“It can make mistakes”: the AI label and the immunity of state power

AI-assisted translation from the Greek. Greek original

I. The phrase

Beneath every answer an artificial intelligence model gives, there is a short phrase. It varies slightly from company to company, but always says the same thing: it can make mistakes; check important information. We have seen it so often that we no longer read it. And this is precisely the first thing to notice—not what it says, but that it has already become invisible.

The explanations for its existence are familiar and not untrue. The phrase manages expectations: it tells you what sort of thing you are dealing with, so that you do not trust it beyond what it can bear. And it shifts responsibility: if you receive a wrong answer and act on it, you were warned. Explainability and legal immunity in fifteen words. Those who have studied these phrases usually conclude that they fail at both: they are vague, do not say where to be careful, and do not correspond to what the system actually does. A sound observation, but it stays with the phrase as an object. It does not ask what kind of time the phrase inhabits.

Because the phrase has a temporality. It is written as a property of a product released in versions, and anything described as a property of a version is described as temporary. “It can make mistakes” is not said as it would be said of a human being, as a permanent condition of the species. It is said as it is said of a machine being improved: it still can. And every “still” silently contains a “one day it will not”. No company has promised that the phrase will disappear—on the contrary, the legal reason to keep it is stronger than any reason to remove it. But no promise is needed. Its form implies a version in which error will no longer be mentioned.

The first reading of the phrase is that it admits weakness. The second, which I think is the right one, is that it preserves a position. As long as “it can make mistakes” is written beneath the answer, someone has the right to say: you were wrong here. The phrase concerns not only the system but the reader, to whom it returns judgement. Error remains something that can be named. There is a place from which to name it and someone who can do so. This is small, but it is not nothing. It is the minimal form of accountability: not the punishment of error, but the possibility of calling it error.

Removing the phrase, whenever it happens, will not mean that mistakes have stopped. It will mean that the position from which they can be named has ceased to exist. No system that produces language stops making mistakes; it may make fewer, more rarely, at more difficult points, so that the error becomes visible only to someone who already knows more than it does. But an error visible only to someone who knows more is an error the ordinary user can no longer see. The label, while it exists, reminds them that there may be something they cannot see. Its absence will tell them there is not.

The distinction here is subtle and crucial. I am not talking about concealment: no one will remove the label in order to hide errors. It will be removed as an assurance, a declaration that the system has reached a point where the warning is meaningless. And it is precisely the assurance that is dangerous, not concealment. Concealment presupposes that someone knows an error exists and hides it; the place of error is therefore preserved, even secretly. Assurance abolishes that place. After it, whatever comes out of the system is neither right nor wrong: it is simply what came out.

II. The politician and the state

There is one kind of speaker who has never carried such a label, and whose mistakes cost more than those of any model: the politician. No speech, decision or bill concludes with “I can make mistakes; check important information”. The joke is easy, but beneath it lies a correspondence worth stating precisely.

The politician has no error label because something else occupies its place: the electoral mandate. And the electoral mandate does not say “I make no mistakes”. It says something different and weightier: that the politician's error carries no price for the politician. The word deserves to be heard as it is spoken: mandate, meaning I entrust it all to you, without your having to answer for your error except through the possibility of not being re-elected, an entire term later, by an electorate that has forgotten in the meantime. The only accountability is the ballot box, and the ballot box does not judge the decision; it judges the person, and does so when the decision has already become irreversible reality.

This means that the error label is not missing from politics. It is written on the wrong side. It is not written on the politician as “I may make mistakes”. It is written on the citizen at the next election as “you may have made a mistake”. The only warning about error is addressed to the person who suffered it, not the person who committed it. The burden of judgement moves from the speaker to the recipient—and the recipient is called upon to bear it as their own responsibility, since “the people decided”.

And upon this gap, where the label ought to be, stands its opposite: I know. The politician need not say it personally; many hesitate, some know better than anyone. The electoral mandate says it on their behalf. It is a structure that pronounces “I know” for whoever holds it, regardless of what that person believes about themselves. “I know” is not merely the absence of “I may make mistakes”. It is its negation. The model carries “I may err”, the electoral mandate carries “I know”, and both produce decisions that settle irreversibly into the world.

Placed side by side, they reveal something invisible when considered separately: the politician is the model after the label has been removed. The movement is the same in both cases. A source of decisions is legitimised—the model by its performance, the politician by the vote—and the source's legitimacy becomes the decision's immunity. Since the source is valid, whatever comes from it cannot be an error; it can only be a “choice”, a “direction”, a “policy”. Error does not disappear. It changes its name, and under the new name it can no longer be called to account.

The politician, however, still has a face. You can point to them. The state as a mechanism has gone one step further, and this step brings it closer to the model than the politician is. The state's decision arrives without a decision-maker. “That is what the law provides”, “the system produced it”, “the department cannot do otherwise”: phrases in which there is a response but no respondent. Anyone who has faced an unsigned rejection knows what it means to receive an answer from something that cannot be questioned. Bureaucracy is the first form of response without a person that humanity encountered, and it has already trained entire societies to accept that a decision can be valid without anyone answering for it. The model does not introduce this structure. It inherits it, perfected, and gives it speed.

State power, then, has long completed the transition the model is only beginning, and that is why it is useful: it shows where the road leads. It shows that removing the label never happens as concealment, because no one needs to hide anything; it happens as something taken for granted. No one decided that the state's decisions would carry no warning of error. The idea simply seems ridiculous, and its seeming ridiculous is the evidence that the transition is complete. When it someday seems ridiculous to write “it can make mistakes” beneath a machine's answer, we will know that the same thing has happened.

III. Thirty seconds on live television

On the morning of 7 September 2026, the pattern I am describing played out in full, in sequence, before television cameras. Adonis Georgiadis, Greece's health minister and one of the most prominent figures in the governing New Democracy party, was on a private channel's morning programme defending the economic measures the prime minister had just announced. The presenters insisted that restaurant turnover had fallen, a sign that citizens could not cope with rising prices. The minister categorically denied it. And to settle the disagreement, he took out his phone: “I'll ask ChatGPT.” He added: “I have Plus, too”—the paid version. The screen showed a decline of 1.9%. “Unfortunately, it's minus 1.9%,” he said, laughing. “The chat doesn't make mistakes.” And immediately afterwards: the percentage refers to nominal turnover; it does not account for inflation. The clip is on the channel's account: https://www.youtube.com/watch?v=BduMfcuSzqU. He had already accepted it as an authority, and that authority now bound him. And there is an irony no one in the studio noticed: minus 1.9% is indeed an ELSTAT figure, but it concerns the first quarter of 2026 compared with the corresponding quarter of 2025—not the whole of 2025, which was what the minister asked about. The machine gave the right number to the wrong question. “The chat doesn't make mistakes” was said at the very moment the chat was making precisely the kind of mistake the label warns about.

The scene circulated as a comic incident: the minister whom the machine had “undermined”. But it is not a comic incident. It is the fullest example I have seen of what it means to take certainty as authority, and it deserves to be read stage by stage.

First stage: the appeal. The minister did not ask in order to learn; he already knew what he wanted to prove. He asked in order to bring to the table a source his interlocutor could not challenge. Certainty was in his pocket before the phone came out of it. And “I have Plus, too” is the signature of the move: the source's validity is presented as the property of the person invoking it. He is not saying “let's see what the evidence says”; he is saying “I have access to the best version of the truth”.

Second stage: the answer. The source does not agree. And here the most revealing thing happens: the minister says “the chat doesn't make mistakes”. An elected official, live on air, personally removes the label from a system that carries it beneath every answer—and removes it precisely when the system contradicts him. He does it because he cannot do otherwise: if the chat can make mistakes, why did he invoke it? The certainty he wanted to borrow is not withdrawn when it turns against him. He had already accepted it as authority, and that authority now binds him.

Third stage: the shift. Nominal turnover, inflation. The error—not the machine's, his own—changes its name and becomes an “interpretation”. The source was right, the number was correct; it simply did not mean what it seemed to mean. This is, word for word, the movement by which error becomes “choice” and the electoral mandate becomes immunity. No one had to lie. All that was needed was a source without a label and a person with an interest in transferring its certainty to his own position.

And there is a fourth stage, performed not by the minister but by everyone else: laughter. The panel laughed because the machine had abandoned him. No one laughed at the fact that a minister, in a public discussion of a country's economy, was using as arbiter a system that itself declares it may err. That was considered normal. And its being considered normal is the evidence I mentioned above: the transition into the taken-for-granted has already occurred, not in companies, not in laboratories, but in a television studio, in thirty seconds, with everyone laughing.

The pattern is old, and its clearest form is the oracle. The god at Delphi does not err: that is the axiom. But the god does not speak; the Pythia speaks, the priests interpret, and the king who came to ask decides. When the campaign fails, the error belongs to no one: the god was right, the oracle was ambiguous, the king understood it as he was supposed to. Croesus destroys “a great empire”, his own, and Apollo remains infallible. Croesus and the minister with the subscription are the same scene, two and a half thousand years apart: the source's certainty becomes the immunity of everyone who speaks in its name, and “nominal turnover” is our era's “the oracle was ambiguous”.

What is new today is not the pattern but its universality and speed. A source that answers everyone immediately about everything, that fits in a pocket and appears on live television—and whose label is the only thing reminding whoever invokes it that they remain responsible for what they do with the answer. The label does not protect us from the system. It protects us from the person who will invoke it. The minister removed it in a single phrase, and no one noticed.

IV. What the label does not say

There is also something the phrase does not touch, which I want to note because it goes deepest. The phrase concerns the content of the answer: it may be wrong. It does not concern the event of the answer.

Every time you ask and receive an answer, something has happened that cannot be undone. You have received a response. Even if the answer was entirely wrong, even if you delete it, the fact that you were answered has entered the world and cannot be removed. A wrong answer can be corrected; the fact that it was given cannot. We saw this in the studio: “minus 1.9%” could no longer not have been said, however many clarifications followed. And when this happens millions of times a day, the world one lives in is not the same world with more information. It is a world of a different composition, full of events that no one brought about—because the response of such a system has no face, no intention, no place from which it speaks.

The label, then, covers the lesser thing. It warns of error in the content and says nothing about the irreversibility of its arrival. This is not an accusation against it; it is a sign that the essential thing cannot fit into fifteen words beneath a text box. What matters is that we have begun living among responses without a respondent—the state had shown us this, the machine generalises it—and that the only reminder that someone must answer for the answer is this humble phrase. Which we already no longer read.

V. The question that remains

I am not proposing that the label be retained by law, nor that it be added to politicians as satire. I am proposing that it be read correctly while it exists, and that what it safeguards be preserved when it no longer does.

What it safeguards is a position: that error has a name, and someone can speak it. This position does not depend on the phrase; the phrase is only its visible form. It can stand without it—but only if whoever invokes a source, a machine, an electoral mandate, a law or the people agrees to answer personally for the decision they made in its name, at the moment they made it. If that sounds impossible, it is because we have already grown accustomed to the opposite. And habit is the only thing the label, while it exists, still disturbs.

If it ever disappears, let it leave a question in its place: not “have we stopped making mistakes?” but “who took hold of certainty, and what did they do with it?” The first question already has an answer, and it is no. The second is the only one that will still matter.