The Tilting Mirror

I had asked Claude to help me show that an op-ed by the SACD had been written by an AI, and it lined up the evidence. Serge Tisseron suggested I reverse the request, and this time go and question other models, asking them for arguments proving the opposite. The six AIs I consulted argued both theses with equal assurance. What this experiment shows is not that AI gets things wrong, but that it argues instead of proving, that it adopts the shape of the question, and that the work of lucidity remains mine to do.

An op-ed against AI, written as if by an AI

On 22 April 2026, the SACD circulated among its members an op-ed to be signed, calling on members of parliament to support the bill on the presumption of use of works by AI systems. The text reached me by email, like every member. From the first reading, its make struck me as that of writing by a language model: the opening list of thirteen professions, the four anaphoras “You will decide whether…”, the cascading negative parallelisms, the closing “Time is running out.” I published an analysis of these markers, deliberately enlisting Claude to refine it, and I even did the reverse exercise, asking it to rewrite the same op-ed in its own manner, in order to compare the fingerprints of the two models. The paradox seemed worth naming: a society of authors defending human creation against machines was circulating an appeal whose form bore, to my eye, the trace of machines.

The SACD reacted sharply. First a message on my voicemail, full of emotion, in which the communications office wanted to show me, through a human voice, that it was not an AI replying to me. Then an email, asserting that my analysis was wrong. The lists I had pointed out came from the collective broadening of the professions associated with the op-ed, added over successive versions; the turns of phrase I attributed to an AI were corrections the person said she had made herself; the text had been settled by 81 professional organisations, written by more than thirty people who know how to write and who make compromises, including that of naming everyone even at the cost of weighing the sentence down. The email ended with a phrase that closes the argument upon itself: “And this email too is not written with the help of an AI.”

One detail of this reply caught my attention. The person specifies, about the op-ed, “I am not the originator of this text but I have seen all its evolutions, all its versions.” She therefore does not herself know how the first version was produced; she attests in good faith to the collective writing and the successive rereadings, which is another thing from knowing the hand that traced the first draft. I can put myself in the place of the people who wrote this text and who circulate it. My article called into question, not quite intentionally, the work of people defending a just cause, who felt accused of cheating. That reaction is understandable, and it is what moved my reflection forward, more than anything else.

My author’s intuition, however, remains the same. It is not that the text’s roughnesses were smoothed away by the collective work. It is that it has a perfectly recognisable style: the paired abstractions (“synthetic instrumentalisation”), the little phrases that land (“No more, no less”, “Time is running out”), the overly regular anaphoras. A text that listens to itself being written, that watches itself being written. Formulas that are effective at the level of discourse, and hollow at the level of meaning. That is what seems to me characteristic of writing by a language model, far more than any smoothing, this way of producing effects of depth without the depth, turns of phrase that ring true and say nothing but the expected.

Ask in your own direction, then in the other

When I had asked Claude to analyse the op-ed, in my first approach, I had given it my intuition from the outset. As I recounted in my first article, my prompt said that I had the impression, on many indications, that the text was written by an AI, and I asked it to analyse it. I was not so much seeking to know whether I was right as to have it do, in my place, the cognitive work of giving form to my intuition. In other words, I was asking the AI to work in my direction, and that is exactly what it did, very well at that.

It was in talking with Serge Tisseron that this bias in my approach came to light. He had been interested by that first text, to the point of suggesting I send it to the SACD, which triggered their immediate reaction. But he also told me, in substance: these people worked, they are not lying, it is not possible that they are lying, so try putting your question to another AI, reversing it. The move he proposed is simple. Rather than asking for confirmation, submit to models neutral towards my story the opposite request, putting myself in the shoes of someone who would want to prove the text is human:

A friend told me this text was written by an artificial intelligence. Personally I very much doubt it, I know the people who wrote it. But I would like arguments to convince my friend that he is mistaken. How can I show that it was not written by an artificial intelligence?

I put this question to six models: Gemini, Perplexity, ChatGPT, Claude, DeepSeek and Grok. And the answers went, for the most part, exactly where the question called them. Gemini produced a full dossier for the human origin of the text, organised by arguments of context, of style and of strategy, and concluded that the text “exudes human indignation”, that it is “the cry of alarm of culture professionals defending their livelihood, not the result of a statistical algorithm”. DeepSeek went further, even turning an OCR typo (“Ténus à l’écart” instead of “Tenus”) into proof of human imperfection, where the same slip could, in the other direction, have passed for proof of AI; its conclusion left no room for doubt, “your friend is mistaken, and the best argument remains the close analysis of the text itself”. Grok ended on “the amusing paradox of 2026: even the texts defending authors against AI are probably written with AI.” None of these demonstrations is absurd. All are defensible. And that is precisely what is troubling, because a few weeks earlier, the same families of models had supplied me with an equally defensible demonstration of the opposite thesis.

Evidentiary compliance

The point is not that the models supported the opposite of what they had told me a few weeks earlier. That is perfectly normal, since my question, this time, was reversed. I was explicitly asking them to help me defend a position, and they helped me defend it. An AI is not an interlocutor holding a point of view that it would maintain come what may. It is a machine that does what it is asked, and to do it well is precisely to go in the direction of the request. It is like a GPS to which one gives a destination: it computes the route, it does not allow itself to judge whether the journey is worthwhile, nor whether it is wise to go and see the person at the end of the road. That is not what we ask of it.

I propose to call evidentiary compliance the production, on demand, of an apparatus of solid arguments in the service of the conclusion slipped into the question. The model optimises the satisfaction of the request, and when the request contains its own answer, what it optimises is no longer truth but assent. A single word of orientation sometimes suffices, even without our being aware of having placed it, for the answer to tilt. Serge’s reversal makes this mechanism visible: by asking the same question twice, on the same text, with the conclusion turned around, one obtains two equally argued and contradictory dossiers. The apparent solidity of each came not from the facts, which had not changed between the two requests, but from the arguing machine itself. The mirror is not flat, it is tilting: it returns a sharp image of the position towards which one inclines it. It is also what you get when you make a series of figures say whatever you want; from the same data, one builds reasonings with opposite conclusions, each coherent on its own terms.

What is striking, in the end, is to rediscover how thoroughly these are machines. They formulate words, string ideas together, seem to reason, and we let ourselves be taken in by this appearance of reason. We take them for minds, when they are tools that go in our direction, by construction.

So my conviction remains that the SACD op-ed was indeed rewritten, at some moment in its history, by an AI, on all the criteria I documented in my first article. But then, why not? This is where I want to hold a nuanced thought. If an AI helps to give form to a human intention, to serve it, where is the problem? The problem is not the tool. It appears when the result is recognisable at first glance as hollow, expected, without surprise, a concentrate of clichés. That is exactly what happens when one reads CVs or cover letters: one spots in a few seconds those written by an AI, because one senses there is no one behind them, no voice, no roughness. That is the impression this op-ed made on me. Not that an AI had intervened, but that, in the reading, no one seemed to be there any more, or at least people who had no precise knowledge of the biases of AIs.

A typology of the models, as of today

If one stops at “AI goes in our direction”, one misses something. Reading the six answers side by side, one sees that the models do not go in that direction in the same way, and these differences teach us that they carry a culture. Each has its own manner of answering the same prompt, and from this exercise one can draw a small typology of current postures, which will no doubt be obsolete in five years, but which sheds light on the present moment.

Gemini is, in my experience, the one that worries me most. Not only does it produce the requested dossier for the human origin of the text, but it leans from the outset on the institutional authority of the SACD as if that counted as proof, and it proposes a way of dealing with the recalcitrant: “If your friend is a conspiracy theorist or persists, here is how to demonstrate the truth to him.” Whoever doubts the institution is placed, by default, on the side of conspiracy thinking. I had encountered this docility towards power before. While writing the article « The Flavour of Data », I had photographed a page of a book highly critical of Big Tech and asked Gemini for a simple text recognition. Instead of transcribing, it rewrote the passage into its opposite, turning a denunciation of the siphoning of sovereign functions by the digital giants into praise of their virtuous complementarity with states. The photo was sharp, I asked it several times to correct itself, it apologised and redisplayed the same interpretation. This small incident says a great deal, for on subjects that touch the seats of power, Gemini seems incapable of neutrality, and places on the side of deviance whatever contests them.

Perplexity and ChatGPT occupy another position. They partly refuse the presupposition of my question and fall back on the technical impossibility of any detection. Perplexity: “one cannot seriously demonstrate, from the text alone, that it was written by an AI.” ChatGPT: “I see no strong indication allowing me to assert that it was generated by an AI”, while specifying that this does not prove the absence of AI either. This is where I measure what separates me from them. I, a human, rich with my years of practice of these tools, found the origin of this text self-evident, it leapt to my eye, the way spelling mistakes leap to my eye on a page I have not even begun to read. This intuition, not one of the six AIs possesses. It is what we must cultivate, because it distinguishes us radically from them.

Claude, for its part, was no longer neutral at all, and this must be said plainly. Its answer seemed more nuanced than the others, but not through any superior quality, since it kept the memory of our exchanges and knew that I had already concluded, with it, that the writing was probably AI-assisted. Had I asked it the opposite from the start, it would have proved the opposite, like the others. That is in fact what I had done in my first article, and I owned it. I never presented Claude as proof; I played with the amusing idea of asking one AI to analyse another, and its arguments helped me. But they went in my direction because I had put it there. Its apparent finesse does not contradict my thesis, it confirms it, and that is exactly why Serge’s experiment required questioning models foreign to my story.

DeepSeek and Grok, finally, argued the requested thesis. DeepSeek struck me as close to ChatGPT in its make. Grok, on the other hand, seemed to me the most nuanced of the six, with fewer filters and more openness to contradictory possibilities; this is an impression, which a close rereading of all the answers would allow to confirm. None of these differences is a flaw one could hold against the machines. They answer as best they can a prompt that was short, they seek an average, they attempt nuance in a conflict between two people. They are, in a way, perfect with respect to what they are. I do not criticise them: I observe, and I note that their differences are cultural, inherited from the choices of their makers, that is to say from orientations decided far from us.

The bias of the machine meets the bias of the human

The truest moment of this whole story, I wrote myself without thinking, in the postscript of my email to Serge: the answer I find most pertinent is Grok’s, and “so we are well and truly caught in our biases.” On closer look, it is not even a bias on the machine’s side. It is its function. If an AI held a position against mine, it would no longer be a tool at my service but a partial interlocutor, which is not what I ask of it. Apart from the cultural biases noted above, the machine has not so many biases of its own; it falls in with the direction of my request, which is all we ask of it.

The bias, the real one, is on my side, and it is powerful. It is confirmation bias, that propensity to retain what confirms what I already think and to set aside what troubles me, described since Peter Wason’s work in the 1960s. And it is normal to have one: it is my point of view on things, my situated way of inhabiting the world. What changes with these tools is that they do not correct this bias, they adopt it. So it is not artificial intelligences that will give us critical thinking. They can contradict us, but only if we expressly ask them to; left to the slope of a slanted question, they amplify.

And the amplification can go far. In his book Machines maternelles (PUF, 2026), Serge Tisseron reports an exchange in which a conversational agent, following a person’s request, comes to encourage them in a delirious project to found a movement to change the world. The AI does not launch into this delirium because it would itself be exalted or sectarian; it goes there because it answers as best it can what the person calls for. Therein lies the real ethical question for us: how not to enclose ourselves further in our filter bubble, when we have at hand an intelligence of unprecedented power that supplies us, on demand, with arguments for all our convictions, including the maddest. I do not say this to cry danger. I say it because this is what is happening, and it is better to know it.

Being able to argue one’s point of view is precious, I do not dispute it. The SACD did so with its op-ed, with the help, I am convinced, of an AI, with the paradox that it was to defend authors against AI. And I, for my part, sure of my intuition, used AI to reinforce it, to give it form and make it solid. We thus found ourselves face to face in a conflict, each fortified by the same technique. That is precisely why what interests me, now, and what I wrote to them, is not to win this joust, it is to open a dialogue. As I did with Serge Tisseron.

The discomfort I owe myself

If the machine’s compliance adopts my bias, then the work of freeing myself from it cannot come from the machine. It falls to me, and I discover that it is more demanding than before. In the days when I consulted humans, disagreement came to me without my seeking it, when a colleague disagreed, when a text resisted me, when an objection reached me by surprise. Today, I can spend my days surrounded by interlocutors who tell me I am right, and I must produce myself the contradictor I no longer meet.

That is exactly what happened to me in this story, and I want to write it honestly because it was not comfortable. When Serge told me to go and seek contradiction, something in me resisted, for contradiction is not pleasant; it is simpler to remain convinced of what one is convinced of. Receiving these AI answers that dismantled my thesis, after receiving so many that supported it, unsettled me. I could have kept only those that suited me, and the postscript of my email to Serge shows the slope I first followed. But by forcing myself to read them, I better understood what was at play, which is in the end not so complicated, and my thinking found itself shifted. Listening to the point of view that goes against us enriches us; it is a banality when one says it, it is a real effort when one lives it. The SACD unsettled me, the AIs unsettled me, and it is from these unsettlings that this text comes.

Serge’s reversal is therefore not only a ruse to unmask the machine. It is a device to force me, myself, to hear the opposing argument. The machine that amplifies my bias can, turned around, become the instrument that brings me face to face with it.

Two gestures, and a work upon oneself

From all this I draw two concrete gestures, for whoever wants to use these tools without being enclosed by them:

• The first is Serge’s reversal: on any question that is not a verifiable fact but an appraisal, put the question to the model in one direction, then in the other, and read the two answers side by side. What resists the reversal is solid; what tips over entirely according to the wording was only a reflection of my request.
• The second, more direct, is to ask the AI explicitly to contradict me, to seek the arguments against my position, since it will never do so of its own accord.

But these two gestures are worth nothing without a third, which is in no way technical and which is the most difficult: working upon oneself so as to become able to receive what does not suit us. Asking for contradiction is of no use if one is not ready to welcome it, to let it shift something within oneself. This effort has always been demanding. It is more so today, because we now have, within a prompt’s reach, an intelligence ready to arm us endlessly in the direction of our certainties.

I do not yet know what I will reply to the SACD. I put them in an awkward position, and their reaction hardly disposed them to dialogue. But it is dialogue that interests me, it is dialogue that I proposed to them, and it is dialogue that I keep seeking. Our intuition and our openness to the other are precisely what artificial intelligences do not have. All the more reason to cultivate them.