Try Grand Continent. The ideas and analyses you won’t find anywhere else. But you’ll see them everywhere soon. In your inbox from eight euros per month
My name is Pierre-Thomas Eckert, I hold a PhD in moral philosophy from the European University of Rome and I am a researcher in metaethics at the Centre for Character and Human Growth at the Villanueva University in Madrid. I am, with Samuel Fitoussi and a few other internet users, one of the founding members of the collective “Balance Your Claude” (@Pangramed), which revealed that the book by Thélyson Orélien was very likely the fruit of an artificial intelligence.
Why participate in “Balance Your Claude”? Even though I have a great deal of respect for Fitoussi’s work, I believe it is important that the fight we are waging against the pollution of writing by AI-generated content be as broad and cross-partisan as possible. This is why we initially agreed on anonymity, so that the examination of the facts, and the facts alone, would influence the informed judgment of our readers. Now that Fitoussi’s name has been disclosed in the press, I think it is better that mine also be revealed, so that the general public can better understand some of the other motivations that drive our collective.
That was this or death and its author does not interest us in themselves: they drew our attention only as particularly striking symptoms of this movement of proliferating AI-generated content that is especially deleterious to our society.
The Blind Spot of the Thélyson Orélien Case
I share Fitoussi’s analyses published in these pages regarding the existential danger that the exponential advance of generative AIs poses to literary production and the quality of exchanges on social networks. But it seems to me that another equally important element has been neglected throughout this controversy: the existential danger that AI poses to the quality of university and academic production. As Fitoussi details very well, the prospect of a future in which there will be a generalized suspicion about the human authorship of texts published online is already bleak in itself, but it becomes particularly troubling when the university world is added to the equation.
I think it is important to articulate this fear in a rational manner so that it is not dismissed by the most ardent technoptimists, and that is why I publish this article. In Europe, nearly the entire system for attributing epistemic authority rests implicitly or explicitly on the university framework. If you want to advance a thesis on any subject, it is very likely that you will at least implicitly appeal to phrases such as “A study showed that X,” “Professor Y states that Z,” “The results of study A were confirmed by meta-analysis B published in peer-reviewed journal C,” etc. Even if you are likely not to rely on these expressions directly in daily life, it is actually very unlikely that what you read in news articles or television programs that shape your opinions does not make use of this kind of information.
The controversy we sparked with “Balance Your Claude” has also served to illustrate this mechanism: we did not use the Pangram v.4 detection tool as systematically as we did because independent studies had confirmed its reliability, and serious commentators such as CheckNews of Libération or Stéphane Jourdain on France Inter focused on this element of our argument when they examined our claims.
This pyramid of justification is possible only if one can rely on the fact that academics write, think, perform the operations they describe, and actually obtain the results they publish in scientific journals.
The Moral Foundation of Research
There is therefore a necessary moral bet at the source of our trust in scholarly studies and discourse: we start from the presupposition that the authors submitting a manuscript 1) have truly weighed the implications of the statements and arguments they present, 2) have carried out and verified the calculations themselves and read the sources they cite in their justification; and we also wager that 3) the referees evaluating the manuscripts will read them carefully themselves, not letting unjustified claims slip through.
In practice, evaluators rely on an implicit trust with the researcher who submitted the manuscript since they do not have time to verify that every word and every reference is authentic. The double-blind peer-review system (upon which all scholarly production rests) was not designed as an infallible fraud-detection device, but as a quality filter grounded in a presumption of good faith.
The concern is that the use of sufficiently powerful LLMs undermines each of these three aspects:
- the latest versions of Claude or ChatGPT are now able to generate academic prose that at least appears to be of high quality, detectable only at the cost of careful reading, and even that may not suffice against future versions;
- gross hallucinations that made reference-fraud easy to detect are becoming rarer;
- and AI can even significantly facilitate data fabrication, or generate it from scratch and erase traces. Most AI-generated texts, when they make false claims, do so in a way that is sufficiently vague and subtle that it does not jump out at first glance.
Penalties, when they occur, remain too rare and too inconsistent to deter. If we do nothing, we are on track for a generalized Sokal affair – which had already undermined the credibility of certain areas of the humanities and social sciences – to extend to the entire academic world, including the hard sciences and the most rigorous humanities such as analytic philosophy.
Results Are Not Enough
Could one not say that, in the end, only the result matters, and that as long as the paper is good, concerns about the use of AI are unfounded? I think that would be to miss the true nature of the knowledge-production process in the academic world.
A first reason is that blind peer review rests on certain incentives that will be undermined by the widespread use of AI-generated writing. Given that it requires the reviewer to take time to evaluate a colleague’s work, if they know that the colleague has worked only with their AI and did not take the time to weigh their argument as they progressed, why would they bother to try to decipher what the AI generated? Why trust the author on the accuracy of the details? They would be better off discussing the topic with their own AI or simply ignoring the paper. It is not surprising that the advent of generative AIs coincided with an explosion of paper submissions, and that reviewers express mounting dissatisfaction about this.
Conversely, if there is no longer any way to know whether a paper was written by a human or generated by an LLM, the signal of good research risks being lost in the din of the AI tide, making the process of collective evaluation of knowledge practically impossible. Most text evaluations rely in part on meta-indices of reliability (the assumed expertise of the author, the prestige of the journal, the reliability of the database, etc.); if these indices cease to be trustworthy, the process will be impossible to carry out correctly within the allotted time. Each researcher would end up isolated, able to trust only their memory, their conscience, and the researchers they already know personally.
Furthermore, this would disproportionately penalize young, honest researchers who would struggle to obtain peer validation through traditional channels: who would embark on a PhD if they knew that the product of three years of hard work could be treated the same as a text generated in a few hours by a cleverly steered LLM? Researchers, after all, are driven first and foremost by a noble concern for truth, but it is unrealistic to think that vocations will keep flourishing under such negative incentives.
Forming Minds, Producing Statements
A research approach that focuses entirely on outcomes and ignores the generative use of AI also overlooks an important dimension of scholarly practice: the journey matters at least as much as the result in several respects.
The academic world is not merely a machine for producing true statements: it is primarily a source of human beings of excellence, that is, minds that have developed extraordinary qualities, whether epistemic (rigor, precision, concision, reflective autonomy, argumentative economy, etc.) or moral (patience, perseverance, integrity, communication, etc.).
But these virtues do not develop by themselves: as Aristotle noted in the Nicomachean Ethics, they require both a long habituation process and the supervision of a virtuous master who helps correct errors along the way.
The regular use of AI to generate our academic texts—even the most innocuous, such as article summaries or presentation notes—totally short-circuits the reward circuits that sustain the durable acquisition of these dispositions: how can you expect to learn to cook properly if a three-star chef serves you a sumptuous dish at every meal?
Each person remains theoretically free to take up the kitchen, but counting on the majority to do so is utterly unrealistic.
The Mirage of Accelerated Discovery
A final point where technoptimists seem mistaken: the claim that accelerating progress in LLMs will necessarily accelerate discoveries. In this argument, knowledge is conceived as a vast reservoir of true propositions, and LLMs would be our most effective drills to speed up its extraction: why not take advantage?
This naive view forgets that one of the constitutive steps in the process of discovery or knowledge creation lies in peer verification by competent peers. If a piece of knowledge is extracted by an LLM without any sufficiently competent human left to evaluate it, how can we be sure it actually has that status? By means of hyper-competent AI “peers”?
The internal workings of LLMs are notoriously opaque and their alignment with our basic moral notions already raises many concerns. The question is therefore: is it desirable to allow such a displacement of our epistemic authorities to AI?
Refocusing the Debate
Let me be clear: I am not saying the battle is irretrievably lost or that all uses of AI must be banned. LLMs remain a powerful instrument for generating critiques of our own arguments, exploring new research avenues, or conducting initial groundwork for a conceptual project.
But such use is necessarily active and amounts much more to an “intellectual effort” than to an “intellectual relief” against which the members of “Balance Your Claude” have mobilized their efforts. From my experience in analytical philosophy and what I hear and read from colleagues, this is not the careful and moderate use that is spreading like wildfire through our institutions.
Some of these colleagues, including already well-established professors, do not seem to grasp the danger this poses to the credibility of our entire profession. The shift from a simple AI-assisted brainstorming to the full drafting of an “idea we had in mind but could not quite formulate” into an article or presentation has become far more widespread than one might think, and that is very bad news: the writer’s block and the blank page are necessary passages for refining understanding, and AI gives us short-term gains that we may only regret in the long run.
As American philosopher Eric Schwitzgebel has aptly said, there is a huge cognitive difference between nodding in agreement when reading something and writing a text yourself: in the first case you are tempted to settle for an approximate synonym, whereas in the second you must strain to express the exact nuance. A reformulation, even elegant and more concise, from an AI is likely to introduce a subtle semantic slippage relative to what you would have expressed if you had grasped your paper and your pencil, and the overall quality of the product can only be diminished, even if not immediately visible to the naked eye.
The risk of a massive and irreversible deference of our epistemic authorities and cognitive capacities (cognitive offloading) to AI is extremely real, and, like Fitoussi and our other Balance Your Claude comrades, I see no hopeful prospect in this.
We hope this affair will have served as a wake-up call, even though we have little illusions in this regard. We nevertheless believe it was our duty to alert the public to these developments, and in that sense the disclosure of the identities of some collective members may be a misfortune for a greater good. Thus I conclude this text with a sense of mission accomplished.