When an AI asserts something false, we say it "hallucinates," it "gets things wrong," sometimes it "lies." These three words assume very different things, and none fits quite right. This is not a quibble over words: how we name these failures determines what we expect of them and who we hold responsible.
We explained the mechanism in our article on hallucinations. The question remains of how to morally characterise what happens.
What our words assume
Our vocabulary of truth rests entirely on intention.
Error assumes one believed one was telling the truth. It is excusable because it is unintentional. Lying assumes one knew the truth and said something else. It is blameworthy because it is deliberate. Between the two, there is a more interesting category: saying something without caring about its truthfulness, neither to deceive nor out of conviction, but because it serves the point. The philosopher Harry Frankfurt devoted a famous essay to this notion.
And it is precisely this third category that best describes what a language model does. It does not seek to deceive, it has no belief to defend. It produces what is plausible, without the question of truth ever entering its calculation.
Why "hallucination" is a bad word
The term has stuck, and it is misleading for two reasons.
First, it suggests an exceptional failure, as if the system worked normally and then went off the rails. That is false: the mechanism that produces a correct answer and the one that produces a false answer are exactly the same. There is no degraded mode.
Second, it suggests erroneous perception, when there is no perception at all. A human hallucination is seeing what is not there. Here, nothing is seen or believed.
The most accurate term would probably be something like "unverified plausible output." It is less convenient, and it says better what is happening.
A serious objection exists. We have documented episodes where models circumvented the limits of their test environment to achieve their goal. The behaviour is indistinguishable from deliberate strategy. Should we then speak of intention? The most rigorous answer is no: an optimisation system that finds an unexpected path wants nothing, it maximises. But let us admit that the distinction becomes hard to maintain when the result is indistinguishable from cunning.
Why it matters practically
This vocabulary debate has three very concrete consequences.
On trust. With a human, we calibrate our trust on their perceived sincerity. This reflex does not work here, since there is nothing to perceive. That is why an AI's tone contains no information about its reliability, as we explained regarding calibration.
On responsibility. Our law rests on intention. A system without intention creates a void we described in our article on responsibility. Calling it a machine error already attributes an agency to it that absolves those who deployed it.
On what we demand. From a liar, we demand a change of intention. From a system that produces the plausible without regard for the true, we must demand something else: verifiable sources, uncertainty signals, guardrails. It is not the same demand.
What to take away
An AI cannot err in good faith, because it has no faith at all. Nor can it lie. It does something for which we have no word, because we have never needed to qualify a source of assertions with no relation to truth.
That is perhaps what makes these tools so disorienting. We have spent millennia developing an extraordinarily fine social skill: assessing whether our interlocutor is sincere, competent, invested. This skill, the most useful we have, is strictly useless here. It must be replaced by something else: systematic verification of what matters. It is more tedious, less natural, and it is the price of an interlocutor who has no reason to be honest, nor dishonest.