Focus of the letter 47


Yesterday, the Académie Goncourt withdrew Thélyson Orélien’s C’était ça ou mourir from its selection, for plagiarism deemed proven, in a case where the suspicion of a text largely written by an AI had also been hanging over it for several days. I haven’t read this novel, I don’t know the rest of its author’s work, and it’s not the “affair” itself that I want to comment on here. The Goncourt, as a prize, also judges the award, that’s not my subject: what interests me is what this story has awakened around it. On the networks, the case quickly went beyond the “Orélian case” to become a broader question, almost a shared fear: what if the AI had written a text in the place of the author I am reading, without me knowing it? This fear goes far beyond literature, even if it is certainly exacerbated by the prestige of the literary prize, we all know it. At work, it takes another form: does using AI mean admitting that you are not competent enough in your job? At school, yet another: is using AI cheating? Three different worlds, the same worry deep down, that of a text, a work, a copy that would no longer really be ours.

Michel Serres had already seen this fear coming, even before generative AI. In Petite Poucette, he described a generation that thinks and writes differently, thanks to or because of digital technology, no less well but differently. And in C’était mieux avant, he recalled that each new technology, writing, printing, the calculator, has provoked the same reflex of decline announced. Socrates already feared that writing would weaken memory. AI belongs to this story of technological delegation, but it introduces a difference of degree that can become a difference of nature: it is no longer content to externalize a function, it can participate directly in the production of language, reasoning and form.

I believe that the question asked like this, in this binary form, whether we use AI or not, is never the right one. What distinguishes a use that poses a problem from a use that does not is not the fact of using it. That’s why we use it, when, and especially how.

This year, I was able to experiment for a long time with co-writing with a generative AI, on a demanding editorial project, carried out over several months. The challenge was not technical. It was editorial: to remain in the position of author in the face of a tool that produces quickly, that formats well, and whose proposals sometimes have the convincing appearance of an already finished text. Learn to distinguish what is correct from what is right. Never let a formulation that did not say exactly what I wanted to say, even when it resembled it.

What I kept from this work was not the writing. It was the framework, the sources, the plan, the judgment on each sentence, the final decision to publish or not. AI produced raw material, quickly and in quantity. I exercised editorial authority, from start to finish. This assumed asymmetry responds exactly to the two fears of the beginning. What can be executed can be delegated. What must be judged cannot be delegated: the boundary is not between two watertight tasks, but between two postures, the one that executes the decision and the one that makes it. Being competent in one’s profession or in one’s studies certainly does not mean typing every character or word yourself: it means keeping that judgment, whatever one delegates elsewhere.

This letter that we are sharing is no exception to the question. Every week, an AI and I talk to build it. Every week, I remain in control of the content. I choose what remains, what jumps, what is reformulated. The game Echo, co-written with the same method a few months ago, was already asking this question by living it rather than describing it.

There remains the question of readers, who have their full share in this contract. When faced with a text that we like, what account more for us, readers: whether it was written by hand from start to finish, with a quill or a keyboard, or whether we appreciate it as it is? I believe that if a text holds, if it touches, the way it was made should not be enough to cancel everything retroactively. An AI does not bear the responsibility for a text alone: it generates from instructions and a context, while the responsibility for use and publication falls to humans. But it does not exempt from anything, it only requires honesty. The reading contract is ultimately about saying what has been done.

Should we be able to verify it from the outside when nothing has been said? This is where it gets complicated. A text can be problematic even if it has not been generated by AI, and a text can be produced with an AI without constituting plagiarism. AI text detectors, which have been in high demand in recent days, are themselves only statistical estimates. Their own publishers acknowledge a declining reliability on short texts, and a Stanford study has shown that they more often identify non-English speaking authors as AI-generated. The score of a detector never replaces a statement. It only circumvents it, more or less well, and sometimes it mainly casts confusion and doubt. There is an irony in the discourse about AI text detectors: it is claimed both that the models produce texts that are indistinguishable from human text and that the detectors identify them reliably. These two statements cannot be true at the same time: the capacity of a detector depends on the language, the length of the text, the model, the share generated, human rewrites and the evaluation corpus. A detection result remains a clue, never a proof, and we must be as wary of the myth of undetectable AI as that of the infallible detector.

At work, the question arises with a reversed front: it is not the absence of a foundation that worries, it is the fear that delegating reveals that one could have done without it. The foundation already exists, but in the end it is the legitimacy to use it that is lacking. A real risk remains and it does not concern everyone in the same way. Distinguishing what is correct from what is right presupposes that we already have a base of benchmarks to compare with. For a student who discovers a notion, for someone who is starting out in a profession, this foundation does not yet exist. The machine produces a plausible, well-formatted, convincing text, and there is nothing in front of it to confront it, no experience to cling to feel that a development rings false. The judgment I am talking about above is not summoned from scratch, it is built up with time and practice. This is perhaps more than in cheating or incompetence that the real risk lies: not in the use of AI, but in its use too early, before having had time to acquire this foundation. And this foundation is not only lacking in novices: it is also lacking in those who do not have the time to exercise it, when the workload transforms the chosen delegation into a forced delegation, and judgment into a luxury that can no longer be afforded.

The question is therefore no longer whether AI has intervened, it is already intervening, in our professions, our schools, our reading and more broadly our daily lives. It is to know what we have kept from the judgment, and what we agree to say about it. It is perhaps this grid that will make it possible to get out of the false debate between “using AI” and “not using it”. The same use can be acceptable professionally, artistically questionable and pedagogically prohibited. It is not the tool that is enough to judge the use, but the context: the part really delegated, the judgment preserved and the transparency that accompanies it.

Echo:

• The game Echo of Letter 30 already recounted this experience of delegation lived from the inside: Echo: when making a game about delegation becomes an experience of delegation.
• Letter 42 asked the question of the time we give ourselves to decide: To generate quickly or to decide slowly?