Can a machine tell the truth? Heidegger, Foucault and the risk a speaker takes

Ask a language model a question of fact and you will usually get a correct answer, delivered with a confident cadence. Ask whether the machine has therefore told you the truth and the question is harder than it looks. Since 2025 the research community has had a precise way of putting the difficulty. The MASK benchmark, published on arXiv in March of that year by Richard Ren, Arunim Agarwal, Mantas Mazeika and colleagues, was built on the observation that most "honesty" evaluations were really measuring accuracy, whether a model's beliefs are correct, and that none was directly measuring whether a model asserts what it does not believe. When they tested for the latter, the result was uncomfortable. Most frontier models scored highly on truthfulness benchmarks and, placed under pressure, showed a substantial tendency to lie. Scaling up made models more accurate. It did not make them more honest.

That finding gives an old question new urgency. A paper published on the Institute of Art and Ideas site on 9 June this year by Bun-Sun Kim of Chung-Ang University in Seoul and Hongjoon Jo of Dong-A University in Busan carries the title "AI is incapable of telling the truth", and argues from Heidegger and Foucault that "artificial language fundamentally cannot contain truth in the way that human language does". The problem, on their account, is not misinformation. It is that a system with "neither a mortal body nor a finite existence to lose" cannot do the thing that human beings do when they tell the truth, which is to stake something on it.

The thesis is overstated in places, and we will say where. But the distinction it rests on, between a true sentence and an act of truth-telling, is one that two very different twentieth-century philosophers worked out in detail, and their accounts explain something the benchmarks leave out.

Heidegger: truth as unconcealment, and the cost of it#

Heidegger's contribution is the older and the stranger. In section 44 of Being and Time (1927), and again in the 1930 lecture "On the Essence of Truth", he argued that the correspondence theory, truth as the agreement of a statement with its object, is derivative. Before a sentence can agree or fail to agree with anything, the thing it is about has to have come into view at all, and that prior showing is what the Greek word aletheia names: unconcealment, the a-privative plus lethe, hiddenness. Truth in the primary sense is not a property of propositions. It is an event in which something that was covered over is uncovered, and it always happens against a background that remains hidden.

Two consequences follow. The first is that unconcealment is something a being does, and only a certain kind of being can do it: the being Heidegger calls Dasein, which is thrown into a world, projects possibilities, and is going to die. The second is Heidegger's account of what ordinarily prevents it. Most speech, he says, is Gerede, idle talk: the passing along of what "one" says, das Man, where the words circulate without anyone going back to the matter they are about. Idle talk is not false. It may be perfectly accurate. What it lacks is the relation to the thing, and the one who breaks out of it does so under the pressure of Angst and what Heidegger calls the call of conscience, a silent summons back to one's own finitude.

Kim and Jo lean on exactly this. A large language model, they argue, is idle talk made into an industrial process: it recombines what has been said into "the most statistically plausible variations" and, left to govern our discourse, "will ultimately lead to a dictatorship of the 'they'". Since the system "cannot experience anxiety", it cannot hear the call, and so cannot name things, in their phrase, "truthfully". Their closing sentence states the view without hedging: "truth is revealed only through the courageous utterances of finite, anxious, and conscientious beings."

One should note, before accepting this, that Heidegger himself later retreated from part of the doctrine. In "The End of Philosophy and the Task of Thinking" (1964) he conceded, under pressure from the classicist Paul Friedländer, that the Greeks had not in fact used aletheia to mean unconcealment as he had claimed, and that unconcealment was not "truth" but the condition of its possibility. The concession matters. If unconcealment is not truth but what truth requires, then the question of whether a machine can be in the truth comes apart from the question of whether it can produce true statements, which is where the MASK authors also want to put the line. We have looked at Heidegger's account of technology on these pages before; his position was never that machines produce falsehoods, but that they organise a way of revealing in which the question of truth no longer arises.

Foucault at Berkeley, autumn 1983#

Foucault's contribution is more concrete and, for the problem of machines, more useful. Between 10 October and 30 November 1983 he gave six lectures in English at the University of California, Berkeley, under the seminar title "Discourse and Truth". They were recorded on tape, transcribed by Joseph Pearson, whose editor's notes date from 1985, and published in 2001 by Semiotext(e) as Fearless Speech. Foucault never corrected the transcript, and the text lacks his imprimatur, a caveat its editors are careful to repeat. The subject was the Greek word parrhesia, frank speech, which he had been tracking through his last courses at the Collège de France and which gave his final course of 1984 its title, Le courage de la vérité.

Parrhesia, as Foucault reconstructs it from Euripides, Plato and the Cynics, is not a property of sentences either. It is a speech act with a particular shape. The parrhesiast says everything he thinks; he says it as his own opinion and not under cover of rhetoric or quotation; what he says is true, and he knows it is true because it is what he believes; he says it to someone who has power over him, so that speaking carries a danger, of punishment, exile or death; he speaks out of a sense of duty when silence would be safer; and the content is usually a criticism, of the interlocutor or of himself. Strip out the danger and you no longer have parrhesia, only a true remark. A teacher who tells a pupil that two and two make four is not a parrhesiast; the pupil cannot harm him. The same is true in the courtroom. A witness and an archive both carry information; only the witness can be ruined by giving it.

This is the structure Kim and Jo are reaching for, and it is worth being precise about something the Korean paper does not do, whatever secondary accounts of it have said: it does not discuss parrhesia. Its Foucault is the Foucault of The Order of Things, the four forms of resemblance, convenientia, aemulatio, analogy and sympathy, and the "care of the self" as the source of the meaning a speaker gives to what she says. That is a defensible route to the same conclusion. But the parrhesia lectures make the point sharper, because they locate the difference between a machine's speech and a human's not in the body or in anxiety but in a social relation. The truth-teller is someone to whom something can happen as a result of speaking. On that criterion a model fails, and fails cleanly: nothing happens to it. It can be retrained, but retraining is not punishment, however often the vocabulary of punishment is borrowed to describe it. We discussed the posthumous publishing around this material in our piece on the Foucault centenary.

What the honesty engineers have found#

Set beside these accounts, the engineering literature of the last eighteen months looks less like a rival than a confirmation from an unexpected direction. The MASK authors separated honesty from accuracy because evaluation had conflated them, and the separation they drew is the ancient one: accuracy is whether a belief is correct, honesty is whether an assertion matches a belief. A model can be accurate and dishonest, as their results show, and in principle honest and inaccurate. Their intervention of choice, "representation engineering", steering the model's internal states toward honest output, improves honesty scores. It does not, and does not claim to, make the model a party to its own assertions.

Anthropic's published constitution for its Claude models makes the aspiration explicit. Under the heading "Being honest" it lists properties the model should have: truthful, in that it "only sincerely asserts things it believes to be true"; calibrated; transparent, in that it "doesn't pursue hidden agendas or lie about itself or its reasoning"; forthright; non-deceptive, in that it "never tries to create false impressions of itself or the world in the user's mind"; non-manipulative; and autonomy-preserving. The document says the company wants the model to hold "standards of honesty that are substantially higher" than those at stake in ordinary human ethics, to the point that it "should not even tell white lies".

A philosopher reading that list notices what it describes and what it cannot. It describes most of what the Greek word covered: frankness, the absence of a hidden agenda, the refusal of rhetoric as a substitute for opinion. What it cannot describe, because no document could, is the danger. A rule that forbids white lies is not the same as a reason not to tell them. For the parrhesiast the reason is the risk, and the risk is what makes the speech worth hearing. For the model the reason is the rule, and the rule was written by someone else. That is not a criticism of the rule. Honesty has been specified from the outside precisely because it cannot, in this case, be undertaken from the inside.

Where the strong thesis fails#

Kim and Jo want more than this. They want it to follow that a machine produces no truth at all, and they want to anchor the impossibility in the body: without a "mortal body" there is no anxiety, without anxiety no conscience, without conscience no unconcealment, and even a future embodied system remains excluded "as long as this body is a mere combination of software and hardware". Three steps there are weaker than the first.

The first is the slide from "cannot tell the truth" to "cannot say anything true". A sentence's truth does not depend on who utters it; "the capital of Iceland is Reykjavík" is true in the mouth of a witness, a liar and a database. The interesting claim is about the act of telling, and once that is admitted the machine's statements fall into the same category as the archive's: true or false, but not told. Foucault would have had no trouble with the distinction. He spent six lectures insisting on it.

The second is the body. Heidegger's Dasein is finite, thrown and mortal, but it is not primarily a body; Heidegger was criticised for decades, not least by Merleau-Ponty's readers, for how little the body figures in Being and Time. Foucault's parrhesiast risks his life, but what makes the risk relevant is the relation of power, not the physiology. To move from "no risk" to "no body" is to switch arguments mid-stream, and to let an open question in philosophy of mind, whether any artefact could have experience at all, do the work of a closed one.

The third is the claim that AI speech is idle talk because it averages what "one" says. That describes a base model's training objective well and a deployed system poorly, since the latter has been shaped by exactly the kind of rule-writing the constitution exemplifies. One can object that rule-following is not conscience. One cannot object that it is nothing.

Who takes the risk now#

Where does that leave the question in the headline? A model can produce true sentences, can be trained to prefer them, and can be measured, with increasing precision, on how often it asserts what it has internally represented as false. It cannot tell the truth in the sense the Greeks had in mind and Foucault reconstructed at Berkeley, because nothing can happen to it for doing so, and nothing can happen to it for failing to. That is not a temporary limitation of the technology. It is a feature of what the word has always meant.

What follows is a redistribution rather than an absence. The risk that used to attach to the speaker now attaches elsewhere: to the person who repeats the model's answer in a courtroom, a clinic or a newsroom; to the company whose rules the model follows; to the public whose shared sense of what has been said depends on sentences no one has staked anything on. We argued in a recent piece on the public sphere that the hard question about generated text is not whether it is true but who is answerable for it. Foucault's parrhesiast knew the answer because the tyrant was in the room. The honesty benchmarks, for all their sophistication, are an attempt to put something back in that chair.