Conscious compromise with artificial intelligence

To critique artificial intelligence today, in most intellectual situations, requires using it at least occasionally. This contradiction is not a moral scandal. It says something about our contemporary condition, provided we pay attention to it.

A recurring situation

For several months now, I have been observing recurring situations in the intellectual and cultural circles I move in. A researcher publishes a rigorous article on the ideological biases of ChatGPT, and the draft was reviewed by Claude. A union circulates an op-ed against the extractive practices of language models, and the text bears all the stylistic signatures of a large language model. A collective of teachers signs a letter calling for a ban on AI in student work, and the report accompanying it was produced with the help of an artificial intelligence.

These situations describe our contemporary condition. They are not the result of individual bad faith. Generative artificial intelligence is now embedded in the tools of intellectual work, as the computer was in the 1990s and the search engine in the 2000s. To critique it requires writing, and writing now passes, at one moment or another, through a device that relies on these technologies.

One can be indignant, denounce hypocrisy, demand pure consistency. I would rather take a different path here, and look at what this situation says about our relationship to the world, in order to draw a few practical consequences from it.

Compromise is not a moral failing

A few months ago, I wrote an article titled Philosophy of Compromise, in which I develop the idea that we constantly live in the gap between our principles and our actions, between our ecological ideals and our daily consumption, between our desire for justice and our accommodations with the system. This dissonance is not a moral weakness one would need to eliminate to become a person of integrity again. It is the ordinary form of our relationship to the real. No situated action escapes effects it does not control, or dependencies it has not chosen. To want purity is to want an act without context.

In my view, the question is not whether we are compromised. We are. The question is whether we know it.

Choosing to look at what we do

In his book Homo sacer (1995), the Italian philosopher Giorgio Agamben works on a figure from ancient Roman law: the homo sacer, the human being deemed “sacred” in the sense of excluded, who could be killed without it being murder, but who could not either be offered in ritual sacrifice. This figure existed in the law, at the margin, without any concept properly thinking it through. It was Agamben who, by naming it and placing it at the center of his reflection, made it a tool for thinking the state of exception, “bare life,” and the contemporary mechanisms of banishment. Bringing something to light does not abolish the situation; it makes it thinkable, and therefore partially transformable.

The same gesture can be made for contemporary compromise. As long as it remains invisible, it acts without our knowing it. We critique AI while using it, and we do not see ourselves doing so. This invisibility is not a deliberate lie. It results from a framework of thought that has not yet integrated the new technological reality.

But naming is not enough. To name presupposes that one has first decided to look. This is, in my view, a philosophical question in itself. To look at what one does, and not only at what one says, presupposes a deliberate act. There is often, on the other side, the opposite movement: not looking, in order not to soil one’s eyes, in order not to have to compromise oneself in thought as one is already compromised in action. Displayed purity is sometimes a refusal to look. It preserves the image one has of oneself, at the price of a form of blindness.

To choose to look is to accept naming, but also understanding, taking an interest in what one depends on, embracing the situation in its entirety. The framework of thought then changes. One does not become consistent. One ceases to believe that one is. This modest operation is enough to open another space for action. One can continue to critique AI while using it, on condition of doing so in the knowledge that one is critiquing it while using it. Consistency ceases to be a precondition. It becomes a practical question: how does one inhabit this gap without being compromised in a serious and inhuman way?

Good conscience as a problem

Étienne de La Boétie, in his Discourse on Voluntary Servitude (1576), perceived that the most tenacious servitude is the one that presents itself as reason. The person who believes themselves free while being enslaved is harder to emancipate than the one who knows they are. The consciousness of servitude is the beginning of freedom.

Good conscience works in the inverse mode. It places an imaginary cordon sanitaire between the person and what they do, and it prevents real work.

Intellectuals who denounce AI while using it are not, in my view, in hypocrisy in the moral sense. They are in good conscience. They hold a critical discourse and believe themselves to be on the side of critique for that reason alone. They do not perceive that their practice informs this discourse as much as their discourse informs their practice. They do their best, without ill will, but with a defect in the consciousness of their own situation.

This defect has two concrete effects. The first is rhetorical. When an op-ed against AI is drafted by AI, the content may be just, but the texture of the text belies the substance. The attentive reader perceives the contradiction. The lobbyists fighting against the op-ed will be able to use it. The cause is damaged by its own voice.

The second effect is deeper. The absence of consciousness prevents one from thinking the situation. As long as one remains in good conscience, one cannot formulate what one is doing. One cannot say “I used AI to draft this text that critiques AI, because we had to act fast, because I am alone in being able to hold this critique without the time it would have demanded, because that is how I was able to act within the deadline I was given.” This sentence, which states the compromise consciously, is harder to write than pure critique. It exposes. It also opens a space for reflection on what one is doing, and therefore for ways of orienting that action.

Knowing what one critiques

There exists a particular variant of good conscience that deserves to be named. It is that of intellectuals who speak of AI without seeking to know it, who may use it here and there, but who have never taken the time to understand technically how it works, or to explore in practice what it can do, what it cannot do, what it modifies in the work of thought.

As I am writing these lines, I have before me a column published in Le Monde on April 26, 2026, titled “If writing allows us to organize our ideas, how will we understand the world if we let AI write in our place?”. The columnist strings together legitimate concerns, citing in succession the American writer Hua Hsu in The New Yorker, an MIT Media Lab study, the linguist Naomi Baron, and Bruno Patino’s recent essay The Time of Human Obsolescence. The tone is grave. There is talk of the “dethronement of the book,” of a “slow atrophy [of the brain],” of a necessary “order of artificial intelligence centered on the human being.” None of these formulations is wrong a priori. None proceeds either from a fine-grained knowledge of the object spoken of. The column addresses an object it does not frequent, and its apocalyptic tone partly stems from this distance. When one knows a little, one becomes prudent. When one does not know, one summons authorities one after another, and holds forth in a discourse from above.

This stance has a social function, which strikes me as little remarked upon. It establishes a legitimacy in the milieu of those intolerant to technology. Critique from afar protects, it places one on the side of lucidity. It no longer serves to know its object, it serves to take a position. Bruno Patino, who proposes in his book an “order of artificial intelligence centered on the human being and at the service of connection,” never says what this formula entails technically, economically, legally. The formula works because it reassures readers already convinced that AI is to be kept at a distance. It is just vague enough not to have to be followed by real work.

This stance contains a more disturbing dimension, which I must name with caution. In the way a part of the intellectual world speaks of AI, one finds, in its cognitive structure, something that resembles a colonial relation to alterity. In the time of slavery and colonization, entire peoples could be exploited on the condition of not really looking at them, of denying them their own human reality, of maintaining a protective distance between them and oneself. I am not establishing any moral equivalence between these situations. A machine is not a human being, and exploiting a technical infrastructure is not of the same order as exploiting persons. But the cognitive mechanism is analogous. Refusing to know what one depends on allows one to continue depending on it without having to modify one’s relation to the world. If one were truly to look at AI, one could no longer treat it as a convenient foreign presence one uses without acknowledging it. One would be obliged to think its status, the conditions of its production, its concrete effects, and to modify in turn one’s own relation to writing and to thought. Good conscience would become harder to maintain.

The attitude I propose is different. It consists in taking an interest. In understanding technically what a large language model is, how it has been trained, what it can do and what it cannot do. In exploring anthropologically what this technology changes in our ways of writing, of reading, of thinking together. In asking philosophically what it says about the human, about language, about collective memory. This curiosity is not adherence. It is, in my view, the condition of a critique that holds.

Three ways of consciously inhabiting compromise

First example. An authors’ society circulates an op-ed about the presumed use of works by AI systems. The text bears the stylistic signatures of a language model. The op-ed is just on the substance and necessary in its purpose. Its authors do their best within a tight parliamentary timeline. Unconscious compromise consists in presenting the text as if it had been written in the human tradition of collective drafting. Conscious compromise, harder to assume, would say something like: “this text was written with the help of an artificial intelligence, within the constrained timeline of the parliamentary calendar; this help does not invalidate the position we are defending, it makes it possible under the conditions in which we work; we ask that this possibility be reserved for the creators we represent, and not confiscated by models trained without their agreement.”

Second example. A teacher refuses the use of ChatGPT by their students in their work, and themselves uses Claude to prepare classes. This asymmetry, formulated bluntly, seems unfair. Formulated consciously, it can be defended. The teacher has already learned to write, the student is in the process of learning, and AI mediation does not have the same formative value in the two situations. This defense supposes that the teacher has reflected on what they do with AI, on what they preserve of their own intellectual work, on what they delegate. Without this reflection, the asymmetry remains arbitrary.

Third example. A cultural institution publishes a manifesto on the necessity of digital sovereignty, and uses Microsoft and Google tools massively in its daily operations. The manifesto is not therefore wrong. It becomes weak if the institution does not specify what it is doing to reduce this dependence, at what pace, with what resources. The work then consists in adding this specification, and publishing the manifesto with it.

A stance for the times to come

This ethics has a scope that exceeds the relation to AI. It says something about the position of contemporary intellectuals. Our era is marked by the speed of technological transformations. No one can be pure under these conditions, and displayed purity tends to become a sign of blindness.

Part of the difficulty lies in the fact that we are in a relation of dependence with these tools, without always being aware of it. Dependence is not in itself an evil. We depend on all sorts of things to live and to think: books, libraries, colleagues, editors. But an unseen dependence acts as a silent norm, it orients the work without our having chosen it. A few elements of method can help to make this dependence more conscious, and therefore more bearable. The first is to name in the very text that one uses AI, when one uses it, and for what. This may seem trivial; it is the operation that distinguishes conscious compromise from masked compromise. The second is to take time, regularly, to take stock of what one has delegated and what one has kept. Was such a revision, such a rephrasing, such a synthesis within my reach without AI, or are they starting to escape me? This taking stock only happens on condition of being decided, and of giving oneself moments where one works without the tool, in order to measure the gap between what one used to do and what one continues to do. The third is to pay attention to the tools one uses, to their hosts, their training conditions, their social and ecological effects. Preferring one tool over another, a host respectful of data over one that is not, is part of the same movement: seeing the dependence in order to recover, within it, margins of choice. The fourth, which conditions all the others, is to continue to take an interest in AI itself, in its technical foundations, its conditions of fabrication, its rapid transformations, and not to remain at the opinion one had at the moment one first formed a position.

I do not believe it is possible to maintain this stance permanently. We are caught in our deadlines, our fatigues, our urgencies, and no one can, at every instant, measure what they delegate to AI and justify it publicly. The consciousness of compromise is a work that is taken up, that slackens, that is taken up again. It is less a matter of reaching a state of consistency than of maintaining an attention. In my view, it is this attention that allows one not to be compromised in a serious and inhuman way. The serious and the inhuman begin when one stops seeing what one is doing. Seeing what one is doing, in the difficulty of daily life, already demands a great deal, and I expect of no one, not even of myself, a perfect consistency. I simply believe that, within the gap, one can keep a modest attention on what one is going through. It is this attention that I am trying here to name.