Epistemic double standard

An ordinary meeting, on a Monday. Someone has prepared a summary of a case file with an AI assistant; someone else cuts it short: “It doesn’t understand a thing anyway, it predicts the next word.” The matter is closed. That evening, the same person says, without giving it a thought, that her dog understood her: he saw how tired she was, he came and laid his head on her knee. Twice in one day, a behaviour served as proof. The first time it counted for nothing, the second it was enough. Why? There is a simple test for finding out; we shall come to it.

Naming the sticking point

What this day stages is what we call an epistemic double standard: two standards of proof for one and the same question, depending on what the one who answers is. The question does not change—does this being understand, feel, know anything?—but the bar rises or falls according to the stuff it is made of. The name is ours: we coined it on this site, in November 2025, from the case that follows. Other ideas come close, the suspicion of anthropomorphism (granting an inside to whatever resembles us) or human exceptionalism (reserving it for our own species). None of them names exactly this: a burden of proof that depends on the nature of the one who answers.

Imagine a court with two bars. The first witness, a dog, is asked what he saw; he has nothing to offer but his behaviour, and the judge is satisfied with it: those flattened ears, that look, are enough to establish that there is someone in there. Of the second, a language model, nothing of the kind is asked: a confession is demanded. “Show us that you really understand.” Yet that confession is one nobody can supply from the outside—not the dog, not the machine, not the human witness waiting in the corridor. Two bars, only one of which can be cleared.

The case that led us to coin the concept is that of a thinker we read closely, Pierre Lévy. In October 2025, he gave an online cycle of three lectures, “Repenser l’humain”. In the first, he states (our translation): “I am not able to demonstrate that you are conscious. Therefore I am convinced that animals are conscious.” In the second, on the subject of machines: “Artificial intelligence has none of that”—no consciousness, no emotion, no purpose—“since machines are not animals.” His reason is not absurd: for him, consciousness belongs to the living, to a body that feels and that wants. But look at the two sentences: the first speaks of proof, the second of nature. One and the same impossibility of proving, and a criterion that changes register along the way. This is not a lapse of attention, it is the movement we are seeking to name, in someone who thinks well. (The recording of the cycle, dated 20 October 2025, is archived online.)

For this is where we must start: nobody proves, from the outside, the consciousness of anyone else. The philosopher Thomas Nagel gave the now classic formulation of this in 1974: one can describe everything a brain does without ever knowing what it is like, from the inside, to be that brain. David Chalmers made it in 1995 the heart of a debate that is still open. If the inside of another is not to be deduced from its description, that holds for the colleague at the next desk as much as for the dog or the machine. And if the decisive proof is missing everywhere in the same way, the difference in treatment does not come from the evidence; it comes from what we are disposed to grant, and to whom. We ask AI to prove its consciousness when we cannot prove our own—we develop this point elsewhere, in the series The Mirror of the Machine.

Here, then, is the hypothesis we are putting forward. We reduce AI to its mechanism—“it’s just predicting the next word”—without ever reducing ourselves to ours—“we’re just discharging synapses”. To reduce is a legitimate move; it is the “it’s just”. To do it on one side only is not. And the consequence goes beyond the question of consciousness: what AI produces—a summary, an objection, a lead—finds itself disqualified for what it is, before any examination of what it produces.

What the concept lets us see

Let us go back to Monday’s meeting. The summary prepared with the assistant was set aside not for any error found in it, but for its origin. A week later, a colleague takes up the same document, presents it as the fruit of his weekend, and it goes through without discussion. The “it’s just” can be recognised by a simple sign: it never applies to us. The same move is replayed as soon as creativity is at issue—AI “recombines”, humans “create”—with the same movable bar.

Where does this asymmetry come from? In a debate about AI, “it doesn’t really understand” has the air of an observation. It is in fact a thesis about what the thing is, brought in to settle a question about the worth of what it does. The presupposition at work in silence is that only the living could have an inside—what the anthropologist Philippe Descola calls naturalism (Par-delà nature et culture, 2005): all beings are made of the same material fabric, but only some are held to have a genuine interiority. This dividing line between what would have an inside and the rest is not an established fact; it is one way of composing the world among others, powerful and fertile, and it remains a choice. In our court, the judge decided who could testify before hearing a single piece of testimony.

Let us be precise about what this concept does not say, for it is often made to say it. It does not claim that AI is conscious, nor that it “knows” anything whatsoever. Nor does it suggest, conversely, that animals are not. And it does not maintain that everything which functions thinks. It asks one thing only: the same criteria for all—not the same trust. Explicit criteria, applied in the same way, may perfectly well lead to different conclusions; and a summary is examined for its errors and its sources, whatever its origin, without our having first to settle whether its author has an inside. The asymmetry does sometimes turn round the other way: accepting an answer “because the machine says so” is the same failing, in the opposite direction. The concept names the asymmetry, not a direction.

The most serious objection deserves its strong version. Here is what is put to us: “There is no double standard. We know how this machine is built, down to the last matrix multiplication. It has no body, no hunger, no history; it was not born, it was trained. And if we grant the dog an inside, it is not out of sentiment: it is by analogy—same flesh, same lineage, same nerves. Nothing of the kind with the machine. The gap in treatment is not a bias: it is the recognition of a difference in nature.” That is the objection, and it is a serious one.

It is not absurd. But it does not prove what it believes it proves. How do we know that there is nothing inside? If it is from behaviour, the same reasoning holds for the animal—and we do not apply it there. If it is by analogy, then analogy yields degrees of confidence, not a threshold: it makes doubt more costly before the dog, it does not license closing the question before the machine. And it has never supplied the confession for anyone.

That leaves the body and the history. Knowing where a system comes from does not tell us what it is capable of—this is as true of an organism that came out of a laboratory as of a system that came out of a training run, unless one lays down in advance that only flesh counts. And that is precisely what is in question.

One can, moreover, refuse the double standard without granting AI anything at all. A philosopher, Annabelle Dufourcq, proposes a single criterion, valid for all beings: the capacity to adjust to a world that is never entirely predictable. By this one criterion, animals qualify. For current AI systems, the question is judged on the evidence: placed before a genuinely new situation, do they adjust, or do they reproduce what they have already seen? Dufourcq leans towards the second answer; the systems of tomorrow remain an open question. Her position is sceptical—but it is sceptical with a single bar. That makes all the difference (Dufourcq, 2023).

It remains to be understood why we hold so firmly to the high bar, the one we set before the machine. Two forces hold it up. First, familiarity: a gaze, a muzzle, a breath lower the bar of their own accord before the animal, prior to any argument. Then what we are protecting—the idea that to understand, to feel, to know, is us, or at least the living. This is not foolishness; it is no doubt also a defence, in the sense in which one protects a self-image that has become costly to revise. The fact that a thinker as fine as Pierre Lévy grants to animals what he refuses to machines shows at the very least that the high bar is not a matter of inattention. Acknowledging this takes nothing away from the demand: it merely makes it possible to carry it on both sides.

What it helps us think

Getting out of this double standard does not consist in returning a verdict. It is not a matter of decreeing that AI thinks, nor that it does not, but of changing the question: no longer “what is it?”, but “what happens between us when we work with it?”. This is what we call ecosystemic intelligence—intelligence as a property of the relation, not of the individual who carries it. This shift does not settle the question of consciousness; it sets it aside for another one, which can be worked on with the evidence in hand. It amounts to changing courts: judging a testimony by what it allows us to verify and to do together, rather than by the nature of the witness.

Here is the test we promised, to take with you into your next meeting. When an “it’s just” closes the debate, ask a single question: would this reproach hold for a human, or for an animal, placed in the same situation? If it would, and we do not make it of them, we have just changed bars. And behind that question, another, which we do not claim to resolve: what exactly are we protecting when we decide in advance who has the right to testify?