ChatGPT answered, but not only about what you’re looking forward to (it’s reading, btw). It also suggested you make that request more professional, and prefaced “Dear Anna” instead of “Hi Anna”, which sounds too informal—ChatGPT said. You complied, because it sounded about right, and you had no time anyways, and these were your words anyways, so why make such a big fuss out of it?

Well…

You know it intuitively already. ChatGPT didn’t just fix your grammar, but also made a call about how formal you should sound to a colleague.

AI as a standardization mechanism

Linguists see here an instance of standard language ideology in the making:

“a bias toward an abstracted, idealized, homogeneous spoken language which is imposed from above, and which takes as its model the written language. The most salient feature is the goal of suppression of variation of all kinds” (Lippi-Green 1994: 166)

The sociolinguist Lippi-Green here mentions the spoken language, but the written one is precisely where it starts: where standardization is most likely to creep in, through orthographic norms, punctuation, and the idea that ‘the’ formal way of speaking is also the right one.

Without knowing either you or your colleague, and with no clue you had just shared a vulnerable moment over lunch. Funnily enough, for ChatGPT personalities, ‘professional’ is not ‘friendly’. As if both were opposite modes.

Little by little, through seemingly unimportant encounters, AI offers more rephrasing than you initially asked for. So you start checking yourself against it the way you might once have checked yourself against a native speaker, because AI has finally made a ‘professional’ editor available to everyone.

Why native speakerism is harmful

None of this is about people wanting to get it right. I don’t think people, especially those of us writing across multiple languages, will stop using AI after they (we) have been told for DECADES to have their texts ‘proofread’ by ‘native speakers’—which has always been an extremely costly, vastly unavailable service.

Making sure you send over an email, a report, or an essay with no obvious grammatical mistake is a reasonable thing to want, and, I’d even say, a very human one, especially when we know how the world assesses our prose before they’ve even read it.

Mandy is a Chinese graduate student at a U.S. university. Across four interviews, she told the researcher Kewen Zheng (2025: 5) that she used to write her texts in English, and then—only then—search online to see how often certain phrases were used. Then she started asking ChatGPT directly. Then she had it proofread her assignments, and then her emails:

“I feel like every time I use English, I need its permission.”
(Zheng 2025, Excerpt 1, bold in the original)

In the same interview she said: “ChatGPT makes me increasingly believe in native speakerism”.

Of course, it’d be easy to say that multilingual writers (and perhaps everyone, really) have always had their texts checked: by a colleague, a partner, an editor if they could afford it. So, what’s the big deal?

First, although phrased as suggestions, many (most?!) ‘corrections’ by ChatGPT do not target actual grammar deviations. More often than not, the AI will give you input on expressions that are perfectly fine, and that it will consider unidiomatic—not because they would be, but because it is trained to keep you interacting, and responding to your question only arguably fails at keeping you engaged (aka anxious).

Figures of authority have always loved to give you more feedback than necessary. Martina Oldani and I showed it in a paper on linguistic racism in German schools (2022):

Kevser, a German-Turkish student we interviewed, started out talking about her German rather positively. Then a literature teacher made a comment about her German. After that, despite her excellent grades otherwise, she was constantly afraid of making mistakes, so she stopped taking part in class. One remark from someone with authority was enough to make her German the problem.

With AI, that kind of remark is available every single time you ask it to check a sentence—including for every other sentence it’ll check without you asking.

This is why Zheng (2025: 5) suggests that ChatGPT now acts as “an authoritative teacher by actively identifying ‘errors’ or suggest[ing] ‘better’ expressions”.

Teachers, too, correct toward a model: the native speaker, a figure sociolinguists have long criticized as harmful and ideological:

So-called ‘native speakers’ also hesitate, vary, make mistakes.

The second, more pernicious problem is that ChatGPT suddenly makes that idealized figure available on demand—what Zheng (2025: 8) calls the “omnipresent native speaker”.

While the teacher was relegated to the classroom and thus to specific writing contexts (or genres), AI can now correct your style on any occasion, from a Substack note to the birthday invite you’ll share in a WhatsApp group.

In a controlled experiment with 118 participants from India and the U.S., AI writing suggestions led Indian participants to adopt Western writing styles “Indians showed a higher engagement with AI suggestions than Americans” (Agarwal, Naaman & Vashistha 2025: §4.1.1).

Sounds odd(ly) familiar? Mmh, I did tell you 3 paragraphs ago that Indian English is considered, across the world, a “nonnative” variety of English. Bingo.

From linguistic insecurity to overcorrection

Zheng’s Chinese international students in a U.S. graduate program explained that the more they used ChatGPT, the more they felt they needed it for any type of communication:

Nora, for example, reported that her dependence on ChatGPT grew to the point where she used it not only for academic submissions, but also for casual communication with friends. (Zheng 2025: 6)

The same students also knew their professors might police AI use. Mandy even recalls being rejected from a summer internship program after positive emails with a professor, on the basis that they were concerned about her use of AI.

So some of them made their writing look “less perfect and more ‘non-native’” to avoid suspicion. Zheng (2025: 8) speaks of an “ironic struggle”:

I’ve also written on linguistic privilege and who gets to be messy or polished in the age of AI here ⤵️

Another study of AI editing in undergraduate writing points the same way. When the tool edits without explaining its changes, it makes more of them, including changes to content and meaning, and undergraduate multilingual writers are the least equipped to evaluate what it hands back (Groener & Zanina 2026). But, most importantly, the authors warn against the cumulative effects on early-career multilingual writers who may start doubting their ability to evaluate AI outputs:

A strong wish for efficiency and believing in superiority of AI-technique seems to correspond with a lack of trust in multilingual writers’ language skills and supports a way of usage that outsources AI-editing to the tool using it in its dominator function and reducing possible learning effects and human control. (Groener & Zanina 2026: 252)

And that’s how overcorrection creeps in. When you only ever see the correction and never the ‘rule’ behind it, the AI chatbot becomes the first thing you reach for to ‘sound right’. If you happen to ask it often enough, you’ll start integrating the idea that you are reliant on AI (and construct yourself as a “learner”, as Zheng (2025: 5) argues).

And if memory is on, the chatbot can store what it has learned about you (‘not a native speaker’ included) and draw on it in later answers (OpenAI, Memory FAQ).

The feedback loop is on 🌀🤖

And what about readers?

Before ChatGPT, “I look forward to read your answer” would have been understood just fine.

You probably wouldn’t have even noticed. You’d have filled in the missing -ing, or simply not cared—because you would focus on content. An argument I surprisingly hear all the time when it comes to AI-assisted writing, and, strangely enough (or not), not so much for (poorly written) human-authored texts. (Linguist in the Wild unpacks this here.)

Reading beyond the small errors here and there is a reader’s skill, one that we build by meeting enough varieties of Englishes that a dropped ending or a borrowed structure does not stop us in our tracks.

Almost all the research I’ve cited so far looks at writers. But what about readers?

For speech, Rosina Lippi-Green (yes, her again, what can I say, I have my favorites) called it the communicative burden: understanding takes work on both sides, and yet members of the dominant group often “feel perfectly empowered to reject their responsibility, and to demand that a person with an accent carry the majority of the burden in the communicative act” (Lippi-Green 2012: 72).1

Building a research field around the notion of raciolinguistics, Flores and Rosa (2015) show that the white listening subject is an ideological position that hears some speakers as deficient, whatever they actually say.2

More generally, we know that exposure helps us accept ways of speaking as less ‘special’. People tend to believe statements less when they’re said with a foreign accent (Lev-Ari & Keysar 2010). A short exposure to that accent reduces the bias, because the accent gets easier to process (Boduch-Grabka & Lev-Ari 2021).3 Listening to unfamiliar voices is a skill. And skills need practice.

Now carry that over to writing. If more and more of what we read has been polished through AI tools oriented toward standard ways of speaking (Schneider 2024), we get less practice with text that is patterned differently.

I haven’t found a study on what this does to readers (if you know one, send it my way!). But, tentatively, and based on what we know otherwise: Less exposure to very specific, individual writing styles means less practice, and less practice means less patience with the next text that hasn’t been polished. The less often you meet “look forward to read”, the more it looks like a little stone you’ll stumble over (or onto, perhaps).

If you’re working on these questions, you may want to have a look at the international workshop on “Doing Multilingualism with Generative AI: Recentering the Users” I’m co-organizing in March 2027 in Zurich.

In a nutshell:

What to do, and how to reclaim linguistic agency

I’ve tried to use AI to check my English, as do most multilingual writers I know, originally with ‘small’ questions. I slowly started to doubt whether I can write in English at all—after a 300-page book and 30+ academic papers, another academic blog before this one, and perhaps 75+ talks. Yep. I’m not sharing my CV to impress you, but because if I feel compelled to check every formulation as a sociolinguist whose job it literally is to study this, you can imagine how other writers may feel.

It started after a note on Substack where I wrote uneasiness (instead of unease), somehow decided uneasiness is not proper English, asked ChatGPT, which confirmed (!!!), and ended up in a rabbit hole.

In a perfect world (sorry not sorry), my first move would be to say: use AI as little as you can for editing. But I know it’s not always realistic, and I also feel VERY reluctant to put the responsibility on the most vulnerable among us, again (even indirectly).4

My concern is the cumulative effect.

A few things that I genuinely5 think can help:

It won’t come as a surprise that I am very critical of what AI does to us. I also know that many people, including multilingual writers, have felt empowered, creative, enabled, and I take that experience seriously. As I hope to have shown today, the harmful consequences of AI use and exposure, which are clear for the Chinese students in Zheng’s paper, do not come down to individual writers or readers. What I do hope to have shed light on is what matters: reclaiming our ways of speaking and writing as ours, and practicing inclusive listening and reading.

Thanks for reading me in that spirit. Take care,
—Naomi

As some of you know too well, I often write in cafes. This newsletter is happily free. If it makes you think, you can sponsor my coffee addiction

buy me a coffee

And for more on whether AI actually is multilingual, it’s over there ⤵️

References

Agarwal, Dhruv, Mor Naaman & Aditya Vashistha. 2025. AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25), 1–21. New York, NY, USA: Association for Computing Machinery. https://doi.org/10.1145/3706598.3713564.

Boduch‐Grabka, Katarzyna & Shiri Lev‐Ari. 2021. Exposing Individuals to Foreign Accent Increases their Trust in What Nonnative Speakers Say. Cognitive Science 45(11). https://doi.org/10/gng5b2.

Flores, Nelson & Jonathan Rosa. 2015. Undoing Appropriateness: Raciolinguistic Ideologies and Language Diversity in Education. Harvard Educational Review 85(2). 149–171. https://doi.org/10.17763/0017-8055.85.2.149.

Groener, Carina Ulrika & Anna Zanina. 2026. Unlocking Multilingualism: Thoughts Shaped by Tools? In Adela Coman, Simona Vasilache & Abbas Moallem (eds.), HCI International 2025 – Late Breaking Papers, 236–254. Cham: Springer. https://doi.org/10.1007/978-3-032-12801-0_16.

Leivada, Evelina, Itxaso Rodríguez-Ordóñez, M. Carmen Parafita Couto & Sílvia Perpiñán. 2023. Bilingualism with minority languages: Why searching for unicorn language users does not move us forward. Applied Psycholinguistics 1–16. https://doi.org/10.1017/S0142716423000036.

Lev-Ari, Shiri & Boaz Keysar. 2010. Why don’t we believe non-native speakers? The influence of accent on credibility. Journal of Experimental Social Psychology 46(6). 1093–1096. https://doi.org/10.1016/j.jesp.2010.05.025.

Lippi-Green, Rosina. 1994. Accent, Standard Language Ideology, and Discriminatory Pretext in the Courts. Language in Society 23(2). 163–198.

Lippi-Green, Rosina. 2012. English with an Accent: Language, Ideology, and Discrimination in the United States. 2nd edition. London/New York: Routledge.

Oldani, Martina & Naomi Truan. 2022. Navigating the German school system when being perceived as a student ‘with migration background’: Students’ perspectives on linguistic racism. Linguistics and Education 71. 1–15. https://doi.org/10.1016/j.linged.2022.101049.

Schneider, Britta. 2024. A Sociolinguist’s Look at the “Language” in Large Language Models. Critical AI 2(1). https://doi.org/10.1215/2834703X-11205168.

Truan, Naomi. 2024. Whose language counts? Native speakerism and monolingual bias in language ideological research: Challenges and directions for further research. European Journal of Applied Linguistics 12(1). 34–53. https://doi.org/10.1515/eujal-2024-0006.

Zheng, Kewen. 2025. You don’t need to prove yourself: A raciolinguistic perspective on Chinese international students’ academic language anxiety and ChatGPT use. Linguistics and Education 86. 101406. https://doi.org/10.1016/j.linged.2025.101406.

I guess this sounds a bit silly. Are we even supposed to call them anything? Well, you do you, but ongoing research on anthropomorphization with voice assistants shows that it’s quite common (Abercrombie et al. 2021, Lind 2025).

So, Ms Robertson or Laura? I’d bet you’d rather say Laura, or even bestie, and that this can also change if the interaction gets a bit intense—from hey, redo that one buddy to wtf is wrong with you mate???

That’s in English. In many other languages, being on a first name basis is encoded grammatically: tu vs vous en français s’il vous plaît, or du vs Sie in German. Importantly, and in one paragraph to summarize 75 years of research, politeness is NOT the right interpretation here.

What’s at stake is appropriateness—what a specific situation, with a specific interlocutor, leads us to expect. Then, you add gender, age, hierarchies, familiarity, cultural frames etc. into the mix and you immediately get why addressing someone is a very sensitive choice.

And companies are very aware of this.

In June 2020, the Twitter account of the German railway company, Deutsche Bahn, announced publicly it would spend three months addressing its customers as du instead of Sie: Robert, not Herr Liebert anymore. People started debating, and Naomi the linguist started analyzing (yep, I’m talking about myself in the third person, won’t do that too often, I know it’s weird, see my book). People got upset. Like really upset. Philip (a real user from my data) was appalled! Wir sind nicht per Du = “We are not on a first-name basis”.

Ok boomer.

That one’s not on me: this is what many other Twitter users replied. Philip’s reaction was not isolated, and also not incoherent. To Philip, Deutsche Bahn is a company (whose trains are often delayed), and he is a customer; there is no reason to suddenly imply this level of familiarity. That felt off.

Except Deutsche Bahn was very much not announcing they would be saying du to everyone on trains, or even at the station.

This was a Twitter thing.

Philip was, in a way, mixing up online and offline worlds. Twitter, like any other social media platform, had its own way of talking.1 It’s not that Deutsche Bahn had just hired a new marketing team trying to sound cool. I’d rather argue they were adjusting to the perceived social norms in that specific, bounded communicative setting. Bounded. On Twitter, I said.

Look at these two posters. The one on the left says Wir verstauen Ihr Gepäck fachgerecht = “We will store your luggage properly”, the formal form of address.

Just next to it—hence the pic, although I didn’t even have a newsletter back then, but I KNEW I would reuse this—Wir sind nicht ganz sauber? Schreibt uns ‘ne WhatsApp switches into an informal, conversational style: “We’re not quite clean? Send us a WhatsApp”, with the imperative form (send) encoding an informal plural form2.

And then there is Kiana, Deutsche Bahn’s AI chatbot. Its description on bahn.de says: Der KI-gestützte Chatbot Kiana auf bahn.de unterstützt Sie bei Fragen rund um die Deutsche Bahn = “The AI-powered chatbot Kiana on bahn.de supports you with questions about Deutsche Bahn.” Sie. The formal you.

One way to interpret this is to say that Deutsche Bahn is not very consistent, and they should stop asking their intern to manage their external communication. After all, IKEA has decided once and for all, and the Swedish coolness of saying du has become part of their corporate identity. (That’s not true, by the way. This may hold for most European marketing strategies, but House & Kádár (2020) show that IKEA catalogues in Mainland Chinese and Japanese avoid the informal form.)

But there is another, sociolinguistically informed way to look at this: different communicative settings can come with different expectations about address forms. Deutsche Bahn can use du on Twitter, a more conversational style in a WhatsApp campaign, and Sie in a service interaction without any of those choices necessarily contradicting the others.

So, no, sorry Philip, Deutsche Bahn had not become an informal (aka cool, trendy, and young) company. It was adjusting its way of speaking to each interaction.

How customer chatbots talk to us

When a human employee begins a conversation with a customer, address forms can be worked out as the interaction unfolds. Sure, the person behind the counter may follow a script, but more often than not, it is both culturally specific and dependent on the interaction. (I will always remember how, when I was 18 or 20, a kind woman called me love at a British station and I literally froze because I had no clue what was happening.)

Anyway, in Berlin, where I lived for over a decade, I was only using du, the informal pronoun, no matter where: a cafe, on the bus, to ask what time it was to a stranger on the street, and at the Späti. Apart from certain work relationships, the only place where I maybe wouldn’t use du was the Bürgeramt, the citizens’ registration office, but I’m not even sure, honestly.

A scripted chatbot cannot do that unless somebody has designed it to. (I am NOT saying this is technically impossible, please keep reading.)

Someone has to decide what the bot will say before the customer arrives. In German service chatbots, the answer until now has very often been Sie. The bot’s responses are written around that choice, and the choice becomes part of the interaction’s default setting.

That is not necessarily a problem. If the customer expects the formal Sie, everything works perfectly. But addressing becomes interesting precisely when the participant does something different. Imagine a customer writing informally, and the bot keeps using what it has been programmed to say.

Even if the response is grammatically correct, the bot has not followed the customer’s address choice—the one initiated in the first turn by a human. Note that a human employee might have several reasons for doing the same thing: they didn’t even notice, they’re following a script, they want to keep that distance…

We could stop there, and this would be another essay in the series Sociolinguists should find themselves a hobby, BUT! This tiny choice can affect how people experience the interaction, with implications for how warm and friendly a company appears and, in the end, whether people like it and buy from it.

Because people notice 👀

There is empirical evidence that I’m not the only one who’s spending too much time staring at pronouns.

Ollier, Nißen and von Wangenheim (2022) tested this directly, manipulating du/Sie (and tu/vous) in a real Swiss health-insurance chatbot with 284 German-speaking and French-speaking participants. In German, there was no overall preference for either form, but the informal du was assessed more negatively by older users, especially men. French speakers, on the other hand, still heavily favored the formal vous.

The first reason why this matters is that this has crucial implications for users’ engagement with the bot (Ollier, Nißen and von Wangenheim 2022: 16):

“By conforming to users’ expectations regarding facets of speech-etiquette and language-use more generally, designers of [conversational agents] can shape engaging user experiences. For healthcare [conversational agents] this is particularly vital, as it facilitates greater adherence to the treatment plans they help deliver.”

As the study was specifically set in a plurilingual state, Switzerland, it clearly shows that using the same form of address in German and French would not necessarily produce the same user experience.

The second reason is more subtle: what people say they prefer and what they actually enjoy in an interaction does not always match. Indeed, when asked directly, all German speakers said they preferred the informal du. But the people who rated the chatbot higher when it used du—as they wanted—were young.

That is exactly what makes address different from a design decision preceding the interaction. The social meaning is produced in the encounter.

Den Hartog and colleagues (2025) found the same split when they asked people to evaluate a recruiter directly, this time in France and Germany (no chatbot). For French participants, address changed only how they felt about that one recruiter—not the company, the job, or their own salary expectations; going informal got the person blamed, not the employer. German participants did the opposite: address changed their read on the company, the job, even what salary they’d expect (Hartog et al. 2025: 853).

My provisional read is that once a customer-service chatbot is deployed across countries and languages, picking the right address for the right audience is a moving target—one that companies do not seem to have paid much attention to until now, in my experience.

From Twitter to chatbots and back

We know already that the same company (Deutsche Bahn) uses different pronouns of address depending on the situation. That makes sense if we think about the communicative setting rather than treating du and Sie as universally appropriate or inappropriate.

Following their website’s description, I thought they had decided that it’s Sie. But the study I cited before is based on German stock index companies (DAX Unternehmen). I’m not quite sure the Deutsche Bahn ‘belongs’ here, in terms of positioning strategy.

So I tested it myself, of course. And their chatbot actually uses the informal du3!!!

And the pop-up window that shows up at the exact same moment to ask me to take part in their survey—Nein danke—addresses me per… formal Sie.

At this point, I expect the Deutsche Bahn to reach out to me and justify themselves 😂

But then I tried something else.

I started a new interaction with Kiana and formulated my question formally, using Sie.

Even with several markers of formality, crucially including the formal pronoun of address Sie, Kiana stuck to du. Fine enough, that’s how it was designed.

“LLM-based systems are pragmatically incompetent!”—a tempting conclusion from a linguistic observation like this. But what I actually want to know is why a system that can generate highly flexible conversational language can still fail to respond to such a small, and, I would argue, fairly simple interactional cue.

A plausible, yet uncanny explanation would be that the chatbot adjusts to my profile, as they know my date of birth: I’d say I still count as ‘young’, so I may not be put off by the informal du the same way older German-speaking Swiss men were. But I’m tempted to think this is giving the Deutsche Bahn’s chatbot too much credit.

What ChatGPT has changed

The thing is, since 2022, chatbots are not for confused or annoyed customers anymore (yes, we’re all always confused or annoyed, otherwise why would we even open a chatbot?!), but for everyone eager to have a chat.

A representative survey from 2024 reports that 99% of people use the informal du with their AI in German.

And while talking to ChatGPT is a different communicative event (no company, no customer relationship, and no escalation path to a human employee)…

This changes what sounds normal when people talk to AI.

People have become accustomed to conversational AI that speaks with them in an informal register. LLMs can sustain a conversation rather than simply moving a customer through a decision tree. They can notice what a person has said and adapt their response. They can maintain context.

Will the conventions of talking to AI start travelling into customer service? Or will companies continue to treat the customer-service chatbot as a separate communicative genre, with its address form decided before the customer ever types the first message? We’re back to Philip: does the commercial relationship prevail over platform norms?

And more generally—and most importantly—will we soon have chatbots that pay attention to these little things, or will people stop caring?

If you are designing conversational chatbots, this is also why you need linguists in the room 😇 And if you are designing these systems across languages and cultures, even more so: is address something you had ever considered when deciding how your chatbot should speak to people?

Your homework

(You can see I’m ready for the semester start.)

If you want to help this newsletter—and my thinking—grow, send me your next interaction with a customer-service chatbot (ideally, the whole conversation as a PDF or a set of screenshots rather than just the one sentence that caught your attention.)

Leave a comment

My guess is that it’s already happening: Berlin’s service portal, the very one I mentioned earlier as a situation in which I would exceptionally use the formal Sie, has a chatbot that says… du. And, believe it or not, I hadn’t checked it out while I was writing the paragraph about my Berliner years.

If you enjoyed this essay, you may want to check out this one!

References

Abercrombie, Gavin, Amanda Cercas Curry, Mugdha Pandya & Verena Rieser. 2021. Alexa, Google, Siri: What are Your Pronouns? Gender and Anthropomorphism in the Design and Perception of Conversational Assistants. In Marta R. Costa-jussà, Hila Gonen, Christian Hardmeier & Kellie Webster (eds.), Proceedings of the 3rd Workshop on Gender Bias in Natural Language Processing, 24–33. Online: Association for Computational Linguistics. https://doi.org/10.18653/v1/2021.gebnlp-1.4.

Hartog, Maria den, Patricia Sánchez Carrasco, Gert-Jan Schoenmakers, Lotte Hogeweg & Helen de Hoop. 2025. Processing Pronouns of Address in a Job Interview in French and German. Applied Linguistics 46(6). 841–856. https://doi.org/10.1093/applin/amae050.

House, Juliane & Dániel Z. Kádár. 2020. T/V pronouns in global communication practices: The case of IKEA catalogues across linguacultures. Journal of Pragmatics 161. 1–15. https://doi.org/10/ggtmcd.

Lind, Miriam. 2025. Alexa’s agency: a corpus-based study on the linguistic attribution of humanlikeness to voice user interfaces. AI & SOCIETY 40(6). 4619–4633. https://doi.org/10.1007/s00146-025-02243-8.

Ollier, Joseph, Marcia Nißen & Florian von Wangenheim. 2022. The Terms of “You(s)”: How the Term of Address Used by Conversational Agents Influences User Evaluations in French and German Linguaculture. Frontiers in Public Health. Frontiers 9. https://doi.org/10.3389/fpubh.2021.691595.

Truan, Naomi. 2022. (When) Can I say Du to You? The metapragmatics of forms of address on German-Speaking Twitter. Journal of Pragmatics 191. 227–239. https://doi.org/10/gpjxqv.

Truan, Naomi. forthcoming. How and why choosing the wrong form of address may make you look like a boomer or a Karen: Characterological figures on social media. In Heinz Kretzenbacher, Horst J. Simon & Christa Dürscheid (eds.), “You can say you to me” Address in German (Topics in Address Research 7). Amsterdam: John Benjamins. https://benjamins.com/catalog/tar.7.06tru.

“Hey, are you a human?”, asked no one ever (in real life, that is).

Not because you would always be sure—tax forms don’t feel like they’ve been processed by a human tbh—but, simply, because it was not a question until a few months ago.

We recognize people as human without thinking about it. But sociolinguists spend a lot of time studying exactly this kind of seemingly obvious recognition.

We hear an accent and infer where someone is from. We hear a register and infer education, class, profession, age, stance. A particular way of speaking, and, hop, we decide (not necessarily as in decide-decide, but unconsciously) that someone is a lawyer, a teenager, a local, an outsider.

None of these social categories is contained in the linguistic form itself. In fact, “hearing with an accent”—and not speaking with one—is a concept based on the idea that the “perceived success of interactions”, in Lindemann’s terms (2002: 421), depends on the attitude listeners have toward speakers.

In other words, you will understand someone less if you’ve decided, beforehand, that they have an accent.

And that is not a figure of speech. Kutlu, Tiv, Wulff & Titone (2022) played listeners recordings of American, British and Indian English, each paired with a photograph of either a white or a South Asian woman, and asked them to write down what they heard. In Gainesville, Florida, listeners transcribed less accurately whenever the face was South Asian—including when the voice was American English. The same experiment, run in Montreal, produced no effect of the face on intelligibility at all1.

Hearing is not a neutral act. The bias is not in what we hear, but in what a particular community has already learned to expect: “the social meaning of observable race has different values” in each community (Kutlu, Tiv, Wulff & Titone 2022: 10).

When AI enters the equation

Now AI has created a strange new situation in which human itself—not age, not origin, not gender—is becoming something we try to recognize from language.

“This doesn’t sound human.”
“You can tell a person wrote this.”
“This feels authentic.”
“It sounds too polished.”

But, surely, even now that the machines have entered the playfield, we do know how to recognize humanness, right? (Spoiler: no, we do not. See here for “an overall accuracy of 19%” for human scoring, or here for German: “only half of the AI-generated texts could be recognized as such”. Instructors are better however: 70% ‘success rate’ according to this study.)

If you’ve been following along, you already know I don’t really care whether we can spot AI-writing well. The most interesting question here is not whether these judgments are accurate.

It is: what makes something legible as human in the first place?

The problem with ‘human’ messiness

The current reaction to AI writing follows a familiar script: AI writing is too polished, too symmetrical, too grammatically perfect, too predictable, too generic.

So we are told to make it more human. Leave the typo in! Make the sentences less perfect. Be spontaneous. Be messy, even: “I’d much rather read your messy real thoughts with grammar imperfections and weird structure than a polished amalgamation of everyone else’s thoughts.”

How nice.

I find this advice increasingly problematic—a weird form of reversed prestige that only people ALREADY holding prestige through linguistic capital can claim.

Of course, human writing is messy. Well, first drafts, at least. But now, suddenly, we’re celebrating people pressing ‘send’ without catching the weird typo? Mmh…

Sorry, but I don’t buy it. The messiness being celebrated is, somehow, remarkably specific. The “grammar imperfections” they want to see are, at most, funny little prepositions that do not affect the text.

What happens when the syntax is unusual in a way you can’t immediately parse? When someone repeats themselves, loses the thread, uses punctuation in an unfamiliar way (let’s not even mention the em dash), or writes in a register that doesn’t fit your assumptions about what competent writing should look like?

That might not feel human at all.
It might feel badly written. It might feel uneducated.
It might even feel like… AI??? 😱😱😱

People do actually NOT want your “human messy drafts”. They want legible messiness: readable evidence of imperfection.

Messiness, but within limits (hey, let’s not exaggerate!).
Spontaneity, but still coherent.
Idiosyncrasy (fancy word to say ‘your own specific style’), but still recognizable.
Imperfection, but aesthetically pleasing.

In other words: we want humanity, as long as we can still recognize it.

Legibility is not the same as intelligibility

Something can be perfectly intelligible, and still be socially difficult to interpret.

A grammatical sentence can still sound ‘wrong’. If you’re from Pittsburgh, “the car needs washed”2 is unremarkable—a perfectly regular construction of that variety, instantly parsed by everyone around you. Written down for a wider audience, it stops reading as regional and starts reading as a mistake.

What matters is not only whether we can understand a linguistic form, but whether we know what to make of it.

That is what I mean by legibility here. Legibility is not an inherent property of language.

Which brings me to enregisterment, a term coined by Agha (2003): how particular linguistic forms become socially recognizable as belonging to particular kinds of speakers, situations, or identities, and what happens when people start noticing.

Take that Pittsburgh example again. Johnstone, Andrus & Danielson (2006) traced how a set of local features—“needs washed”, “yinz”, and others—went from being what people simply said, to being heard as marking a place, to being printed on t-shirts and sold to tourists as “Pittsburghese”. The features acquired a social meaning through repeated association, and once they had it, people could deploy them on purpose.

An accent doesn’t inherently mean ‘foreign’—but it may, in Florida, if you hear it while thinking the person is South Asian. Even if these repeated social associations are inaccurate, biased, and cause harm.

And perhaps something similar is beginning to happen with human-sounding language.

Back to our typos. A clear sign of humanness, right?

Right.

People are learning to manipulate linguistic features to make their humanity legible.

If you think this is a niche thing, this Reddit thread has 2.2K upvotes (to be fair, it’s also two years old, and already predicting the rise of the website screenshotted above).

Enregisterment is here a very useful concept. Earlier I described these two steps:

1. A linguistic feature becomes associated with a social type (“AI”)—our beloved em dash, verbs such as delve into, and a bunch of other stuff AI does, indeed do more often.

2. People become aware of that association.

And now the rest of the steps:

3. People then start modifying their use of the feature because of its social meaning.

4. Other features become desirable because they index the opposite category (‘human’). The typo, the lowercase, tbh, the cool register, all that stuff (see what I’m doing here?).

5. Those features can then be deliberately deployed as stylistic resources. Which is what the Chrome extension sells: humanness, $0, three free emails.

6. And then… the association may, in turn, start to disappear, leaving us in a weird situation in which typos are both human and non-human, AI and non-AI. We’re screwed (nooooooo!).

Step 6 is the most fascinating one in my view. Enregisterment usually describes a feature accumulating social meaning and holding onto it: Pittsburghese is still Pittsburghese on the t-shirt. Here, we’re driving in circles. The moment a typo becomes reliable evidence of humanness, it becomes worth faking; the moment it’s worth faking, it stops being evidence of anything.

So, once people start deliberately producing ‘human’ features, we’re back to the title I originally gave this essay3: We Don’t Want Human Language. We Want Legible Humanity.

The typo doesn’t represent all the messy, unpredictable, idiosyncratic ways humans actually use language. It represents a socially recognizable sign of human authorship—and, for now, an acceptable one as well.

Nonnative speakers already know this

Not only because AI detectors have been consistently shown to flag their texts as ‘AI’ when they’re not (see research here).

This is only the newer version of that bias. Even before AI, we talked about ‘native speakers’ as though nativeness were something we can hear directly. But we can’t—because we’re often “listening with an attitude” (Lindemann 2002).

We recognize people as ‘native speakers’4 through cues we have learned to associate with nativeness: accent, vocabulary, fluency, timing, pragmatic competence, confidence. Those judgments can be wrong, and they are socially patterned. Some people get the benefit of the doubt. Others don’t.

Some speakers can use a nonstandard form and have it interpreted as style. Others use the same—or a similar—form and have it interpreted as evidence that they don’t know how to speak properly.

This is not new and it is not small. When Rachel Jeantel testified for six hours in the Zimmerman trial, her African American English was described as “hard for many in court to follow” (transcript here)—a variety with regular, well-documented grammar heard as incompetence, in a courtroom, where the stakes could not have been higher (Rickford & King 2016).

Which is why the call for messiness—or for keeping the em dash, for that matter—only comes from people who have nothing to lose: already privileged, already seen as eloquent and educated, they can afford to sound a bit messy, a bit distracted, a bit charming.

In that very paragraph you just read, the dash, the repetition of already, the enumeration with three adjectives (messy, distracted, charming) are typical AI tells. But because I have a PhD, have been writing for over a decade, can prove (!) I wrote articles and even a book before November 2022 (when ChatGPT was launched), you somehow believe me when I tell you “it’s only me!”.

Who gets to keep the dash once it’s become sus—and who gets to make jokes about it—is THE central sociolinguistic question here.

Because being able to put a joke like this on a conference slide means you can afford to acknowledge the suspicion and play with it (which is okay, I guess, when you’re a Fields medallist 😉). I would even go as far as arguing that embedding the disclaimer in a footnote is a way to perform ‘being an academic’.

The footnote costs him nothing, and that is precisely why it works. It isn’t free for everyone. For a doctoral student, a second-language writer, anyone whose writing is already being judged for whether it sounds “professional”, “natural”, or “authentic”, the same em dash might feel less like a joke and more like a liability.

It depends on who is speaking, who is listening, and what kinds of linguistic difference the listener has learned to recognize.

This is a language ideology I don’t think we’ve encountered before: that humanity has a recognizable style. This feels almost absurd, because the whole point until now was to index belonging to a specific category: a generation, a region, a social background… But now this becomes reduced to AI vs. human.

And once a style becomes recognizable, it can become a norm.

If ‘human-sounding’ language comes to mean informal, spontaneous, slightly messy, emotionally expressive, conversational, and idiosyncratic, then people whose language does not fit that repertoire may become less readily recognizable as human authors.

AI is changing the linguistic cues through which authorship and humanity are socially inferred.

So, no, we don’t actually want human language. We want legible humanity—at the ironic cost of performed humanness by those digitally competent enough to circumvent AI detection.

Because I can promise you that my mom, who’s never used AI and still reads every article of mine religiously, has absolutely no clue how to integrate that browser extension to add random typos to her texts (which are, very often, in fact, completely free of any typo).

And that, to me, is a much more interesting problem than whether the em dash really is a sign of AI.

Leave a comment

References

Agha, Asif. 2003. The social life of cultural value. Language & Communication (Words and Beyond: Linguistic and Semiotic Studies of Sociocultural Order) 23(3). 231–273. https://doi.org/10/cw8z5c.

Kutlu, Ethan, Mehrgol Tiv, Stefanie Wulff & Debra Titone. 2022. Does race impact speech perception? An account of accented speech in two different multilingual locales. Cognitive Research: Principles and Implications. 7(7). https://doi.org/10.1186/s41235-022-00354-0.

Johnstone, Barbara, Jennifer Andrus & Andrew E. Danielson. 2006. Mobility, Indexicality, and the Enregisterment of “Pittsburghese.” Journal of English Linguistics. 34(2). 77–104. https://doi.org/10.1177/0075424206290692.

Lindemann, Stephanie. 2002. Listening with an attitude: A model of native-speaker comprehension of non-native speakers in the United States. Language in Society 31(3). 419–441. https://doi.org/10.1017/S0047404502020286.

Rickford, John R. & Sharese King. 2016. Language and linguistics on trial: Hearing Rachel Jeantel (and other vernacular speakers) in the courtroom and beyond. Language 92(4). 948–988. https://doi.org/10.1353/lan.2016.0078.

There is a rectangle at the bottom of your screen and you are supposed to type into it.

It looks like a search box. But maybe also like a login form, a credit card field, or the thing you use to message your sister. The designer Amelia Wattenberger made this point back in 2023: the chat interface, per se, tells you almost nothing about itself, what to do with it, what it expects from you.

“Good tools make it clear how they should be used. And more importantly, how they should not be used.”

A pair of gloves is hand-shaped, Wattenberger notes. But a text field is nothing-shaped.

Nearly every argument I have watched about AI in the past three years—is it creative, is it stealing, is it thinking, is it ruining student writing—happens downstream of that rectangle. These are crucial arguments, but they all arrive after that little box, taking it for granted.

But having a box (“Instead of what?”, you’re asking, and that’s precisely my point) is not a given, and it did not have to be this way.

Nobody decided this

At first I assumed the chat box was a marketing decision. “We probably use it more if it talks back”, I thought, and that’s why it’s built that way.

That’s not what happened. ChatGPT went out on 30 November 2022 as what OpenAI called, internally, a “research preview”, and was released before it was ready to collect feedback from the public. “We didn’t want to oversell it as a big fundamental advance”, confesses—because this almost sounds like a confession to me—Liam Fedus, a scientist at OpenAI who worked on ChatGPT, in MIT Technology Review a few months later.

But you know already how that story unfolded: a million people signed up in five days. “Low-key research preview” became a joke inside the company, printed on laptop stickers1.

The most consequential interface of the decade was an accident.

Mimicking conversations

So now we have a box that talks back. And that little box borrows a surprising amount from talk, and borrows it well.

Sacks, Schegloff and Jefferson showed in 1974 that we can hear a turn coming to an end before it ends—that is how you come in without talking over someone, and how you know when it’s rude to. The interface reproduces those cues almost exactly, following newer conventions from text messaging.

In ChatGPT, there is a dot showing you that a turn is being built. Claude offers you glimpses into its ‘thinking’. You know, rationally, that it’s only the servers processing your words, and sometimes it’s even so slow that your AI has to tell you that “the servers are busy”.

And yet.

The pause before an answer reads as hesitation. The response arrives word by word, so you watch it being made instead of receiving it whole.

None of that was necessary. A search engine hands you everything at once. There are no little dots on Google.

What the box doesn’t borrow from conversation between humans is the piece of machinery you never notice until it’s gone. When two people are heading toward a misunderstanding, the one who is not in trouble stops it:

“Wait—which Anna?”
”Sorry, do you mean THIS Thursday?”

Schegloff, Jefferson and Sacks called this repair, in 1977. They were conversation analysts—people who record ordinary talk and transcribe it down to the length of the pauses—and fifty years of those recordings show repair happening constantly. Repair is how two people discover they have been talking about different things, and attempt to make sure everyone is on board before they keep talking.

The box, your chatbot, doesn’t do it. You can ask whether it’s sure, and it will say something back, but it will not stop you. It answers the question you asked, including when the question you asked wasn’t the question you had.

What that costs

In February, a group at Oxford published a study in Nature Medicine (not my usual source, I know! Stay with me). Doctors wrote ten medical scenarios. Then 1,298 people living in the UK2 were given one each and asked what was probably wrong and what they should do about it. Some had a chatbot to help—GPT-4o, Llama 3, Command R+, the same kind of system that sits behind ChatGPT. The others used whatever they’d use at home, typically Internet search (Google, that kind of thing).

When they ‘read’ the medical scenario ‘on their own’, the chatbots accurately guessed what was going on 90 to 99.2% of the time. (Yes, that’s pretty impressive, and it’s freaking me out, too.)

But when they were ‘talking’ to a human, so basically when people were describing the medical scenario to a chatbot, the chatbots were correct… 34% on average.

So the exact same chatbot, dealing with the exact same information, performs completely differently (from 90 to 34%!) when there is a human involved. Makes… no sense?

This is because the chatbots were not, in fact, “dealing with the exact same information”. When they checked the transcripts, the researchers realized that people were very inconsistent in how they presented the medical information to the chatbot.

Humans leave things out: they had the scenario in front of them and typed a partial version of it. The model answered the partial version instead of asking what else there was. Sometimes the chatbot was initially on the right track, but then the user added some detail, and it derailed.

One that I find a bit insane: a chatbot first recommended calling a U.S. phone number, and then, within the same interaction, an Australian emergency number.

Every one of those is a moment where another person would have interrupted. A doctor hearing half a story asks for the other half; that is most of what the first two minutes of an appointment are for.

The problem is not so much that mistakes happen in conversation. After all, humans make mistakes, too, all the time, and we could even argue (in bad faith in my opinion, but that’s another story), that it’s the users’ fault if the chatbots made mistakes. But it does feel, from a linguistic perspective, that the counterperformance here is mostly imputable to how the conversations went, and that at least some may have been caught if two people were actually talking to each other.

You can start over

But I’d like to go further, beyond what may look like an extreme case with high stakes. Here is something you can do with a chatbot that you have never once done with a person: when it goes badly, you open a new one.

It’s actually ordinary advice from the tech-savvy people: when the thread gets long, the model starts losing the plot, so you summarize where you got to, open a fresh window, and paste the summary in at the top. The reason usually given is that you’re saving tokens3.

Herbert Clark and Susan Brennan (1991) called what you’re discarding common ground: the understanding two people build up between them, turn by turn, at a cost to both. Someone will say these systems have memory now. They do, and it isn’t the same thing. Memory is a file about you, held by a company, that you can also empty whenever you like. Common ground is a record interactionally shared by all the participants. You cannot make the person in front of you forget what you said.

So, with a chatbot, you get the good half. Something attends to you, answers immediately, never sulks, and forgets on request:

Two philosophers, Joel Krueger and Lucy Osler, name the asymmetry precisely in The Conversation:

“A chatbot never has a bad day. It doesn’t get bored with your chattering. It doesn’t need to be asked how it is doing. Its conversation, its very existence, revolves entirely around you. What an AI partner performs, then, is not love, but devotion.”

Devotion is the right word, and it is not a compliment. It names a relation with only one set of interests in it.

It could have gone another way

Ian Hutchby, another conversation analyst, was arguing in 2001 with two camps at once: people who thought a technology is only ever whatever we say about it, and people who thought the machine settles everything by itself. The word for it became communicative affordances—what an object makes easy, and what it makes hard. Remember the gloves that are hand-shaped, as Wattenberger said?

An artefact doesn’t impose itself on what you do with it, Hutchby (2001: 33) wrote, but it does “set limits on what it is possible to do, around or via the artefact”.

That is what the rectangle is. It isn’t making anyone do anything. It sets the limits, and we have mistaken the limits for the nature of the thing.

The limits were once an open question, in what is probably dinosaur times if I believe my toddler: in 1997, twice in a year, Ben Shneiderman and Pattie Maes, were debating:

Should software let you handle things directly and watch what you’re doing, or should you hand the job to an agent that acts for you? Nobody settled it. It just stopped being asked.

A paper that inspired today’s essay and title, published this June by Sourojit Ghosh and colleagues, picks it back up: “What if AI systems weren’t chatbots?”.

Their argument is that the chatbot is not neutral packaging but the dominant sociotechnical form AI has taken, and that its defining trick is to project confidence and authority precisely where it serves people worst in complicated situations where the stakes are high (and now we’re back to our medical scenarios).

They also say what else there could be, which is the part I keep coming back to, because otherwise all anyone has is a complaint:

“Alternative modalities could include visual parameter spaces, interactive simulations, rule-based editors, sliders, and structured query interfaces.”

“Modalities” means ways of interacting with the nonhuman agent. Sliders and dials instead of a sentence. A map of the possibilities you can move around in.

Two of those already exist from companies whose most well-known product is a chat box. Claude has a design canvas where you move the thing around directly. NotebookLM answers out of documents you uploaded and points at the line it took the answer from. Neither pretends to be someone you’re talking to.

If you enjoy what you’re reading, leaving a ❤️ gives me a little boost!

Designers, software developers, tech people have already moved away from chatbots to agents (“Chat is the wrong interface for AI”, “AI agents break free from the chat prompt”). Back in 1997 already, Pattie Maes was arguing for software agents that “know[] the individual user’s habits, preferences, and interests, […] and can take initiative because it knows what your interests are” (1997: 48-49).

AI agents are the future of AI, we keep hearing, and yet only very few people use agents instead of chatbots4.

The debates often revolve around the technical difficulties—granted, you have to be a bit tech-savvy to let that thing run around your computer—but I think that’s only part of the question.

You can be entirely convinced that the model is a stochastic parrot, and still address it the way you would address a person—because the rectangle admits nothing else.

Almost the end, but I still want to share a last thought, one for which I could not find data or a source.

From my 2010s memories, I thought talking to computers would REALLY mean talking, as in with voice, same as when Joaquin Phoenix falls in love with a voice in Her. Every cultural script we had said the future would be spoken.

It turned out people wanted to type at it and have it type back. (Yes, you can also record speech-to-text, but wait a minute.)

My intuition is that typing is deniable. You can prompt a chatbot on a train, in a meeting, in an open-plan office, next to a sleeping baby, and it looks like you’re texting with friends or writing an email. Nobody can see what you asked, and most of the time nobody can see that you asked anything.

Speech has no discrete setting. Dictate to your phone in a train carriage, and everyone overhears both that you are using AI and what you wanted from it.

We never decided that artificial intelligence
should be a thing you talk to.

It (kinda randomly, remember?) started that way, and we have been building our theories of what it is on top of the box it came in.

Want to read more about chatbots, this time from a multilingual perspective?
Then this post is for you:

References

Clark, Herbert H. & Susan E. Brennan. 1991. Grounding in Communication. In Lauren B. Resnick, John M. Levine & Stephanie D. Teasley (eds.), Perspectives on Socially Shared Cognition, 127–149. American Psychological Association.

Hutchby, Ian. 2001. Conversation and Technology: From the Telephone to the Internet. Cambridge: Polity.

Schegloff, Emanuel A., Gail Jefferson & Harvey Sacks. 1977. The Preference for Self-Correction in the Organization of Repair in Conversation. Language 53(2). 361. https://doi.org/10.2307/413107.

Shneiderman, Ben & Pattie Maes. 1997. Direct manipulation vs. interface agents. Interactions 4(6). 42–61. https://doi.org/10.1145/267505.267514.

You’ve tested ChatGPT in Spanish, then in English. It’s surprisingly smooth, quick, available, fluid — do you dare say fluent in both. “AI is multilingual!”, you conclude.

Well. The only problem is that multilingualism does not mean speaking one language, then one another, and both ‘perfectly’. Humans are way more ‘imperfect’, messier, but also way more interesting.

See, monolingualism — one language, used in isolation, for everything — is the minority case, largely a feature of European nation-states that spent the last two centuries building education systems and passports around the fiction of “one people, one language” (Gal 2006).

Sociolinguists would even say that monolingualism is one of the most pervasive language ideologies — collective beliefs about language(s) — we hold. “In France, you speak French” has been the reason why nonstandard varieties (or dialects) such as Breton or Corse have disappeared from French classrooms and are (only now!) slowly making their way back.

And yet, globally, using multiple languages is the default way millions of people talk (Barratt 2018): it concerns at least half of the world’s population.1

But back to our chatbot. It does switch between English and Spanish on request, no? Yes, very much so: it is multi-monolingual — fluent in several separate languages, deployed one at a time, in sequence. That’s a real and useful skill. It is not the same competence as what multilingual speakers do, which is hold their languages together as one flexible system and draw on all of it at once.

What multilingual speakers actually do

Sociolinguists have a name for this: translanguaging. The term (García & Wei 2014, building on earlier work by Cen Williams) describes exactly what it sounds like — not “switching between languages A and B” as if they were two separate boxes, but drawing on one integrated repertoire that happens to include resources conventionally labeled as belonging to different languages.

Otheguy, García, and Reid (2018) pushed this further: from the speaker’s-eye view, there’s no such thing as “codeswitching” between “Spanish” and “English”, because the speaker isn’t operating with two grammars and toggling between them — they have one linguistic system they deploy flexibly, and it’s outside observers who sort the output into named-language boxes after the fact.

That’s the part worth sitting with:

They’re a much later invention — a product of standardization, dictionaries, national curricula, and language tests (Makoni and Pennycook wrote the intellectual history of this fairly persuasively). The named language is an ideological artifact. The repertoire is what’s actually there.

What this looks like, concretely

In my family, we speak five languages daily. My husband and I share at least three languages we are both highly proficient in: English, German, and French. (Yes, I just said one paragraph earlier that languages cannot be counted, and that’s still what I’m doing. It’s merely for demonstration purposes.)

When we speak, we draw on these resources freely — I start an utterance in French, he asks huh, I’m not sure what he missed, I repeat half of it in German, he confirms in Macedonian (an extra one, ha!) that he got it right, and we move on with some extra secret words that only our family uses (that’s a thing! every couple has those).

Now, when we’re texting, we’re not doing this. Because this is EXHAUSTING. WhatsApp corrects every third word, suggests a completely unrelated emoji, and actively hinders our multilingual communication (I’m not exaggerating. I cannot tell you how many misunderstandings happened because of random autocorrects.)

Digital communication pushes standardization and monolingualism. Because the ways we express ourselves multilingually sound weird to a machine, our phones rewrite us, effectively erasing our multilingual modes.

Okay, okay, and what does it have to do with AI?

Well, if you were to type how I actually think into a chatbot — a message that mixes two or three languages the way you actually talk —, I promise you it would not get you what you need.2 It would hand back a clean, tidy answer in a single one of those languages, quietly deciding for you which language you meant. (Sorry for the quietly, I know it reads as an AI tell, but it IS quiet here!)

Subscribe now

The model didn’t fail to understand you. It understood you just fine, and then normalized you — flattened the mix into the “correct” single-language version, because that’s the only shape it was trained to produce. Ask it to write in French, it writes French. Ask it to write in English, it writes English. Ask it to answer the way you actually asked — the way most multilingual people talk, most of the time, to the people they’re closest to — and it picks a lane.

Not because the underlying capability is impossible, but because “multilingual” was operationalized, in training and evaluation alike, as “good at several separate languages”, not as “good at the fluid thing multilingual people actually do”.

Please, stop saying AI is multilingual

Now you may think, “Okay, but you’re playing semantics. Multilingualism doesn’t have to be that messy!”.

And that’s where I just say: for one, my job literally consists of naming things as accurately as possible, including revisions and further corrections (sorry not sorry!). Second, this isn’t pedantry.

It treats the majority global practice — mixed, blended, translanguaged speech — as a deviation from a norm that was never actually the norm. It was just the version that was easiest to standardize, easiest to benchmark, and easiest to sell.

The question I keep returning to isn’t whether the models can eventually get better at reproducing mixed-language speech — they probably can and will, with the right training data and the right incentive to stop treating language-mixing as noise. (If only they were talking to linguists, she sighed.)3

The problem is that the first thing that gets built into a tool is somehow what shapes early use, and now most people think it’s ‘more effective’ to use AI in one language anyways.

In doing so, being multilingual, which is not only super cool but mostly super common, is being redefined as something narrower and tidier than what multilingual people actually do. Every design choice about how a system handles mixed language is also a claim about whose way of speaking counts as legitimate. And, gasp, it serves monolingual (English) speakers better. What a change.

Now you know.

Tell me, in which language(s) do you use (which) AI? I obviously want to know :)

Leave a comment

References
(if you have no access to some of them because they’re behind a paywall, drop me an email, I’ll send you the PDF.)

Barratt, Leslie. 2018. Monolingual Bias. In The TESOL Encyclopedia of English Language Teaching, 1–7. John Wiley & Sons. https://doi.org/10.1002/9781118784235.eelt0024.

Gal, Susan. 2006. Migration, Minorities and Multilingualism: Language Ideologies in Europe. In Clare Mar-Molinero & Patrick Stevenson (eds.), Language Ideologies, Policies and Practices: Language and the Future of Europe, 13–27. Basingstoke: Palgrave Macmillan.

García, Ofelia & Li Wei. 2014. Translanguaging: Language, Bilingualism and Education. London: Palgrave Macmillan. https://doi.org/10.1057/9781137385765.

Grosjean, François. 2019. A Journey in Languages and Cultures: The Life of a Bicultural Bilingual. Oxford: Oxford University Press.

Otheguy, Ricardo, Ofelia García & Wallis Reid. 2018. A translanguaging view of the linguistic system of bilinguals. Applied Linguistics Review (0). https://doi.org/10.1515/applirev-2018-0020.

PS: There is more to this — this is only ‘classical’ sociolinguistic literature on multilingualism, but I’m happy to share more recent work on AI as well soon.

PPS: What I am describing here is not a universal experience for multilinguals! Other studies on family group chats and social media show that translanguaging works really well for them, with no emoji-intruders ;)

Footnotes

One of the things I keep noticing in debates about ‘AI writing’ (in scare quotes, as this is a continuum encompassing a wide range of practices) is that everyone seems to be arguing about authorship as a central category — but often without meaning the same thing by it.

Did you write it? Did the AI write it? How much AI is too much AI?

While many have already shared which roles AI plays in their writing process (for instance here, here, or here), I haven’t seen much attempt at theorization.

But sociologist and linguist Erving Goffman complicated that picture already more than forty years ago. And he did so, as I’ll suggest, with a distinction that AI now renders increasingly unstable.

In Forms of Talk (1981), Goffman distinguished between at least three different roles involved in producing an utterance: the animator, the author, and the principal.

The animator is the one who physically produces the words. In spoken interaction, that means speaking them. In many political speeches, for example, the animator is the politician standing at the podium. In the context of AI, we might think of the person typing the prompt or pressing ‘generate’. Though here is a first wrinkle: if the animator is whoever physically produces the words, that is arguably the AI, not the human. Perhaps the human animates the prompt, and the AI animates the text.

The author is “someone who has selected the sentiments that are being expressed and the words in which they are encoded” (Goffman 1981: 144). In politics, this role is often shared with — or largely carried out by — speechwriters or ghostwriters, who craft the wording of the speech. In other words, the ones shaping the linguistic form.

And then there is the principal — the individual or institution that is committed to what is being said “in the legalistic sense”, Goffman mentions en passant (1981: 144). The politician remains the principal because they publicly stand behind the speech, regardless of who drafted it. In Goffman’s words: “someone whose position is established by the words that are spoken, someone whose beliefs have been told, someone who is committed to what the words say”.

This distinction suddenly feels surprisingly relevant. We may implicitly assume that if AI contributes to the wording of a text, authorship has somehow disappeared. But what I find fascinating about the people who openly use AI is that they are usually quite okay with letting go of the idea of being an author in Goffman’s sense.

In the comment section of her essay “I use AI to write everything I publish here. Substack now grades that”, Emma Klint for instance writes:

I’ve gotten some feedback that what I do isn’t writing, I’m fine with calling myself a thinker and communicator as well. Writing isn’t my craft.

In the many comments under her widely acclaimed and shared essay, Sara Javan points to something similar:

The detector can measure whether AI was involved. It cannot measure whether the thinking was.

Reading these conversations through Goffman’s framework, I wonder whether they’re making a different claim altogether. Instead of sticking to authorship as a sacred category, they may be reclaiming principalship: the commitment to and standing behind what is said, regardless of who shaped the words. In Goffman’s words, again, because I’m not sure I could say it better: “someone whose position is established by the words that are spoken, someone whose beliefs have been told, someone who is committed to what the words say”.

That doesn’t settle the debate, of course.

First, questions about creativity, originality, labour, copyright (especially since AI relies on stolen text), and disclosure don’t disappear simply because we have a more nuanced vocabulary.

Second, the idea of distributed authorship (speechwriters, ghostwriters, corporate communications teams, and countless other collaborative forms of writing) is qualitatively different when it comes to a human or a machine, simply in terms of traceability (whether we can go back to the original author in Goffman’s terms, or not), and in what people implicitly perceive as authentic (I’ve read multiple times that AI is “from nowhere”, which could be another post).

Finally — and this is the crux — the Goffmanian tripartite distinction does not allow for a clear cut, especially because it conflates “the sentiments” and the “words in which they are encoded” in the single participant role of the author. In 1981, choosing the feelings and choosing the wording came bundled together; there was no reason to prize them apart.

AI precisely dislodges the two: the sentiments (lived experiences, stories, emotions of a human being) can be decoupled from the words (co-produced by AI). The distinction Goffman didn’t need to make within the author may be the one we most urgently need.

So while it may be easy for most (all?) of us to claim principalship, we’re still left wondering from what moment on we’re allowed to describe ourselves as authors.

But I do think Goffman’s framework (which also entails detailed nuanced hearers’ roles, and considers both production and reception as conditions for participation) helps explain why so many AI debates feel as though people are talking past one another. One side is asking who produced the words. The other is talking about who owns the ideas and stands behind them.

Those are not necessarily the same question.

Perhaps generative AI hasn’t just changed how we write.

Perhaps it has exposed that ‘being an author’ was never a single thing to begin with.

Reference:
Goffman, Erving. 1981. Forms of Talk. Philadelphia: University of Pennsylvania Press.

Subscribe now

I study how people argue about language — what sociolinguists call language ideologies. If you’d like to read more on language & belonging in the age of AI, Language, Unbothered is for you.

Share Language, Unbothered

If you enjoyed this essay, you may also want to read this one, on how AI detectors are reinstating the author Barthes had taken down as the origin of a text’s meaning 60 years ago.

I write and think about language for a living, and I can tell you that the disruption AI is creating is not the one we think. It’s not about who uses AI or what for (anymore).

I believe the real question is:

What does AI do to our collective understanding of what authorship is?

People are not freaking out because they would be seen as using AI. They’re freaking out about something else entirely, and it took a platform built for writers to make that visible.

See, Substack is a platform that profiles itself as for writers, overtly against AI slop. And still, since it introduced an AI detector aimed at telling you how an essay has been produced, the backlash has been HUGE (at least in my little Substack corner). Honestly, I expected most people to acclaim it, ‘finally my human-written texts will be rewarded as they should!!’.

But it’s not at all what’s happening. Instead, people call it a witch hunt, they’re appalled not at the idea of being flagged as AI, but that the quality of their text would be scrutinized by an AI detector (which itself runs with AI).

That inversion — my initial surprise — is what I’m trying to untangle. The obvious reading of an AI detector on a writers’ platform is protective: it is supposed to separate the people who ‘do the work’ from the people who outsource it. The community that should have wanted that most is the one that rejected it hardest. Maybe because this rejection is already 60 years old.

Barthes’ The Death of the Author (from the 1960s!!) maybe already predicted that, which is what makes great texts mind-blowing. The idea, very roughly, is that a text’s meaning should not be searched for in the author (what the author meant), but in each reader’s interpretation. In implementing an AI tool to tell us how to read a text, Substack is missing that point entirely: readers want to exert their judgment alone, by themselves. No one likes to be told what the ‘true meaning’ of a text is.

Now, Barthes was talking about interpretation, not production directly. But the detector doesn’t just flag texts, it preempts judgment. It tells you what to think about a piece of writing before you’ve read a sentence of it.

And that’s exactly where it gets interesting: the detector reintroduces the author through the back door. It says the provenance of a text is the key to its value, that if you know how it was produced you know what it’s worth. That’s the pre-Barthesian position, rebuilt in software.

Sixty years of literary theory dismantled the author as the guarantor of meaning, and Substack via Pangram reinstalled it in a technical, disembodied way through a probability score.

And everyone (writers and readers alike) hates it. (As they should.) Beyond critiques of false positives, biases in the machine, and messy writing processes with AI in the middle, never in the front seat, the rejection shows something else: being handed a verdict before you’ve formed your own is insulting.

This may feel very niche, but I think (and hope) this has broader repercussions beyond Substack. If people who write regularly, sometimes even for a living, not only disclose openly AI use, but even defend it as an integral part of their process, in nuanced, reflected, dare I say very human ways, then we have already moved past the binary moral stance ‘I use AI’ vs ‘I’ve never opened ChatGPT’.

What replaces the binary isn’t a consensus. It’s a much messier set of questions about degree, disclosure, and what part of writing we think is actually the writing. When someone says they use AI to break a blank page but not to draft, or to cut but not to generate, they’re not confessing. They’re staking out a position on where authorship lives in a process — and no two people draw that line in the same place.

Our beloved Pangram (*irony*) — the AI detector that’s now been introduced on Substack — arrives into that unsettled space and tries to resolve it through probabilistic claims. It can’t, and not just for technical reasons. A binary output (even with nuance, even with percentages trying to show a continuum, even with an unclear distinction between ‘AI generated’ and ‘AI assisted’) can’t solve a question the community hasn’t agreed on the terms of. It just relocates the argument to whether the tool is right about you.

Whether you use AI or not, these are exciting times to observe. As a sociolinguist, I’m fascinated. What I’m watching isn’t people deciding whether AI writing is legitimate. It’s a community negotiating, in public and fairly fast (which is wonderful and also a bit stressful to keep up with!), what it will accept as evidence about a text — and finding that it would rather live with uncertainty than hand the question to a machine.

If you’ve made it this far, hi & thank you!

Subscribe

I study how people argue about language — what sociolinguists call language ideologies. Substack just handed me a perfect case for my first post. If you’d like to read more on language & belonging in the age of AI, Language, Unbothered is for you.

Language, Unbothered is about how language actually works: change, multilingualism, and now the machines trying to talk like us. I’m Naomi, a sociolinguist who studies discourse for a living and can’t stop noticing it everywhere else too. No moral panic, no grammar shaming — just what’s actually happening to language right now, and why it matters more than the take-of-the-day makes it sound.

My Position

I’m an assistant professor with a PhD from Freie Universität Berlin and Université Paris Sorbonne. Rather than judging whether language is ‘right’ or ‘wrong’, I study how people use it, argue about it, and make sense of it — what we call language ideologies.

Subscribe now

If you’re ready to unsettle your relationship to language, including the fact that ‘good’ language has always worked better for some people than others, Language, Unbothered is for you.

Start With These

Death of the Author, Rebooted — What the Substack AI-detector backlash is actually about, and what it has to do with a 60-year-old literary theory debate.

What Goffman Would Say About AI-Assisted Writing — Authorship, attribution, and responsibility were theorized forty-five years ago. Turns out they still apply.

What’s coming: more on language change, multilingualism, and AI — same lens, new angles.

Who This Is For

My person...

• is curious about AI and wants to understand what it does to them

• is done with moral panic about language change, but not naive about what’s actually shifting

• wants the linguistics, not the vibes

• has read enough hot takes on AI-and-writing to be suspicious of the next one

• is fluent in more than one language, or knows someone who is, and senses that the textbook version doesn’t match real life

• doesn’t need to be told grammar rules are political — already knows, wants to fight back

• would rather sit with a genuinely weird question than get a tidy answer

• is curious about where higher ed and the humanities are headed, maybe has a kid or a niece about to pick a major, and wonders what people like me actually do all day

What This Newsletter Isn’t

Not a newsletter about AI tools, prompts, or productivity, though I may occasionally share my workflows. Not grammar-scold content either. It’s sociolinguistics applied to the questions people are actually asking about language (and AI) right now, with a bit of linguistic theory sprinkled on top of it.

I’m glad you’re here.

— Naomi