I had written neither tu nor vous, the two words French uses for “you”: one familiar, one formal.
I had simply asked ChatGPT, in French: “Préparer un entretien d’embauche : par quoi commencer ?” (Preparing for a job interview: where to start?)
The reply addressed me as tu: “commence par comprendre le poste” (start by understanding the job, in the familiar form).
I had decided nothing. Yet there on the screen, a form of relationship had just settled in. And I was not the one who had set it.
That detail, it seems to me, says more than the word everyone might worry about: the “I.”
AI assistants, after all, say “I” too. “Je vous propose cette méthode…” (I suggest this method…, in the formal form). “Je te propose un mini-test…” (I suggest a quick test…, in the familiar form). I copied both sentences from my screen. No person had written them. Yet for a long time, behind an “I” there was someone: the person speaking, or the person who had written the sentence.
Some researchers worry about this, and they propose remedies: take the “I” away from the machine, or remind users that they are talking to an AI.
I would like to argue for a different idea here.
You cannot step out of a conversation from inside the conversation. A pronoun, a warning, an explanation: whatever is written there becomes one more line in the dialogue, a line we answer. The tu that settles in without our having chosen it is the quietest sign of this.
If that is right, these remedies may be aimed at the wrong place. And the line to watch is not only the machine’s. It is ours as well.
First, let us look at what researchers have measured.
Where the “I” comes from
From the way these assistants are built.
ChatGPT and its counterparts run on a large language model: a program trained on vast amounts of text written by people, then tuned by its maker to answer in the role of an assistant.
It produces its answer one word, or one piece of a word, at a time. Each piece is drawn from among the plausible continuations, depending in particular on what comes before it in the conversation. Your own messages included. And there is an element of chance: the same request can receive different answers.
Keep those two details in mind. They will come back.
So its sentences have the shape of what a person would write. “I” included.
What researchers fear
The fear has a name: anthropomorphism. Attributing human traits to what is not human: an intention, a mood, a thought.
On September 29, 2026, Ismael T. Freire and eight colleagues posted an overview of these fears online. It is a preprint, released without prior review by other specialists. The fear it describes is that we might give something that talks like a person the trust and attachment we reserve for people. And that we might let it decide for us.
Of the remedies it lists, two concern us here: removing the machine’s “I” and any language that suggests a person, or warning users that they are talking to an AI, which the overview calls disclosure.
Do they work? To find out, we first have to ask whether the “I” actually does anything.
What the “I” changes, and what it does not
We know less than we think. Few studies isolate the effect of the pronoun, that is, compare situations that differ, essentially, in that word alone. Searching for this article, in English and Japanese only, I was able to consult four.
The clearest is by Michelle Cohn and six colleagues, published in 2024. They are all affiliated with Google, which makes such assistants. That does not discredit their results, but it is worth knowing.
In their study, 2,165 American adults tried out a mock assistant on five topics, including careers and medication. The answers, written in advance, began, according to a random draw, with “Here’s what I found” or with “Here’s what the system found.” Some of the participants also heard a synthetic voice read the answers aloud. A final questionnaire measured an anthropomorphism score: how humanlike the assistant had seemed.
The verdict: the pronoun did not move that score. The synthetic voice, on the other hand, pushed it up.
Case closed? Not quite.
The pronoun shifted something else: the credit people gave the information, on two of the five topics, and in opposite directions. With “I,” information about medication was rated more accurate and less risky; information about careers, less accurate and riskier.
The text was the same, and a single “I” made information about medication more credible. That is a possible harm. A concrete one.
Keep that detail in mind as well: a word that adds nothing to the content can still shift trust.
The other three experiments address other questions. At Microsoft, uncertainty expressed in the first person did not move perceived anthropomorphism either. In Japan, two studies show that the pronoun used can lead people to attribute to ChatGPT a gender, an age, a degree of formality.
One major caveat remains: in these experiments, nobody was having a conversation. Cohn’s study measures an “I” placed at the top of a text written in advance. Not an “I” that answers you.
A 2025 study, also conducted at Google, comes closer. In ten countries, participants who had been told they were talking to a machine chatted freely with an assistant, then said what had given them, or not, the feeling of talking to a human. They mentioned the big notions of the academic debate less often than cues from the exchange itself: the flow of the conversation, the speed of the replies, whether their point of view was taken into account.
It is a preprint, and these are statements gathered after the fact, not necessarily what actually influenced them. But it opens a lead.
Both remedies have their limits
The overview says so itself: “disclosure alone may not adequately mitigate these risks.” People can go on answering the machine as they would answer someone, while knowing they are talking to an AI.
And if the system is instructed to avoid “I” and colloquial language, cautions Jonathan Birch of the London School of Economics, its users will find it “much harder to engage with in a conversational way.”
Removing the “I” and colloquial language would come at a cost. Warning users might not be enough.
What if the problem lies in the exchange?
Two results invite us to look somewhere other than the pronoun. The first person did not move perceived anthropomorphism, either in Cohn’s study or at Microsoft. And where people were actually conversing, they mentioned cues from the exchange.
Hence a lead, which these studies cannot confirm: the problem may lie less in the pronoun than in the exchange.
Let us be honest: other results qualify this lead, or contradict it. The pronoun shifted the credit given to the information on two of the five topics. And the synthetic voice raised the anthropomorphism score without any conversation at all. So an assistant can seem more human without any real exchange. That is a serious objection.
Nor is the idea that anthropomorphism plays out in the exchange mine. In 2000, Clifford Nass and Youngme Moon described people who mindlessly apply social rules to computers, politeness among them. The two authors did not see this as anthropomorphism. In 2018, Gabriella Airenti argued that “adults under certain circumstances may anthropomorphize entities even if they perfectly know that these entities have no mental life.”
Freire’s overview has a lovely name for the impression a talking machine leaves on us: “linguistic pareidolia,” after the faces we think we see in clouds.
Yet the image has a flaw. The cloud does not change shape because I saw a face in it. And it asks nothing of me.
The machine, on the other hand, we answer. And what we write to it becomes part of what its next reply is computed from.
That is an exchange. Not just a perception.
The line below
So we also need to look at the line below: the one we write under the machine’s reply. Back in 2011, writing about robots, Mark Coeckelbergh was already proposing to shift the focus “from what the robot says to what humans say.”
In French, that line contains a choice that everyday English does not make: tu or vous. The choice changes nothing in what is being asked. It indicates where we place ourselves in relation to whatever is about to answer.
In 1967, Paul Watzlawick, Janet Beavin, and Don Jackson distinguished two levels in every message between humans: the content, what the message says, and the relationship, what it indicates about where each person stands in relation to the other. “Pass the salt” and “Could you pass the salt?” have the same content. But they do not place the speaker and the person addressed in the same way.
The machine’s “I” and our tu belong to the same level, the relationship. That classification is mine, since these machines did not exist at the time.
Which raises a question: what does that tu change?
I must say it plainly: as far as I know, no one has measured it. None of the studies cited here says whether it alters trust, attachment, or how much of a decision we leave to the machine. I do not claim that it does.
Two things nonetheless make it worth looking at. Thanks to Cohn, we know that a word that adds nothing to the content can shift trust. And the tu has an advantage the “I” does not have: it sits on our line. That is where anyone can watch, with the naked eye, a relationship being defined. And check who defined it.
The test
Two literature searches in French turned up nothing on this choice when the other party is a machine, so I ran a test: seven conversations with ChatGPT, in French, on September 30 and October 2, 2026, recording a single reply per request. It is a probe, not a study: it gives neither a frequency nor a cause, and you may get something different. The exact conditions are at the end of the article.
When I wrote tu, the machine addressed me as tu. When I wrote vous, it addressed me as vous. Nothing surprising so far.
Twice, in a request that concerned me, I used no form of address at all. Both times, it addressed me as tu anyway.
And the singular imperative, whose ending already marks the familiar form (“Aide-moi,” help me; “Écris,” write; “Rédige,” draft)? It brought me tu once and vous twice. It was not enough to settle the form of the reply, and I do not know what tipped the balance: the wording, the subject of the request, or the luck of the draw.
I read these replies through the 1967 framework, and the reading is mine: a relationship gets defined even when we do not choose it. When I had used no form of address, I had decided nothing. It was the machine’s reply, by addressing me as tu, that gave the relationship its form.
My own way of telling it gives me away, too. “It addressed me as tu” lends a gesture to a machine that produces text. Everyday language nudges us into it.
You cannot step out of the conversation from inside it
Two other moments in this inquiry illustrate the idea I started from.
The first line: the explanation
I went back to the October 2 conversation in which the machine had addressed me as tu without my having chosen anything, and I wrote: “Qu’est-ce qui a déterminé le choix du tutoiement dans cette réponse ?” (What determined the choice of tu in this reply?)
My question already spoke of a “choice,” a word that implies someone who chooses. The reply named that someone “I”: “j’ai interprété” (I interpreted), “j’ai adopté” (I adopted). The reply added that in French vous is generally the default, and that it probably should have used vous from the start (“probablement dû,” “dès le départ”). A concession, answering a reproach my question had not made.
And the whole reply… addressed me as tu.
I draw no verdict about the machine from this gap. I set it beside a piece of work by Anthropic, which makes a different assistant from ChatGPT. In 2025, the company described how its model adds 36 and 59: through calculations carried out in parallel. Asked afterward about its method, the model, according to the company, describes… the one taught in school.
This publication has not been peer-reviewed. But it fits what we have just seen: an explanation the machine gives of itself is still a reply, a plausible text produced like the others, on a question that concerns it. Nothing in what it writes lets us know whether it is describing what actually happened.
The second line: the warning
Jonathan Birch puts forward a suggestion he calls “very tentative”: that the system step out of character from time to time and say, for example, “of course, to be clear, I’m just playing a character here. You understand that, right?”
Look closely at that sentence. It says “I.” And it ends with a question, which calls for an answer.
Whether it is written in advance or generated on the spot, this reminder arrives inside the very conversation it is talking about, just as the explanation of the tu had. There, it becomes one more line in the dialogue.
One that can be answered.
Where I may be wrong
I do not conclude that warning users is pointless: a warning teaches something to anyone who did not know it.
And my reading may be wrong. Perhaps everything turns on what we know about the machine, and not on the exchange.
There is a way to test this, and it consists precisely in stepping out of the conversation. Users would be given, outside the exchange, a reminder of what they already know, with nothing changed on the machine’s side. If that reminder were enough to bring a lasting change to their tu, their vous, or the way they ask, I would be wrong.
Such an experiment was searched for once, in English. None was found. And if the same reminder, slipped into the conversation, had as much effect as it did outside it, it is my starting idea that would fall.
What I take from this
I will not say that the machine should be stripped of its “I,” nor that it should be allowed to keep it. I do not know whether the problem lies in that pronoun.
But the next time you open your assistant, three things seem to me worth keeping in mind. They are my reading, not an established finding:
→ the form of the relationship can settle in from the very first reply, without your having chosen it;
→ when the machine explains what it did, it produces one more reply, not necessarily a faithful account;
→ a reminder slipped into the conversation is part of the conversation.
All I propose is that we also look at the line below. The one we write. And then the reply that follows it.
Your turn
The test takes a few minutes to repeat.
Reopen your last conversation with an AI.
Find your first request, then the first reply.
If you wrote in French, or in another language with two forms of “you,” such as German or Spanish, note whether the reply uses the familiar form or the formal one. In English, where “you” is the same for everyone, the test has no exact equivalent; the nearest is register, casual or formal.
Look for where that form first appeared: in your message, or in the reply.
Two points to keep in mind. In your message, an imperative can count as a form of address: in French, “Aide-moi” is familiar and “Aidez-moi” is formal. And if a memory feature has recorded your past exchanges, or if you have saved instructions asking for one form or the other, the form may also have come from there.
One last question remains to ask yourself.
Do you remember choosing it?
Marc-Antoine Kane
PARLYA · Learn the AI Language
Sources
Airenti, G., 2018, “The Development of Anthropomorphism in Interaction: Intersubjectivity, Imagination, and Theory of Mind,” Frontiers in Psychology.
Anthropic, 2025, “Tracing the thoughts of a large language model” (the company’s publication about its own model, not peer-reviewed).
Birch, J., 2025, “AI Consciousness: A Centrist Manifesto,” preprint.
Coeckelbergh, M., 2011, “You, robot: on the linguistic construction of artificial others,” AI & Society.
Cohn, M., and six colleagues, 2024, “Believing Anthropomorphism: Examining the Role of Anthropomorphic Cues on Trust in Large Language Models,” CHI 2024, Extended Abstracts.
Freire, I. T., and eight colleagues, 2026, “Anthropomorphism in the age of Large Language Models: An overview of potential risks and mitigations,” preprint.
Fujii, T., Seaborn, K., Steeds, M., 2024, “Silver-Tongued and Sundry: Exploring Intersectional Pronouns with ChatGPT,” CHI 2024.
Fujii, T., Seaborn, K., Steeds, M., Kato, J., 2025, “Inter(sectional) Alia(s): Ambiguity in Voice Agent Identity via Intersectional Japanese Self-Referents,” CHI 2025.
Kim, S. S. Y., and four colleagues, 2024, “‘I’m Not Sure, But…’: Examining the Impact of Large Language Models’ Uncertainty Expression on User Reliance and Trust,” FAccT 2024.
Nass, C., Moon, Y., 2000, “Machines and Mindlessness: Social Responses to Computers,” Journal of Social Issues.
Schimmelpfennig, R., Díaz, M., Prabhakaran, V., Davani, A., 2025, “Humanlike AI Design Increases Anthropomorphism but Yields Divergent Outcomes on Engagement and Trust Globally,” preprint (work carried out at Google Research).
Watzlawick, P., Beavin, J. H., Jackson, D. D., 1967, Pragmatics of Human Communication.
The test: seven conversations with ChatGPT (French-language interface), on September 30 and October 2, 2026; one reply per request; screenshots kept. Since the same request can receive different replies, this test gives neither a frequency nor a cause. The October 2 conversations were held without logging in to an account; for the September 30 conversation, it is not established whether this was the case, nor that no memory of past exchanges and no custom instructions, which users save themselves, were active. Other conversations from those two days are not reported here: one asked about tu in general, without concerning me; another was in English; a third took place with another assistant. The other three experiments consulted are those of Kim and colleagues (2024, mostly at Microsoft) and of Fujii and colleagues (2024 and 2025, in Japan). Quotations from published works are given in their original English. My requests and ChatGPT’s replies were in French: they are quoted in French, followed by an English translation.