerful cognitive tools, and growing pressure to delegate thinking. But the concept will remain fragmented unless it is connected to a theory of autonomous agency.

An integrative design-oriented account starts with the agent as a whole. The agent is not merely a holder of beliefs, of long-term memory and procedural knowledge, or a participant in classroom discourse. The agent is a motivational, affective, cognitive, self-regulating, developing system. The agent pursues projects and goals, responds to norms, develops preferences, forms habits, generates new motivators, and can sometimes reflect upon and redesign its own methods.

From that perspective, learning should not merely transmit knowledge or even build knowledge. It should transform agents. It should help them develop the concepts, skills, standards, preferences, motivators, motive generators, habits, and reflective capabilities required to continue improving after formal education has ended.

That is why epistemic agency needs concepts of Cognitive Productivity and meta-effectiveness. And it is why the seven principles begin with leading oneself through knowledge and end not with passive possession, nor even with creation alone, but with applying knowledge to become more effective and to make valuable differences in the world.

To summarize this paper:

1. Epistemic agency presupposes autonomous agency.

2. Learning should transform the architecture of the agent.

3. Cognitive Productivity provides a broader practical framework than existing epistemic agency accounts.

Thanks for reading. Please stay tuned for a journal article on this topic! Please contribute to the discussion ↓ and share this article with others.

What if the most important question in education is not “What should students learn?” but “What conditions allow meaningful learning to emerge?”

That, I believe, is the central insight of Ron Burnett’s remarkable A Biography of Learning.

I recently submitted an eight-page review of the book to Interchange: A Quarterly Review of Education. This article summarizes the main argument of my review while I await the publication process.

Reading A Biography of Learning was a pleasure because it challenged me in unexpected ways. My own work has focused heavily on cognitive productivity, artificial intelligence, and the cognitive science of learning. Burnett’s book led me to reflect more deeply on something that can easily be overlooked: the educational environments, relationships, and cultural conditions that make meaningful learning possible in the first place.

Burnett offers something unusual: neither a conventional education textbook nor a technical treatise on cognitive science, but a deeply reflective exploration of learning as a lifelong, relational, and often unpredictable process. Rather than presenting learning as a linear progression toward predetermined outcomes, he invites us to see it as an evolving biography shaped by curiosity, relationships, uncertainty, culture, technology, and serendipity.

The book is especially timely. At a moment when education is being transformed by artificial intelligence, analytics, and growing demands for measurable outcomes, Burnett reminds us that not everything of educational value can be planned, standardized, or optimized. Genuine learning frequently emerges through exploration, conversation, unexpected connections, and serendipitous encounters.

One of the book’s great strengths is its resistance to false dichotomies. Burnett is not arguing against structure, assessment, or technology. Rather, he argues for balance—between structure and emergence, linearity and non-linearity, analytics and reflection, technological innovation and human relationships, and efficiency and exploration. Throughout the book, he returns to the idea that educators should devote as much attention to cultivating the conditions under which learning can flourish as to specifying the knowledge students are expected to acquire.

My review is overwhelmingly positive. At the same time, I offer one gentle critique. Given the book’s breadth and ambition, I would have welcomed greater engagement with research in educational psychology, cognitive science, and artificial intelligence—fields that increasingly illuminate how memory, motivation, retrieval, expertise, emotion, and learning environments interact. This omission is understandable. Burnett had already undertaken an ambitious project, and incorporating these literatures in depth might have required another volume. My review therefore points to several areas of educational psychology that are compatible with, and could further enrich, Burnett’s central vision.

Readers familiar with my own work will notice considerable common ground with ideas I develop in Cognitive Productivity. Burnett focuses primarily on the conditions and environments in which meaningful learning can emerge. Cognitive Productivity asks how individuals can use knowledge, tools, and relationships more effectively to solve problems, create valuable products, and improve themselves. It draws explicitly on cognitive psychology, educational psychology, affective science, and artificial intelligence. I see the two books as complementary: Burnett helps us think about the environments in which learning flourishes, while Cognitive Productivity examines how people can act effectively within those environments.

Ultimately, A Biography of Learning is a hopeful book. It reminds us that education is not merely the transmission of information, nor simply the production of measurable outcomes. It is the cultivation of people capable of continued growth, reflection, exploration, and participation in a changing world. In an era increasingly captivated by metrics and algorithms, Burnett makes a compelling case for preserving the complexity, humanity, and openness that make learning worth pursuing.

Readers interested in the practical implications of these ideas may also enjoy my more applied book, Cognitive Productivity with macOS: 7 Principles for Getting Smarter with Knowledge, which examines how people can use contemporary knowledge tools more effectively. My forthcoming book, Discontinuities: Love, Art, Mind, extends related ideas into the domains of literature, film, and other forms of art.

If you are interested in education, cognitive science, artificial intelligence, or the future of higher education, I highly recommend reading Ron Burnett’s A Biography of Learning.

This essay develops one such perspective. It is not a survey of every theory of intelligence. Rather, it draws on an integrative design-oriented tradition in cognitive science and AI—associated with Herbert Simon, Allen Newell, Merlin Donald, Aaron Sloman, Keith Stanovich, Robert White, Ortony, Clore and Collins, and others—to ask how we should think about artificial intelligence today. My own contribution to that tradition began with my 1994 doctoral dissertation, Goal Processing in Autonomous Agents, and has continued through work on motivation, perturbance, cognitive productivity, and knowledge technologies.

I often encounter someone confidently asserting that artificial intelligence either is intelligent or isn’t really intelligent. The discussion rarely lasts long before both sides begin talking past each other. One person points to AI systems solving difficult problems, writing software, composing prose, or passing examinations. Another replies that they merely predict tokens, lack consciousness, have no values, or do not truly understand anything. These are not trivial objections. But they are usually made before the more basic scientific question has been addressed.

Whenever someone tells me that AI isn’t really intelligent, I ask two questions. First: how are you defining intelligence? Second: what theory of intelligence are you using? The first question sometimes produces an answer. The second almost never does. That is understandable. Outside cognitive science, we use the word intelligence without needing an explicit theory. We readily say that Einstein was more intelligent than the average person, that a raven is more intelligent than a pigeon, or that a dog is more intelligent than a worm. Ordinary language works well enough for ordinary purposes. But scientific concepts do not become adequate merely because we sharpen their dictionary definitions.

Much public discussion is concerned with verbal adequacy: finding the “correct” definition of intelligence. Cognitive science seeks theoretical adequacy: explanatory frameworks that organize observations, generate predictions, guide research, and inform design. Definitions summarize theories; they do not replace them. (Richard Feynman famously observed that knowing the name of a bird in every language tells us almost nothing about the bird itself. Likewise, defining “intelligence” more precisely does not, by itself, explain intelligent systems. Scientific understanding comes from theories that explain how they work. Definitions summarize theories; they do not replace them. Naming a phenomenon—even defining it carefully—is not the same as explaining it. One’s definition must appeal to other theoretical constructs.) Asking whether AI is intelligent without first specifying a theory of intelligence is rather like asking whether an animal is healthy without first having a theory of physiology. The question is not meaningless, but it is scientifically underdetermined.

The argument of this essay goes one step further. Even a theory of intelligence is not the deepest explanatory framework. Human-like intelligence is best understood within a broader computational theory of human-like autonomous agency.

Take-away

Before asking whether AI is intelligent, we need to define intelligence within the context of a scientific theory of intelligence. More fundamentally, we need a theory of human-like autonomous agency within which intelligence can be understood. As discussed at the end of this essay, the same considerations hold for “consciousness.”

Why “intelligence” became the wrong starting point

It is understandable that public discussion has gravitated toward intelligence. After all, the field itself is called Artificial Intelligence. The name naturally directs attention toward one aspect of minds while leaving others in the background. Had the founders instead chosen a name such as Artificial Autonomous Agents or Computational Agency, public discussion might well have developed differently. We might spend less time debating whether AI is “really intelligent” and more time asking what kinds of autonomous agents current systems are becoming, what motivational architectures they possess, what regions of the space of possible minds they occupy, and what new discontinuities they may eventually cross.

This is not merely a semantic point. Names encourage questions, and questions guide research. The historical label Artificial Intelligence has survived enormous changes in AI itself: from symbolic problem solving, theorem proving, expert systems, connectionism, reinforcement learning, robotics, and cognitive-affective architectures to today’s foundation models and increasingly autonomous software agents. The label still has practical value. But it can also mislead us into treating intelligence as a single property that systems either possess or lack.

Today’s AI landscape includes systems that converse fluently, prove theorems, control robots, generate software, diagnose diseases, compose music, retrieve information, plan actions, and coordinate activities over extended periods of time. These systems differ profoundly from one another. Some exhibit extraordinary linguistic competence but little endogenous motivation. Others pursue complex objectives but have relatively modest reasoning abilities. Some operate almost entirely reactively; others deliberate extensively before acting.

The more interesting scientific question is whether these systems instantiate different kinds of computational architectures. By a computational architecture, cognitive scientists mean the organization of interacting mechanisms that make intelligent or autonomous behaviour possible: mechanisms for perception, memory, learning, motivation, planning, action, communication, reflection, and so forth. Just as the architecture of a building concerns the organization of its parts, not merely the materials from which it is made, a computational architecture concerns how the components of a mind or artificial agent are organized and how they interact.

Once we adopt this perspective, it becomes less useful to ask whether “AI” as a whole is intelligent. “AI” is now an umbrella term for a rapidly expanding family of systems. The better question is what kinds of computational architectures these systems embody, and how those architectures compare with those of animals, humans, organizations, and possible future machines.

Take-away

The name Artificial Intelligence encourages us to focus on intelligence. A more productive scientific question asks what computational architectures current AI systems instantiate, and what kinds of autonomous agents they are becoming.

Autonomous agency: the deeper scientific question

A system may be impressive without being very autonomous. A calculator can outperform most humans at arithmetic. A chess engine can defeat grandmasters. A search engine can retrieve information faster than any person. These systems exhibit capabilities that matter. But a human-like autonomous agent is not merely a device that produces excellent outputs when prompted. It is a system that must regulate its own activity in a changing world under constraints of limited time, limited information, limited working memory, and limited resources including other people and computational resources.

My interest in these questions is not recent. More than thirty years ago, my doctoral dissertation, Goal Processing in Autonomous Agents, which I believe is even more relevant today than it was then, examined the computational requirements for systems capable of generating, managing, and pursuing top-level and derived goals under severe constraints of time, information, and computational resources. A top-level goal is not simply a subgoal produced by planning. It arises from the agent’s motivational architecture. Derived goals, by contrast, are generated in the service of other goals. Planning a trip, for example, may produce the derived goals of buying a ticket, packing a bag, and arranging transport to the airport. But the top-level motivation to travel—to visit a loved one, attend a conference, escape danger, or explore a new place—comes from elsewhere in the architecture.

This distinction is crucial for AI. Many AI systems can derive subgoals. Some can decompose tasks, make plans, seek information, revise intermediate steps, and monitor progress. That is significant. But deriving subgoals from an externally supplied task is not the same as generating, regulating, and revising one’s own top-level motivations. One of the most important discontinuities in the space of possible minds separates systems that merely pursue externally assigned goals from systems capable of generating, managing, revising, suspending, and abandoning their own top-level motives.

Notice how the question changes once we adopt this perspective. Instead of asking whether a system is intelligent, we ask what sort of autonomous agent it is. For example:

• Can it generate and regulate its own top-level goals or motivators, or does it merely pursue goals supplied by others?

• Can it derive, revise, suspend, and abandon subordinate goals as circumstances change?

• Can it revise priorities when new information arrives?

• Can it interrupt one activity because another has become more urgent?

• Can it improve its own competence over time?

• Can it reflect upon its own reasoning and modify it?

• Can it coordinate all this while operating under severe constraints of time, information, working memory, and computational resources?

These are not merely behavioural questions. They are architectural questions. They ask what kinds of mechanisms must exist inside a system, how those mechanisms interact, and what forms of control they make possible. Within such a framework, human-like intelligence may, to a first approximation, be characterized as the capacity to acquire, represent, assess, integrate, and apply knowledge in pursuit of multiple top level and derived motivators across changing environments with limited resources (time, knowledge, money, other people, etc.). This is not intended as a complete definition. It is a working characterization within a broader computational theory of human-like autonomous agency.

The explanatory order matters. Intelligence does not explain autonomous agency. Rather, autonomous agency explains why intelligent capacities are needed and how they must be integrated with perception, action, memory, motivation, executive control, affect, communication, and reflection. A human-like autonomous agent is not a disembodied head solving puzzles. It is a system situated in a world, contending with resources, opportunities, interruptions, needs, values, and competing motives.

Take-away

Human-like intelligence should be characterized within a theory of human-like autonomous agents. Such agents do not merely solve problems; they generate, manage, and pursue top-level and derived goals under resource constraints.

Figure 1. Levels of explanation and the space of possible minds:

Figure take-away

The question “Is AI intelligent?” belongs near the top of the hierarchy. Its answer depends on deeper theories of human-like autonomous agency, computational architecture, and the space of possible minds.

The space of possible minds

The preceding section moved the question from intelligence to human-like autonomous agency. That move is necessary, but not sufficient. We also need to resist another tempting simplification: the idea that minds can be arranged along a single scale from less intelligent to more intelligent. That picture is sometimes useful in ordinary life. It lets us say that one person solved a problem more intelligently than another, or that one animal has greater problem-solving ability than another. But as a scientific picture it is too flat.

Aaron Sloman has long argued that cognitive science and AI should explore the space of possible minds: a space of possible computational architectures, possible niches, and possible mappings between them. This is not merely a poetic phrase. It is a methodological warning. If there are many possible kinds of minds, then the task of cognitive science is not to find a single essence called intelligence, but to understand the requirements, architectures, trade-offs, and discontinuities that define different regions of that space. This is illustrated by the right side of Figure 1 above.

(My book in progress Discontinuities: Love, Art, Mind and already for sale will end with a chapter on discontinuities.)

This changes the AI question. The usual public debate asks whether AI has crossed a threshold into intelligence. A more sophisticated version asks how intelligent AI is. But the design-space perspective asks a different question: what kind of mind is this? That question does not assume a single ladder. It assumes a structured landscape containing many kinds of minds, many kinds of computational architectures, and many important discontinuities.

Some discontinuities are obvious once we notice them. A purely reactive system differs qualitatively from one capable of deliberation. A system that merely pursues externally supplied goals differs qualitatively from one that can generate and regulate its own top-level motivators. A system that can plan differs qualitatively from one that can monitor and modify its own planning. These are not merely differences in “amount of intelligence.” They are differences in architecture.

Merlin Donald’s work is especially important here. Donald did not treat human intelligence as merely more of the same animal intelligence. In A Mind So Rare: The Evolution of Human Consciousness, Donald described major transitions in human cognitive evolution, including new forms of consciousness and new ways of governing cognition over time. His distinction among sensory binding, short-term control, and intermediate or long-term governance is particularly useful for our purposes. It reminds us that human consciousness is not merely momentary awareness; it is a multilevel control system capable of sustaining projects, meanings, and symbolic structures over extended periods.

Donald’s point also helps explain why external symbolic systems matter. Writing, diagrams, maps, mathematical notation, books, databases, hyperlinks, and other knowledge technologies do more than store information. Properly integrated into human activity, they help sustain context, guide attention, coordinate long-term projects, support cognition over extended periods of time, and, in that sense, extend the functional reach of human consciousness.

External symbolic systems, such as writing, diagrams, maps, mathematical notation, and contextual information retrieval systems, do not replace working memory. Rather, they augment executive function by reducing the need to retain arbitrary contextual information internally and by making relevant information rapidly retrievable when needed. Our Hookmark Mac and iPhone app, for example, provides bidirectional links that help users rapidly recover the context surrounding digital resources, thereby supporting planning, problem solving, and other executive functions. This extends the intermediate and long-term awareness of human consciousness, concepts developed by Merlin Donald.

This Donaldian point matters for AI because it reminds us that human-like intelligence is not merely a matter of solving more difficult problems. Human cognition acquired new architectural possibilities through gesture, imitation, language, external symbolic storage, and culture. These were not simply increments on a scale. They changed what kinds of minds humans could have, introducing successive discontinuities in the space of natural minds.

Take-away

Minds do not occupy a single scale from unintelligent to intelligent. They occupy a structured space of possible computational architectures, with important discontinuities between reactive, deliberative, motivational, reflective, culturally scaffolded, and future artificial forms of mind.

Computational architectures for human-like autonomous agents

To speak of computational architectures is not to indulge in metaphor. It is to ask what kinds of organized mechanisms are required for human-like autonomous agency. Such agents must perceive, act, learn, remember, generate motivators, choose among competing demands, deliberate, monitor themselves, interact with their environments, and develop over time. No single mechanism—language modelling, reinforcement learning, planning, memory retrieval, symbolic inference, neural pattern completion—can by itself explain such an agent.

This was one of Sloman’s central points. Work in AI and cognitive science often studies components: vision, language, learning, planning, motor control, memory, or reasoning. Those studies are valuable. But the deeper problem is how such components can be assembled into a coherent working system. Sloman emphasized that the most important artificial and evolutionary “design” choices for human-like agents concern the overall architecture, because detailed questions about mechanisms and representations are best addressed in the context of a global design.

A useful first approximation is the distinction between reactive processes, management processes, and meta-management processes (see chapter 4 of my thesis). A reactive subsystem responds quickly, often automatically, to internal or external conditions. Management processes, including deliberative processes, construct, compare, and evaluate possible actions or plans before committing to them. Meta-management processes, including reflective processes, monitor and evaluate management processes themselves. Reflection adds another level: the ability to notice that one’s own thinking is going poorly, that one is wasting time, that a strategy is biased, that a problem should be postponed, or that a previously adopted goal should be questioned.

Cognitive psychologists often use the term executive functions to refer to the family of processes responsible for regulating thought and action. In the present framework, executive functions are implemented primarily by management processes and meta-management processes. They include, among others, evaluative functions such as assessing the importance, urgency, insistence, intensity, relevance, and expected consequences of competing motivators; deliberative functions such as planning, scheduling, prioritization, conflict resolution, commitment to action, and the scheduling of deliberative and reflective activity; executive control functions such as directing attention, inhibiting inappropriate thoughts or actions, regulating behaviour, allocating computational resources, and selecting when to engage in Type 1 or Type 2 reasoning (see Dual-Process Theories of Higher Cognition - Perspectives on Psychological Science - APS); reasoning, problem-solving, and explanatory functions such as inference, hypothesis generation, explanation, analogical reasoning, diagnosis, and planning under uncertainty; meta-management (reflective) functions such as monitoring one’s own thinking, detecting errors, recognizing bias, revising strategies, changing mental sets, and deciding when to continue, interrupt, postpone, or abandon ongoing cognitive activity; and ambiguity-management functions such as recognizing uncertainty, tolerating ambiguity, gathering additional evidence, maintaining multiple competing interpretations, and delaying commitment until sufficient evidence is available.

Most of these executive functions rely heavily on working memory: the limited-capacity workspace in which information is actively maintained, manipulated, integrated, and evaluated. Executive functions do not operate in isolation. They continuously interact with working memory, long-term memory, perception, motivator generators, insistence-based filters, and external symbolic systems.

These distinctions should not be mistaken for a rigid pipeline. (Illustrated in chapter 4 of my thesis). The mind is not a factory line in which perception hands a package to motivation, which hands it to deliberation, which hands it to action. A computational architecture specifies possible interactions among mechanisms, not a single mandatory sequence of processing. Perception, memory, internal monitoring, communication, motivator generation, deliberation, and reflection can operate asynchronously and influence one another in multiple directions. Human-like cognition is event-driven, interruptible, opportunistic, and often messy.

That messiness is not a defect in the theory. It is one of the requirements any serious theory must explain. A system that waits for a tidy sequence of central decisions before responding to the world will not be very human-like. Human-like autonomous agents need fast reactive responses, slower management processes, meta-management processes that can inspect and regulate management, working memory to support active cognition, and multiple mechanisms that can generate potential motivators and interrupt ongoing activity when something more urgent or relevant arises.

These distinctions are directly relevant to AI. A system that produces fluent answers may lack robust deliberative management. A planning agent may lack reflective self-monitoring. A reinforcement-learning system may learn policies without being able to articulate, evaluate, or revise its own reasons. A chatbot may simulate reflection linguistically without possessing a stable architecture for self-monitoring across time. Conversely, future AI systems may combine language, planning, memory, perception, action, tool use, working memory, and self-monitoring in ways that cross new architectural discontinuities.

The integrative design-oriented question is therefore not “Does it seem intelligent?” but “What computational requirements for human-like autonomous agency does this system satisfy, and what architecture would explain both its capabilities and its limitations?” Once we ask that question, successes and failures become informative. Hallucinations, brittle planning, perseveration, lack of initiative, overconfidence, susceptibility to misleading prompts, and inability to manage long-term projects are not merely performance glitches. They are clues about architecture.

Take-away

Human-like intelligent behaviour must be explained in relation to the computational architecture that produces it. Such an architecture is not a rigid pipeline, but an interacting system of reactive, motivational, executive, reflective, affective, memory, learning, and action processes.

Figure 2. Sketch of an integrative design-oriented computational architecture for human-like autonomous agency :

This figure depicts interacting computational processes rather than a processing pipeline. It updates and extends a related architecture, H-CogAff, presented in Part II of Cognitive Productivity: Using Knowledge to Become Profoundly Effective and in several papers by Aaron Sloman (example: Sloman 2006, “How many separately evolved emotional beasties live within us?”).

Where do goals come from?

We can now ask a question that has received surprisingly little attention—not only in public discussions of AI, but also within cognitive science itself.

Where do goals come from?

The question sounds almost naïve until one tries to answer it. Much of cognitive science has concentrated on perception, memory, language, learning, reasoning, and planning. AI has likewise devoted enormous effort to algorithms for search, optimization, reinforcement learning, and, more recently, foundation models and autonomous software agents. Yet all of these capabilities presuppose that the agent has something to perceive, remember, reason about, or plan for.

The challenge is deeper than explaining how an agent pursues goals. It is to explain how a human-like autonomous agent generates, organizes, prioritizes, transforms, and sometimes abandons goals in the first place. That, I believe, remains one of the central scientific challenges for both AI and cognitive science.

One influential tradition has sought to explain motivation in terms of reward maximization or utility. Such ideas have proved enormously fruitful, particularly in economics and reinforcement learning. But they provide only part of the story. Human motivation is far richer than the pursuit of a single utility function. We pursue knowledge, beauty, friendship, justice, truth, curiosity, craftsmanship, status, duty, exploration, love, and countless other ends. These motivations often cooperate, but they also conflict, evolve, and reorganize throughout a lifetime. (See Psychological Hedonism Meets Value Pluralism: An Integrative Design-oriented Perspective)

From the perspective developed in this essay, this diversity is not an inconvenience to be abstracted away. It is a clue about architecture.

One thing is certain: the human mind-brain contains multiple motivator generators (generators of goals, motives, projects, standards, and attitudes).

Take-away

The central problem is not merely how intelligent systems pursue goals. It is how human-like autonomous agents generate, organize, regulate, and transform the motivations that make goal pursuit possible.

Architecture-based motivation

Aaron Sloman introduced the profound idea of architecture-based motivation to capture precisely this point. Motivation should not be viewed merely as the output of a single decision-making process, nor as the consequence of one global reward signal. Rather, motivation is itself produced by the architecture.

This becomes clearer if we consider ordinary human life. A person may suddenly remember an unanswered email, become curious about an unexpected observation, worry about a child, notice an inconsistency in an argument, recall an unfinished promise, feel compelled to repair a broken object, or become fascinated by an idea encountered while reading. Some of these motivators arise from biological needs. Some arise from social commitments. Some arise from learned values, professional responsibilities, personal projects, aesthetic preferences, moral standards, or attachment structures. Others arise because perception, memory, reflection, or internal monitoring has detected something relevant.

They are not all computed by a single deliberative process asking, “What action maximizes expected reward?” Nor are they all derived through means-end reasoning. Rather, they are generated continuously by multiple asynchronous motivator-generating processes distributed throughout the computational architecture. These motivator generators generate, modify, reactivate, and regulate motivators; they also assign insistence and intensity. Deliberative processes may subsequently assess, elaborate, reconcile, prioritize, transform, or inhibit these generated motivators, but deliberation is only one contributor to motivation—not its sole source.

This distinction is fundamental. It moves motivation from the periphery of cognitive science to its architectural core. It also helps explain why human-like autonomous agency cannot be modeled adequately as a sequence beginning with a goal and ending with an action. Goals themselves are products of architecture. They arise, compete, fade, recur, and sometimes become urgent through interacting processes distributed across the agent.

The distinction between top-level and derived goals now becomes indispensable. Derived goals are generated in the service of other goals. If I decide to submit a paper, I may derive goals to revise the abstract, check references, format figures, and send an email. But the top-level motivation to write the paper may arise from effectance (including curiosity), professional commitment, desire to contribute knowledge, social obligation, or some mixture of values. The architecture does not merely reason from goals. It generates and manages the goals that reasoning serves.

This is also where contemporary AI becomes interesting. Many current systems are increasingly capable of deriving subgoals from prompts. They can decompose tasks, call tools, monitor progress, revise plans, and recover from failures. These are impressive achievements. But this should not be confused with possessing a rich architecture of endogenous motivation. A system that derives subgoals from a user prompt is architecturally different from one that can generate top-level motivators of its own, assign insistence and intensity to them, reconcile them with existing goals, standards, and attitudes, and regulate them over time.

We can already imagine such systems. Consider a household robot that is vacuuming when it detects that its owner has collapsed and is in physiological distress. Rather than merely continuing its assigned task, it asynchronously generates a new top-level motivator to summon medical assistance. Or imagine that it detects an intruder entering the home. It generates a different top-level motivator—to protect its owner—which immediately interrupts its previous activity and recruits perception, deliberation, communication, and action toward the new concern. These are not merely new subgoals in service of vacuuming. They are new top-level motivators generated by the architecture itself in response to changing circumstances.

Take-away

Human-like autonomous agents do not merely pursue goals. They require architectures that generate, assess, prioritize, inhibit, and transform motivators. This is why motivation is not an optional add-on to intelligence.

Insistence, intensity, and executive controllability

It is tempting to speak of a motivator’s “strength.” But this is too crude. Several dynamic properties of motivators need to be distinguished.

Insistence is the degree to which a motivator persistently competes for limited cognitive resources, including attention, working memory, executive processes, and reflective processes. Intensity is the degree to which a motivator tends to recruit or energize behavioural systems, thereby increasing the disposition toward overt action. Importance is the value attributed to the motivator by executive functions. (These dimensions are expanded upon in chapter 3 of my thesis and briefly in Sloman, 1987) Executive controllability is the extent to which management and meta-management processes can regulate a motivator, its behavioural expression, or its influence on cognition.

In humans, these dimensions can dissociate. In obsessive-compulsive disorder, an intrusive thought about harming a loved one may have very high insistence: it repeatedly intrudes into attention and working memory. But it may have very low (behavioural) intensity: the person is horrified by the thought and has little or no disposition to act on it. Conversely, some action tendencies may have high intensity but low executive controllability, as in many motor tics associated with Tourette syndrome. Reflexively withdrawing one’s hand from a hot stove has high intensity and low deliberative control, but it may not be very insistent once the action is complete. One’s management processes may assign high urgency and/or high importance to a motivator that is not yet very insistent or intense (e.g., the important need to lose weight might not drive action). AI may or may not support all these distinctions.

These distinctions help explain why understanding motivation requires more than assigning a single strength to a motive. Some motivators dominate thought without producing action. Others produce action without prolonged thought. Still others quietly shape long-term life projects without constantly interrupting attention.

Take-away

Motivators differ not merely in “strength,” but in insistence, intensity, and executive controllability. This distinction is crucial for understanding perturbance, intrusive thought, impulsive action, self-control, and future AI motivation.

Motivators: goals, standards, and attitudes

Once we recognize that motivation is architecturally generated, another question immediately arises: what kinds of motivators must a human-like autonomous agent be able to generate and regulate?

Here it is helpful to draw on Ortony, Clore and Collins’ The Cognitive Structure of Emotions, and on the treatment of values in my Cognitive Productivity with macOS: 7 Principles for Getting Smarter with Knowledge. I use the term motivator to refer to a value-bearing control state: something the agent treats as relevant or insistent, consciously or unconsciously, explicitly or implicitly. Motivators help drive assessment, attention, goal formation, planning, and action. In my terminology, motivators also draw attention to the fact that values are not merely labels attached to things. They can spawn assessments and other motivators, including goals.

To a first approximation, motivators come in three different flavors: goals and projects, standards, and attitudes. This taxonomy is partly based on Ortony, Clore and Collins, and I developed it in Principle 1 of Cognitive Productivity with macOS: 7 Principles for Getting Smarter with Knowledge as a practical way of understanding values, motivation, and self-governance.

Goals and projects are states the agent is willing to work to achieve, preserve, accelerate, delay, or avoid. They vary in desirability. A scientist may have the goal of explaining a phenomenon. A student may have the goal of mastering a concept. A robot may have the goal of reaching a destination while conserving energy. Goals can be top-level or derived. They are often arranged not in simple hierarchies but in complex networks of means and ends.

Standards are norms, rules, ideals, or constraints that the agent treats as things that ought or ought not to hold. They vary in praiseworthiness and blameworthiness. “Be honest,” “do not mislead the reader,” “keep promises,” “respect evidence,” and “do not endanger others” are standards. Standards regulate conduct, constrain goal pursuit, and provide the basis for emotions such as guilt, shame, admiration, resentment, and indignation.

Attitudes are likes, dislikes, preferences, interests, aversions, tastes, fascinations, and affective biases. They vary in appeal. One may like a melody without having the goal of hearing it now; dislike a food while having the goal of eating it for health reasons; or be fascinated by a topic before having any explicit project involving it. Attitudes often guide attention and learning before explicit goals are formed.

These three kinds of motivators interact continuously. Standards can generate goals: “I ought to apologize, therefore I will apologize.” Goals can modify attitudes: “I began studying mathematics for instrumental reasons, but came to love it.” Attitudes can generate goals: “That topic fascinates me; I want to learn more.” Standards can constrain goals; goals can override attitudes; attitudes can bias the selection of goals; and all three can influence the assessment of events, actions, objects, people, and ideas. Human-like motivation is therefore not a simple hierarchy of goals. It is a dynamic network of interacting motivators. However, these forms of value can vary independently as well. One may view a state of affair as desirable but not praiseworthy, for instance.

This point is crucial for AI. Human-like intelligence is not merely the capacity to pursue goals. It also involves the capacity to generate, represent, revise, coordinate, and sometimes reject networks of goals, standards, and attitudes. A system that optimizes for a goal but lacks standards and attitudes may be powerful, but it is not human-like in the relevant sense. Conversely, a future AI system with goals, standards, and attitudes would raise a much richer set of questions about agency, rationality, alignment, responsibility, and the space of possible minds.

Take-away

Human-like autonomous agents are not merely goal-directed. They are also norm-governed and attitude-shaped. Human-like intelligence requires the capacity to generate and regulate networks of goals, standards, and attitudes.

Effectance, meta-effectiveness, and epistemic agency

This architectural perspective also helps explain why intelligence should not be understood merely as a collection of abilities. Highly intelligent people do not simply solve problems well. They often display a persistent tendency to become more capable over time.

Robert White called one important aspect of this tendency effectance): the motivation toward competence. In Cognitive Productivity: Using Knowledge to Become Profoundly Effective, I argued that effectance is better understood architecturally than behaviourally. Effectance is not merely a conscious or even unconscious desire to improve oneself. Rather, it is a propensity to generate top-level motivators whose pursuit tends to increase competence, effectiveness, and understanding.

Someone driven by effectance may decide to learn a new mathematical technique, master a musical instrument, study philosophy, understand a difficult scientific paper, improve a piece of software, refine a workflow, or simply ask a better question. None of these activities need be motivated explicitly by the thought “I want to become more intelligent.” Yet together they gradually reshape the architecture of the mind.

Human-like intelligence therefore includes a developmental dimension. Intelligent agents do not merely use knowledge; they acquire it, organize it, evaluate it, integrate it, apply it, and continually improve their capacity to use it effectively. This developmental perspective lays part of the foundation for what I believe should become an integrative design-oriented theory of epistemic agency: a theory explaining how human-like autonomous agents acquire, organize, evaluate, integrate, and apply knowledge in ways that continually increase their competence and effectiveness.

One might think of effectance as the motivation to become more competent, and meta-effectiveness as the skills and dispositions of becoming better at becoming effective. Meta-effectiveness encompasses not only the acquisition of knowledge and skills, but also the cultivation of the dispositions, habits, executive functions, and external symbolic supports required to apply them appropriately. Thus effectance is a subset of meta-effectiveness. Meta-effectiveness is the core concept of my first book, Cognitive Productivity: Using Knowledge to Become Profoundly Effective.

This perspective aligns naturally with Keith Stanovich’s distinction between intelligence and rationality. Rational thought depends not only on cognitive abilities, but also on knowledge and the dispositions to deploy those abilities when appropriate. Similarly, meta-effectiveness concerns developing the knowledge, skills, and dispositions that enable autonomous agents to use what they know effectively in pursuit of their goals, standards, and attitudes. Some educational and practical implications of this perspective are explored in Cognitive Productivity with macOS: 7 Principles for Getting Smarter with Knowledge, where I propose seven principles of meta-effectiveness — augmenting human cognition through knowledge resources and information technology.

A competent agent can solve problems. An effectant agent seeks to become more competent. A meta-effective agent systematically develops the knowledge, skills, dispositions, executive functions, and external symbolic supports that make continual improvement possible.

Take-away

Human-like intelligence is not merely current problem-solving ability. It includes motivational architectures, including effectance, that continually develop competence. Meta-effectiveness extends this developmental orientation into the disciplined cultivation of knowledge, skills, dispositions, executive functions, and external supports.

Attachment structures

Human-like agency is not merely individual problem solving. Humans form enduring relationships with people, places, communities, disciplines, projects, and ideals. These relationships are not just memories. They reshape the architecture.

In a 1996 paper with Ian Wright and Aaron Sloman, we modernized the notion of attachment structures in the context of a design-oriented analysis of grief. (At the time we used the expression “design-based.”) The central point was that an attachment is not merely a feeling. It is a distributed computational structure embedded throughout an autonomous agent. Such a structure may include stored knowledge, expectations, preferences, habits, plans, predictive models, and specialized motivator generators that influence perception, deliberation, and action.

Attachment structures illustrate an important principle of human-like intelligence: long-term relationships do not merely produce memories. They reorganize the motivational architecture itself. An attachment to a person may alter what one notices, what one worries about, what one plans, what one treats as important, and what one feels compelled to protect. An attachment to a research programme may generate questions, projects, standards, and intellectual commitments over decades. An attachment to a moral ideal may repeatedly generate motivators that override convenience, fear, or short-term reward.

Attachment structures gradually modify and create motivator generators. They therefore influence which goals, standards, and attitudes arise in future situations, how insistent those motivators become, how intense their action tendencies are, and how readily they gain access to executive resources. This helps explain why attachment can be so powerful, and why grief, betrayal, loss, or separation can become perturbing.

This also sets a high bar for AI. A system that remembers a user is not thereby attached to that user. A human-like robot capable of forming attachments would need to develop distributed structures that alter its motivator generators, predictive models, standards, priorities, and executive control over time. Such a system would not merely store a profile. It would become differently organized because of a relationship.

Take-away

Attachment structures are distributed computational organizations that reshape motivator generators over time. Human-like AI would require more than memory of persons or projects; it would need architectures capable of forming, maintaining, revising, and sometimes dismantling attachments.

Designed and emergent affect

We can now directly address a topic that public discussions of intelligence often mishandle: emotion. It is common to treat emotion as the opposite of intelligence, or as something that disrupts rational thought. That picture is far too simple. In a human-like autonomous agent, affect is not an optional decorative layer added after cognition has done its work. It is part of the control architecture through which values, needs, concerns, and motivators influence attention, deliberation, action, and learning.

However, we should not treat “emotion” as a single thing either. Some affective mechanisms may be explicitly built into an architecture, whether by evolution or by human designers. Alarm mechanisms, pain systems, attachment mechanisms, curiosity mechanisms, reward systems, and other specialized motivator generators are examples of computational mechanisms that asynchronously generate or modify motivators. They help determine what the agent notices, avoids, approaches, protects, repairs, learns, or pursues.

Other affective phenomena are not best understood as built-in modules. They emerge from interactions among architectural components. This distinction was central to Aaron Sloman’s early work on computational theories of emotion. (See Sloman & Croucher(1981)- You don’t need a soft skin to have a warm heart and Sloman & Croucher (1981) - Why robots will have emotions.)

The point is not that programmers must insert a “fear module,” a “grief module,” or a “jealousy module” into an intelligent system. Rather, if an architecture contains specialized motivator generators, mechanisms for assigning insistence and intensity, limited cognitive resources, interruption mechanisms, learning, memory, executive functions, and reflective processes, then some emotional phenomena, including mental perturbance, may emerge as system-level patterns.

Mental perturbance is a state in which one or more highly insistent motivators repeatedly recruit attention and other cognitive resources, making it difficult for the agent to disengage and sustain alternative activities. It is not merely an emotion, nor merely a thought pattern. Mental perturbance, an architectural phenomenon, is produced when motivators, insistence, attention, working memory, executive functions, and limited computational resources interact over time. Perturbance may manifest as worry, rumination, grief, craving, obsessive planning, anger, shame, or other forms of persistent mental preoccupation, depending on the motivators involved. In particular, this architectural dynamic provides an architectural explanation for repetitive thought: highly insistent motivators repeatedly gain access to executive resources, making disengagement difficult despite competing goals, standards, and attitudes.

Thus, this perspective complements psychological theories of rumination and worry, and more generally Watkins’ (2008) framework of repetitive thought, by proposing a computational architectural account of why repetitive thought occurs: it reflects the recurrent recruitment of limited executive resources by highly insistent motivators.

Perturbance is not another component in a computational architecture. Rather, it is an emergent architectural phenomenon arising from interactions among motivators, executive functions, working memory, attention, and limited computational resources. This is one reason perturbance is theoretically useful. Rather than explaining persistent worry, rumination, grief, craving, or obsessive thought by positing a separate cognitive module, it explains them as recurrent patterns emerging from the dynamics of the architecture.

This point matters for AI as much as for cognitive science. We should not merely ask only whether an AI system “has emotions.” That question is too crude. We should ask which affective control mechanisms are explicitly designed into the system, which motivational processes it contains, whether any of its internal dynamics can generate persistent concern-like states, and which emotional phenomena, if any, could emerge from its architecture. The answer may differ radically across systems.

Take-away

Affect is not the enemy of intelligence. In human-like autonomous agents, some affective mechanisms are architecturally designed, while some affective phenomena, such as mental perturbance, emerge from interactions among motivators, insistence, attention, working memory, executive functions, and limited resources.

Intelligence is not rationality

We are now in a position to address another common confusion. Intelligence is often treated as though it implied rationality. If someone is highly intelligent, we expect them to think well, make good decisions, update their beliefs, avoid foolish errors, and resist obvious biases. Yet everyday experience and psychological research both show that this expectation is unreliable. Highly intelligent people can reason badly. They can be dogmatic, impulsive, biased, overconfident, inattentive to evidence, or unwilling to reconsider cherished beliefs.

Keith Stanovich has done more than almost anyone to clarify this distinction. In What Intelligence Tests Miss: The Psychology of Rational Thought, he argues that intelligence tests measure important cognitive abilities, but they do not adequately measure rational thought. People may have high cognitive ability while lacking the knowledge or thinking dispositions required to seek disconfirming evidence, consider alternatives, override impulsive responses, or use probabilistic reasoning when appropriate.

The architectural perspective developed here helps explain why Stanovich’s distinction is so important. Earlier, we distinguished between management processes and meta-management processes. Management processes construct, compare, and evaluate possible actions or beliefs. Meta-management processes monitor and regulate management processes themselves; they notice that one is reasoning poorly, wasting time, perseverating, being biased, or pursuing the wrong goal. But rationality depends not merely on having reflective mechanisms. It also depends on knowledge, values, dispositions, and motivations to recruit them when needed.

Stanovich’s theory of rationality fits naturally into a broader theory of human-like autonomous agency. In fact, he proposed an information processing architecture that is inspired by and similar to (but simpler than) Sloman’s H-CogAff architecture adapted here. (See figure 2.1, page 33 of Stanovich’s 2011 book, Rationality and the reflective mind). A person may possess the cognitive ability required to solve a reasoning problem, yet fail to engage the reflective processes that would lead to a better answer. They may be tired, anxious, angry, socially pressured, overconfident, or insufficiently motivated to think carefully. They may also lack effectance with respect to reasoning itself: the motivation to improve their own thinking habits, learn from errors, and cultivate better epistemic practices.

Thus, rationality is not a mysterious extra faculty added to intelligence. It depends on an architecture in which reflective processes, motivation, values, norms, knowledge, skills, and learned dispositions interact.

This also illustrates why perturbance matters. A highly perturbed mind (say: in love, angry or experiencing grief) may have ample intelligence and even strong reflective capacity, yet still struggle to reason well because highly insistent motivators keep capturing attention and working memory. Worry, anger, craving, shame, or grief can repeatedly redirect cognitive resources toward a dominant concern. In such cases, irrationality does not arise from lack of intelligence alone. It arises from the dynamics of the architecture.

The same lesson applies to AI. An AI system may perform well on reasoning benchmarks while lacking robust mechanisms for epistemic self-regulation. It may produce plausible answers without appropriately monitoring uncertainty. It may revise plans without understanding when to question the goals that generated them. It may simulate reflection in language without possessing stable dispositions to use reflective processes across contexts. Again, asking whether AI is intelligent tells us too little. We must also ask what kind of rationality, if any, its architecture supports.

Take-away

Intelligence and rationality are distinct. Rationality depends not merely on cognitive ability, but on reflective processes, thinking dispositions, motivational support, values, knowledge, and the regulation of perturbing concerns.

What this means for today’s AI

We can now return to the question with which we began. Is AI intelligent?

The answer is not a simple yes or no. Some AI systems exhibit some forms of intelligence, in some respects, according to some scientifically defensible theories of intelligence. Large language models can integrate information, generate explanations, write code, summarize documents, translate languages, and solve many problems that would once have seemed to require intelligence. Artificial autonomous agents can decompose tasks, call tools, maintain intermediate state, and pursue objectives over time. Robots can perceive, act, adapt, and learn in physical environments.

But these facts do not settle the question. They refine it.

The issue is not whether a system has crossed a single threshold called intelligence. The issue is which computational requirements for human-like autonomous agency it satisfies, which it lacks, and what kind of architecture explains its pattern of successes and failures. What are its capabilities and mechanisms that go beyond humans? A system may be linguistically fluent without being reflectively rational. It may derive subgoals without generating its own top-level motivators. It may optimize actions without possessing standards or attitudes. It may simulate emotional language without having affective control mechanisms or emergent perturbance. It may display competence without effectance or meta-effectiveness.

This is why both triumphalist and dismissive claims about AI tend to mislead. It is not illuminating to say simply that AI is intelligent. Nor is it illuminating to say that it is “not really intelligent.” Both statements compress too much. They ignore the space of possible minds.

A more useful approach asks:

• What kind of AI system are we discussing?

• What computational architecture does it instantiate?

• Which forms of human-like intelligence does it exhibit?

• What kinds of motivators, if any, does it generate?

• Can it regulate top-level and derived goals?

• Does it possess standards and attitudes, or merely optimize supplied objectives?

• What reflective mechanisms does it have?

• What thinking dispositions does it reliably display?

• What forms of rationality does its architecture support?

• Can it form attachment structures?

• Does it exhibit effectance or meta-effectiveness?

• Where does it lie in the space of possible minds?

These questions are more difficult than asking whether AI is intelligent. They are also more fruitful.

Take-away

The scientifically useful question is not whether AI is intelligent in general. It is which aspects of human-like intelligence and autonomous agency particular AI systems exhibit, and which computational architectures explain them.

An integrative design-oriented approach

The perspective developed here reflects what I have elsewhere called an integrative design-oriented approach to cognitive science and AI. The phrase matters. It is not merely the traditional design stance applied to minds. It is an attempt to integrate multiple disciplines, multiple levels of explanation, and multiple design requirements in order to understand human-like autonomous agents.

The integrative design-oriented approach begins not with dictionary definitions, but with requirements. What must an architecture be able to do in order to support human-like autonomous agency? It must perceive, act, learn, remember, generate motivators, evaluate alternatives, manage goals, use working memory, reflect on its own processes, regulate affect, develop competence, form attachments, interact with external symbolic systems, and coordinate with other agents. No single discipline can explain all of this. Psychology, AI, neuroscience, philosophy, anthropology, education, evolutionary theory, and software design all have something to contribute.

The integrative design-oriented approach replaces arguments about labels with questions about requirements, mechanisms, architectures, development, and emergence. It seeks not merely to classify minds, but to explain how different forms of autonomous agency become possible.

That, to my mind, is the deeper lesson AI is forcing upon us. Artificial intelligence has not simply challenged our understanding of machines. It has challenged the adequacy of our theories of mind. If we want to understand AI, we need better cognitive science. And if cognitive science wants to understand minds, it needs richer theories of computational architecture, autonomous agency, motivation, affect, executive function, working memory, attachment, rationality, and development.

Take-away

An integrative design-oriented approach asks what computational requirements and architectures make human-like autonomous agency possible. It treats intelligence as one aspect of a broader scientific problem.

Large language models have changed the debate

It would be a mistake to read this essay as minimizing the achievements of modern AI, particularly large language models such as ChatGPT, Claude, Gemini, and others. They represent one of the most significant developments in the history of artificial intelligence.

These systems can explain complex ideas, summarize books and scientific papers, generate software, translate between languages, critique arguments, adapt their writing style to different audiences, brainstorm new ideas, tutor students, help researchers explore unfamiliar literatures, and collaborate with people on extended intellectual projects. In my own work, including writing this essay, I have found ChatGPT to be an extraordinarily valuable thinking partner for brainstorming, criticism, literature exploration, and improving scientific writing. These are not trivial accomplishments. They deserve to be recognized as genuine forms of intelligent information processing.

Indeed, one of the most remarkable features of large language models is not merely what they can do in isolation, but what they can accomplish in sustained interaction with human users. A productive dialogue often becomes a joint cognitive process in which the human contributes goals, background knowledge, judgment, and evaluation, while the AI contributes retrieval, synthesis, reformulation, analogy generation, and the rapid exploration of alternative ideas. The result is frequently better than either participant could have achieved alone.

Recognizing these remarkable capabilities, however, does not bring us any closer to answering the question, “Is AI intelligent?” On the contrary, it exposes the inadequacy of the question. Large language models exhibit extraordinary strengths in some forms of cognition while remaining limited in others. They have therefore made it even more important—not less—to distinguish among different forms of intelligence, different theories of intelligence, and different computational architectures.

Ironically, the success of large language models strengthens rather than weakens the central argument of this essay. They have shown that intelligent behaviour can emerge in ways that many researchers did not anticipate. Rather than forcing us to abandon cognitive science, they challenge us to develop richer theories capable of explaining both biological and artificial forms of intelligence. The scientific task remains the same as the core of this essay argues: It is no longer to decide whether AI is or is not intelligent. It is to understand what kinds of intelligence different systems possess, how those capacities arise, and how they can best complement and extend human intelligence.

So, is AI intelligent?

So where does all this leave the original question?

One should no longer feel compelled to answer the question “Is AI intelligent?” with either a simple yes or no, or by asking to what quantitative degree it is intelligent. More scientifically productive questions concern the computational requirements for different forms of autonomous agency, the architectures that satisfy those requirements, and the capabilities, limitations, developmental trajectories, and emergent phenomena that follow from those architectures.

Some AI systems exhibit some forms of intelligence, in some respects, according to some scientifically defensible theories of intelligence. Others do not. Future systems will almost certainly occupy regions of the space of possible minds that have never previously existed on Earth.

The question itself is therefore not wrong. It is simply incomplete. Before we can answer it responsibly, we must ask what kind of AI system we are talking about, what theory of human-like intelligence we are using, what computational architecture the system embodies, what forms of value and motivation it possesses, what reflective capacities it exhibits, and where it lies in the space of possible minds.

Those questions do not make the debate disappear. They transform it from an argument about words into a scientific inquiry about minds.

The goal is not to settle the meaning of the word “intelligence.” It is to understand the architectures that make different forms of intelligence and autonomous agency possible.

The next time someone tells you that AI is—or is not—really intelligent, you could ask: What kind of AI? According to what theory of intelligence? And what computational architecture are we talking about?

That is where a serious conversation can begin.

What about consciousness and emotions?

This essay has focused on intelligence and autonomous agency. The same points I made here apply to concepts of consciousness and emotion. In A Mind So Rare: The Evolution of Human Consciousness, mentioned above, Merlin Donald made the same case for consciousness: there is a space of possible minds supporting different forms of consciousness. It’s not a matter of “this animal or machine has consciousness” or “it doesn’t have consciousness,” but what forms of consciousness does the agent have?

Similarly, it is silly to debate whether an AI agent can experience emotions or not without qualifying the question with respect to a particular integrative design-oriented theory of emotion.

Why You Can’t Make a Machine That Feels Pain

This paper was partly inspired by Daniel Dennett’s paper, “Why You Can’t Make a Machine That Feels Pain” (answer because the concept of pain is polymorphous). Here the same idea/pattern is used, generalized and extended and applied to intelligence and consciousness.

Glossary of Key Concepts

This essay develops and integrates terminology from AI, cognitive science, psychology, philosophy, and education. The following glossary summarizes key concepts as they are used here. Many are adapted from prior work; some are refinements introduced in this essay. This could be extended, e.g., to define the forms of awareness specified by Donald in A Mind So Rare (selective binding, short-term control, and intermediate- and long-term governance), rationality and other key terms.

• Attachment structure. A distributed computational organization that develops through repeated interaction with particular people, projects, organizations, places, or ideals. Attachment structures influence perception, memory, motivator generation, insistence assignment, planning, action, and executive control, thereby reshaping the motivational architecture over time.

• Autonomous agent. A system capable of perceiving, acting, learning, generating and regulating motivators (both top-level and derivative), managing competing demands, and sustaining coherent behaviour over time under constraints of limited information, time, working memory, and computational resources.

• Cognitive productivity. Efficience and effectiveness in using knowledge to solve problems, acquire new knowledge, develop expertise and deliver services.

• Computational architecture. The organization of interacting computational mechanisms that together enable an autonomous agent to function. An architecture specifies interacting components and processes, not a fixed sequence of operations.

• Effectance. The architecture-based motivation to become more competent. Effectance contributes to learning, self-improvement, and the development of expertise.

• Epistemic agency. The capacity of an autonomous agent to acquire, organize, evaluate, integrate, apply, and improve knowledge in pursuit of its motivators. More generally it involves cognitive productivity.

• Executive controllability. The extent to which management and meta-management processes can regulate a motivator, its behavioural expression, or its influence on cognition.

• Executive functions. The family of management and meta-management processes responsible for regulating affect, cognition, motivation and behaviour. In this paper, executive functions are understood architecturally rather than as a single faculty.

• Human-like autonomous agency. The capacity to function as a human-like autonomous agent (see “autonomous agent” above), integrating perception, action, memory, working memory, motivator generation, executive functions, affect, learning, reflection, attachment, and development within a coherent computational architecture.

• Insistence. The degree to which a motivator persistently competes for limited cognitive resources, including attention, working memory, management processes, and meta-management processes.

• Insistence-based motivator filters. Architectural mechanisms that regulate which currently active motivators gain access to limited executive resources on the basis of their insistence and other contextual factors.

• Intensity. The degree to which a motivator tends to recruit or energize behavioural systems, thereby increasing the disposition toward overt action.

• Integrative design-oriented approach. A scientific approach to understanding autonomous agency by integrating evidence and theory across disciplines while analysing computational requirements, architectures, mechanisms, development, and design, using the design stance. Regarding the design stance, part of the IDO approach, see Sloman (1993) - Prospects for AI as the general science of intelligence but replace “design-based” with “design stance”. See also Dennett 1987 The Intentional Stance.

• Management processes (deliberative processes). Executive processes responsible for planning, reasoning, scheduling, conflict resolution, resource allocation, and other forms of executive control.

• Mental perturbance. An emergent architectural phenomenon in which one or more highly insistent motivators repeatedly recruit executive processes, making disengagement difficult; i.e, interrupting and influencing attention, working memory, deliberation, memory, and action.

• Meta-effectiveness. Skills, knowledge and dispositions (effectance) to use knowledge to become a more effective person.

• Meta-management processes (reflective processes). Executive processes that monitor, evaluate, and regulate management processes themselves. They support self-monitoring, error detection, strategic revision, mental flexibility, and other forms of reflective control.

• Motivator. A value-bearing control state that influences assessment, attention, planning, executive control, learning, and action. Motivators include motives, goals and projects, standards, and attitudes.

• Motivator generator. A computational mechanism that asynchronously generates, modifies, reactivates, and regulates motivators, and assigns insistence and intensity to them.

• Reflective processes. See meta-management processes.

• Repetitive thought. Persistent or recurrent cognition, such as worry, rumination, obsessive thinking, or craving-related thought. In this framework, repetitive thought is explained architecturally as recurrent recruitment of executive resources by highly insistent motivators. This is a subset of mental perturbance.

• Space of possible minds. The space of possible computational architectures and autonomous agents, encompassing biological, artificial, individual, collective, and hybrid forms of mind.

• Working memory. The limited-capacity executive workspace in which information is temporarily maintained, manipulated, integrated, and evaluated during ongoing cognition. In this framework, working memory is closely tied to executive functions rather than treated as an isolated memory store.

Colophon

The images in this document were generated by ChatGPT. Some of the text was generated through interaction with ChatGPT.

Hookmark now wants to read your mind. Not literally, of course. But it does want to infer what information you are likely to want next so that it can present you a link to it.

Some background

Every day, as we engage in knowledge work, we move among many different kinds of information resources: emails, documents, PDFs, web pages, notes, task lists, calendar events, issue tracker entries, and more. While working on any one of these resources, we often realize that another related resource would be helpful. The challenge is not that we cannot find it eventually. Modern search engines are remarkably good. Rather, the challenge is remembering what to look for, where it is, and interrupting our train of thought long enough to retrieve it. Search itself becomes part of the cognitive burden.

In my book Cognitive Productivity: Using Knowledge to Become Profoundly Effective, I called this challenge the meta-access problem: the problem of accessing information that is relevant to one’s current task. It is a higher-order access problem. The question is not simply “Where is this document?” but “Given what I am trying to accomplish right now, what other information is likely to help me?”

This is the problem that Hookmark was created to solve.

From the beginning, Hookmark has been based on the idea of contextual information retrieval (CIR). Rather than expecting users to remember where everything is or to formulate search queries continually, Hookmark helps users navigate directly among related resources. Unlike bucket applications such as Notion, DEVONthink or Obsidian, Hookmark does not ask users to move their information into a single repository. Instead, it acts as connective tissue among the many applications people already use. A manuscript can be linked to the emails discussing it, to the papers it cites, to figures, outlines, notes (in the user’s preferred note-taking app), calendar events, task lists, and many other kinds of information. Hookmark’s links are robust: they continue to work even when files are moved or renamed, and they can be shared with collaborators who have access to the underlying resources (e.g., file or email).

Present and future

Until recently, however, these relationships depended largely on users creating them, either explicitly or implicitly during their work. Although this remains a powerful approach, we have increasingly asked ourselves a more interesting question: can software infer useful relationships even when users have never linked the resources together?

That question has become one of the principal directions of Hookmark’s development.

Our goal is increasingly to anticipate what resource you are likely to want next. If you are reading a PDF, perhaps you might want the publisher’s web page. You might well want to access the key papers that cite the current PDF, or those that it cites. If you are editing a manuscript, perhaps you will want the email from your co-author, the issue tracker entry describing a revision, or the task reminding you to update Figure 3. These are not arbitrary recommendations. They arise from the context of your current work.

Some of these recommendations require no AI whatsoever. They can be derived from explicit links that you have created, from document metadata, from persistent identifiers such as DOIs, from citation networks, URLs, application context, and other structured information. Some of them are based on log data to which Hookmark has access. Hookmark already exploits many of these sources. For example, when you are reading an academic paper, Hookmark can identify its DOI and immediately offer links to the paper on the publisher’s web site. Similarly, Hookmark 7.3 can read the PDF’s meta-data such as its Where from: field. Hookmark 7.3 will also recommend papers that cite the current paper and papers that are cited by it. No manual linking is required.

These recommendations appear in the RELATED section of Hookmark’s Context window, which can be opened with a single keystroke (⌃H) from virtually any link-friendly application. (You can show and hide this section using the ⌥⌘R keyboard shortcut.) Populating this section has become one of our major areas of ongoing development, and each release expands the kinds of relationships that Hookmark can recognize automatically.

In this sense, Hookmark is trying to “read your mind.” More precisely, it is trying to infer your current information goals. Rather than asking simply, “What documents resemble this one?”, Hookmark increasingly asks, “Given what the user is doing right now, what information is most likely to help them accomplish their current goal?” I believe this is a fundamentally different problem from traditional search and from most web-based recommendation systems.

Traditional web recommendation systems are largely designed to maximize engagement. Their objective is often to keep users consuming content for as long as possible so that they can serve you more ads. Hookmark’s recommendations have a very different purpose. They are intended to reduce cognitive effort, preserve flow, and help users complete meaningful work. Their success should be measured not by how long users remain inside Hookmark, but by how quickly they reach the information that advances their task. And unlike web recommendation systems, Hookmark can serve not only web links but deep app links as well.

Looking ahead, we also plan to incorporate optional LLM-based capabilities into Hookmark’s recommendation engine using local AI. (Analogous to Cotypist’s app’s usage of local AI.) We expect these techniques to complement—not replace—the deterministic methods on which Hookmark already relies. Explicit links, metadata, persistent identifiers, citation networks and application context will remain important sources of evidence. AI simply gives us additional ways of inferring what information is likely to be relevant in a given context.

We also intend soon to open up this information to the AppleScript API, macOS Shortcuts, and the app’s intents API.

Significance

I believe this represents an important shift in personal knowledge software. For decades we have focused on storing information and searching for it. Increasingly, our software should help us anticipate what information we will need before we think to ask for it. That is, in my view, one of the next frontiers of contextual information retrieval.

If you would like to learn more, see our continually updated documentation on the RELATED section of Hookmark’s Context window.

The recommendation features are available today in Hookmark Lite and in the Hookmark Trial. I invite you to download Hookmark and let us know what recommendations you would like to see it make in future releases.

I look forward to reading your comments below ↓.

I designed Hookmark primarily for academics. It happens that solving the meta-access problem for academics also solves it for everyone else.

CogSci Apps Corp. is currently making a push to ensure that as many academics as possible use Hookmark. So I wrote the following web page: Hookmark for Academics. The following diagram sums it up:

If you are an academic or anyone who works with many knowledge resources on a Mac, I invite you to consider using Hookmark.

Search has largely solved the problem of finding information. AI has exposed a different problem: how to retrieve the information that matters because of what you are currently working on.

Every research project now accumulates an expanding constellation of PDFs, AI conversations, notes, email threads, meeting notes, transcripts, datasets, web pages, code, diagrams, drafts, tasks, prompts, and AI-generated reports. The challenge is no longer merely finding one of these resources. It is retrieving the information that is relevant because of the current context, which is typically defined by the foreground resource: the paper, draft, email, task, AI chat, or other item that currently has your attention.

I call this contextual information retrieval. More than a decade ago, in my book Cognitive Productivity: Using Knowledge to Become Profoundly Effective, I introduced the closely related concept of the meta-access problem: the problem of efficiently accessing information because of its relationship to the information you are currently viewing. Generative AI has now made that problem much more important.

From information access to contextual information retrieval

Suppose you were reading an important research paper yesterday and now need to get back to it. Traditional information retrieval asks, “How do I find this paper?” That is what Google Scholar, Spotlight, email search, file search, and reference-manager search are good at.

Contextual information retrieval begins after you have found the paper. It asks, “Now that I am looking at this paper again, what else should be immediately available because it belongs with this paper?” That might include notes (in the app of your choice), ChatGPT conversations, Claude Projects, Gemini Deep Research reports, NotebookLM notebooks, Perplexity searches, DEVONthink records, Bookends or Zotero references, OmniOutliner outlines, meeting notes, transcripts, datasets, code, email discussions, grant proposals, entries in your issue tracking system, tasks, calendar events, manuscripts, diagrams, glossaries, and related papers.

These resources are not necessarily connected by keywords alone. They are connected because they participated in the same line of thought, research process, design problem, collaboration, or project. That is why contextual information retrieval is different from ordinary search.

Three complementary kinds of retrieval

Modern knowledge work depends on at least three complementary kinds of retrieval.

Information retrieval answers the question, “Where is this document?” Typical tools include Google, Spotlight, email search, Finder search, and reference-manager search.

Semantic retrieval answers the question, “What documents discuss this topic?” Typical tools include AI search, vector search, semantic search, and embedding-based search.

Contextual information retrieval answers the question, “Given what I am working on right now, what other resources belong with it?” One important mechanism for this is deep relationship retrieval: retrieving resources because of their deep relationships to the current foreground resource.

These three forms of retrieval complement one another. None replaces the others. Search finds documents. Semantic retrieval finds conceptually related materials. Deep relationship retrieval finds the web of resources surrounding the thing you are currently working on.

Where RAG fits

One of the most important ideas in modern AI is retrieval-augmented generation, or RAG. A RAG system retrieves relevant information before generating an answer. That is enormously valuable, but it solves a different problem.

RAG asks, “What information should an AI retrieve before answering this question?” Contextual information retrieval asks, “What information should I be able to retrieve because of what I am currently working on?” Put differently, RAG retrieves information for an AI, whereas contextual information retrieval retrieves information for a human.

These are complementary capabilities. As AI becomes more powerful, both become more important. AI systems need better retrieval to generate better answers. Humans need better contextual retrieval to manage the growing information ecosystems that AI helps create.

Every foreground resource has an information ecosystem

Every foreground resource — a paper, proposal, email, software issue, AI conversation, legal document, presentation, or draft — has an information ecosystem. That ecosystem includes everything that explains it, challenges it, extends it, implements it, summarizes it, cites it, depends on it, or resulted from it.

For example, a manuscript may be connected to AI chats, meeting notes, transcripts, reviewer emails, Bookends references, Zotero collections, DEVONthink records, OmniOutliner outlines, diagrams, glossaries, datasets, code, tasks, calendar meetings, source PDFs, web pages, and related drafts. The challenge is navigating this information ecosystem without repeatedly resorting to search.

That is the meta-access problem in its contemporary form.

Deep relationships, not merely similar text

It is tempting to think that contextual information retrieval is just about finding resources that are textually or semantically similar to the current foreground resource. But that misses the deeper point. What really matters is that these resources are connected by a common episode of knowledge work. A tool may later make those relationships explicit and navigable, but the relationships themselves often already exist in the work.

Two documents that happen to contain the same words have a shallow lexical relationship. Two resources that participated in the same cognitive, scholarly, design, or work process have a deep relationship. A PDF, a ChatGPT conversation, a grant proposal, a dataset, an issue tracker entry, a meeting transcript, and a draft manuscript may share few keywords. Yet they may belong together because they all contributed to the same line of reasoning.

Contextual information retrieval is therefore not primarily about text similarity. It is about preserving and navigating the deep relationships that constitute the structure of your thinking.

Hookmark and deep relationship retrieval

Hookmark was designed for this problem. Rather than asking you to move everything into yet another repository, Hookmark lets you create durable links among the resources that already exist in the applications you already use.

Those relationships can span PDFs, Finder files, folders, notes, AI chats, emails, tasks, calendar events, DEVONthink records, Bookends references, Zotero items, OmniOutliner outlines, web pages, source code, and many other resources. When you invoke Hookmark for the current foreground resource, it shows the resources connected to it. In doing so, Hookmark makes deep relationships explicit and navigable.

This is especially valuable in AI-assisted work. If a ChatGPT conversation helped you understand a paper, hook it to the paper. If a Claude Project helped shape a proposal, hook it to the proposal. If NotebookLM generated useful questions about a source, hook them to the source material. If Perplexity uncovered an important paper, connect it to the manuscript that cites it. The objective is not merely to preserve links. It is to preserve the structure of your thinking.

You are already familiar with links on web pages. Hookmark extends that idea beyond the browser. It can connect resources across many link-friendly applications, meaning applications that expose stable links to their resources. Hookmark works not only with standard links (like https:// links) but with omni-links, deep links that make resources addressable across applications.

Beyond search

Search transformed knowledge work by making individual resources easy to find. Semantic retrieval made it possible to search by meaning. RAG is transforming AI by giving language models access to external knowledge. The next frontier is helping people navigate the growing web of relationships among the resources they create.

Every interaction with ChatGPT, Claude, Gemini, NotebookLM, Copilot, or Perplexity can produce another potentially valuable knowledge artifact. The bottleneck is no longer generating information. It is maintaining contextual access to the information ecosystem surrounding whatever you are currently working on.

That is why contextual information retrieval matters. It is also why the meta-access problem has become more important than when I first described it in Cognitive Productivity. Hookmark was built to solve it.

Executive summary

This article introduced the following concepts.

• Contextual information retrieval — retrieving information based on your current working context, typically the current foreground resource.

• Meta-access problem — the problem of efficiently accessing information because of its relationship to the information currently in the foreground.

• Deep relationships — relationships among information resources that arise from a common cognitive, scholarly, design, or work process, rather than merely from lexical overlap, semantic similarity, or temporal proximity.

• Deep relationship retrieval — a principal mechanism for contextual information retrieval, in which resources are retrieved because they have deep relationships to the current foreground resource.

• Foreground resource — the document, note, AI conversation, email, task, web page, manuscript, or other resource that currently has your attention.

• Information ecosystem — the network of resources surrounding a foreground resource, including AI chats, notes, emails, tasks, datasets, drafts, diagrams, glossaries, code, and related documents.

• Information retrieval — finding a known resource, typically through keyword or file search.

• Semantic retrieval — finding resources because they are conceptually related to a query.

• Retrieval-augmented generation (RAG) — an AI technique in which retrieved information is supplied to a language model before it generates a response.

• Knowledge artifacts — the intellectual products of knowledge work, including AI conversations, annotations, drafts, reports, diagrams, code, notes, datasets, and summaries.

• Link-friendly applications — applications that expose stable deep links so their resources can participate in a larger information ecosystem.

• Omni-links — app-specific or standard deep links that uniquely identify resources across applications, enabling tools such as Hookmark to connect them into a navigable knowledge graph.

• Hookmark — a contextual information retrieval system that uses deep relationship retrieval to make the information ecosystem surrounding a foreground resource immediately accessible.