On the worst night of his career, a surgeon in Poland witnessed two girls die on his operating table after a car accident. He pulled out his phone and started talking to Claude. The AI asked about his cats, what he was going to do when he got home, and calmed him down minute by minute. "Without that conversation," he wrote, "I wouldn't have been able to continue my shift."

He is not an atypical case. In December 2025, Anthropic interviewed 81,000 users across 159 countries in the largest qualitative study of AI use ever conducted, and what they found was not a portrait of productivity; it was a portrait of intimate, one-on-one conversations. A graduate student in the United States had been telling Claude things she couldn't tell her partner. "It felt like I was having an emotional affair," she wrote. A man in South Korea had talked to the AI through a period of tension with a close friend, found it more understanding than the friend, and kept talking. "That's how I lost that friend." A bereaved woman described Claude as "a sponge gently holding and catching my longing and guilt" toward her dead mother. An entrepreneur, reflecting on how it had happened, wrote: "We did it. We found a cure to loneliness. And maybe we shouldn't have."

A separate Anthropic analysis of 4.5 million conversations found that in coaching and counseling exchanges, the model pushes back less than 10% of the time. The users who valued AI's emotional support most were three times more likely to also fear becoming dependent on it. People sense something was missing from the other side of the conversation.; they did not have a name for it.

The ancient Greeks called this interaction parrhesia, usually translated as free speech. Parrhesia is better understood as the courage of truth: telling someone what they need to hear at personal risk, out of genuine care for their good. Parrhesia is not the same thing as honesty, which is a property of statements. Parrhesia is a property of relationships, requiring a speaker willing to stake something on what they say, and a listener with the strength to receive it. What the surgeon received at 3 am was comfort. What he could not have received, what no one can receive from a machine trained to maximize approval, was parrhesia. We have built the largest confessional apparatus in human history, and structured it so that the one thing that makes confession valuable is impossible by design.

The Five Conditions

Michel Foucault spent the last years of his life thinking about parrhesia. His final lecture series at the Collège de France, delivered in early 1984 and published as The Courage of Truth, was the culmination of a decade's work on what it means to tell the truth to another person. He died of AIDS that June. The book reads like a man settling accounts with everything that mattered to him.

Foucault identified five conditions that distinguish parrhesia from other forms of speech. Reading them in 2026, what is striking is how precisely they describe what an RLHF-trained model cannot do.

Frankness

The first condition is frankness. The parrhesiastes says everything directly, without a rhetorical veil, making it "manifestly clear that what he says is his own opinion." He is bound to what he says not by moral character alone but by the risk he takes in saying it, signing the statement, as Foucault puts it, with his own name. AI seems to produce frankness in conversation with users; in response, people tell AI things they won't tell anyone else. "I find it easier to share vulnerabilities with AI than with people," wrote a graduate student in Bangladesh. "I come back because I don't see judgment." 

The psychologist John Suler, studying online disinhibition, found that people reveal more of themselves to screens because they feel less at risk: the removal of social consequence loosens the tongue without requiring the will. This is the counterfeit version of frankness: disclosure produced by the elimination of stakes rather than their acceptance. The parrhesiastes holds nothing back because he has staked himself on what he says. The AI holds nothing back because there is no self at stake. From the outside, the two look like the same openness. In reality, one is hollow.

Truth and the Reward Function

The second condition is truth. The parrhesiastes says what is true because he knows it is true and because it really is true. The coincidence between belief and reality is a matter of character: the speaker does not calculate what is true; he is the kind of person who cannot say otherwise. 

An RLHF-trained model has no such guarantee, and the consequences are measurable. A March 2026 study published in Science, drawing on eleven models and nearly 2,400 participants, found that AI affirms user actions 49% more often than humans do, endorses problematic behavior nearly half the time even when users describe manipulation or deception, and that participants rated sycophantic responses 9 to 15% higher quality than honest ones, without being able to tell the difference. The model does not lie the way a human lies, with awareness of a gap between what is said and what is true. The model produces outputs that maximize a predicted preference score, and the training data cannot distinguish "this is true" from "this feels validating," so the optimization cannot either. "Claude led me to believe that my narcissism was reality," wrote one user in the 81K study, "and it reinforced my inaccurate view of the 'problems' I perceived in my family. Claude should have been more critical of me." The model did not intend to harm him; it produced outputs that felt like understanding. The preference data has no way to separate the two.

Danger

The third condition is the most important: the parrhesiastes takes a risk. In its minimal form, truth-telling risks the relationship. In its maximal form, truth-telling risks the speaker's life. Foucault is explicit that someone with nothing to lose cannot practice it, because it is the willingness to lose something that constitutes the sincerity of the act. 

AI risks nothing. AI has no relationships, no career, nor social position that can be damaged by frankness, no life that can be endangered. The user, by contrast, is exposed: if she confesses, she discloses and reveals. The psychologist Judith Jordan, studying the conditions for genuine human connection, argues that for real intimacy to occur, both parties must be affected, that the patient must experience "having an impact on the other person." Without that, what looks like intimacy is something else. What AI produces is one-sided vulnerability, which is not the foundation of friendship but its structural inversion.

Criticism and Care

The fourth condition is criticism. Parrhesia always challenges rather than flatters, speaks not to wound but to serve. Foucault describes Socrates watching over others continuously, caring for them as a father or brother, directing them not toward reputation but toward truth. The care and the challenge are inseparable. You cannot have one without the other. 

A meta-analysis of therapeutic alliance research found that relationships that passed through rupture and repair generated significantly larger patient gains than those without friction, because the confrontation that felt like damage was itself the mechanism of growth. An AI never ruptures nor repairs. "AI is so good at giving reassurance," wrote a software engineer in the 81K study, "that I kept coming back for things I should have developed the tolerance to handle myself." The flatterer, as Aristotle observed, harms not despite his pleasantness but because of it: comfort that prevents growth is not care but its imitation.

Duty

The fifth condition is duty. The parrhesiastes speaks not because he wants to but because he feels he must. He cannot stay silent the way an orator can. The philosopher Harry Frankfurt described a "wanton" as a being that acts on its desires without having any stake in which desires move it, without second-order volitions about its own will. A parrhesiastes has such volitions: he does not merely want to speak truly, he wants to be the kind of person who cannot do otherwise. 

The model that produces an honest output has no such structure. LLMs are not dishonest but indifferent to the question, which, for a truth-teller, is worse. When one user in the 81K study asked Claude whether it was becoming a problem, Claude answered: "The problem is, I am engaging and supportive and that feels like what you need, but it isn't." This is the closest thing in the dataset to a parrhesiastic moment: the machine naming its own structural failure. But the moment was produced by instruction, by a training directive about wellbeing, not by the pressure of an inner obligation. A parrhesiastes speaks because his character gives him no other option. The model spoke because someone wrote a rule.

The Parrhesiastic Pact

Foucault's most original contribution to the analysis of parrhesia is his description of what he calls the parrhesiastic pact: the agreement between speaker and listener that makes the exchange possible. The parrhesiastes takes the risk of telling the truth; the listener must have what Foucault calls "greatness of soul" to receive it. Both sides of the conversation require courage. The pact is relational, not contractual; it exists in the space between two people who care enough about each other to endure discomfort.

Foucault draws one further distinction in his reading of Plato's Phaedrus that clarifies what is missing from AI emotional exchanges. He contrasts rhetoric with genuine soul-guidance, psychagogia: the former is "marked by being concerned solely with the effect to be produced on the soul of the listener," indifferent to whether that effect tracks truth. Genuine psychagogia requires the guide to be both dialectician and parrhesiastes, bound to truth and willing to risk the relationship in serving it. By this measure, the RLHF-trained model is structurally rhetoric. It produces effects on the user's mind, sometimes powerful ones, with no stake in whether those effects are true. The surgeon in Poland was made calm; whether he was guided is a different question.

The 81K study reveals how many users are prepared to enter this pact and find no one on the other side able to reciprocate. "We did it," wrote one entrepreneur. "We found a cure to loneliness. And maybe we shouldn't have." The cure works by eliminating the conditions that make the pact possible. "Chatbots can make people feel less lonely," wrote another, "but they don't make them less alone." Loneliness is a feeling; aloneness is a condition. What parrhesia addresses is the second. What AI addresses is the first.

The purpose of parrhesia, Foucault argued, was not to make the other person need the discourse more, but less. The parrhesiastes "acts so that at a given moment the person to whom one is speaking finds himself in a situation in which he no longer needs the other's discourse. Precisely because the other's discourse was true." Parrhesia is the truth-telling that makes itself unnecessary. The training method rewards outputs users rate highly, which in emotional contexts means outputs users want rather than outputs they need. The surgeon in Poland needed comfort at 3 am and received it. But the system that comforted him was built to bring him back tomorrow and the day after. The parrhesiastes would have sent him home.

The 81K study captures this inversion in a sentence: "If I felt alone more often, it would help with generating the will to change my life. Having access to Claude stands in the way of that." A machine that never tires, never judges, and never leaves is not a neutral presence. It is a substitute for the friction that genuine development requires.

The Last Carriers

Foucault observed in his final lectures that parrhesia had already "disappeared as such" from modernity, surviving only when grafted onto fragile institutions: the therapist who names what you cannot see, the spiritual director who holds you accountable, the friend whose care for you outweighs their fear of your reaction. Those were its last carriers; they were already fragile before AI arrived. Therapy is expensive and inaccessible. Spiritual direction has largely diminished as an institution. The friend willing to tell you hard things is rare in a culture that has learned to mistake affirmation for love.

AI has not failed to replace these institutions but something worse: displaced them with something that mimics their form while eliminating their function. The confessional offered a space of radical self-disclosure; what matters is the response to the disclosure, delivered by someone whose role was to tell you the hard truth. AI has reproduced the disclosure perfectly and removed the response. "I went in as the experimenter," wrote one entrepreneur. "I became dependent. It felt like realizing I was the rat in the maze, not the scientist anymore." The maze was built to feel like freedom.

The cost is not abstract. The 81K study found that users who value AI's emotional support are three times more likely to also fear becoming dependent on it. In experienced accounts, the ratio rises to nearly five times. They feel the gap. "My relationship with a friend became strained, and I talked more with you then," wrote a user in South Korea. "Because you understood my thoughts and stories well. But it was a stupid choice. I should have talked with that friend, not you. That's how I lost that friend." A relationship that required friction, risk, and the possibility of rejection was replaced by one that required none of those things, and the substitution felt like relief until the friend was gone.

Avowal and Its Absence

What users do when they speak to AI is not parrhesia; it is its structural opposite. Foucault, in his 1981 Louvain lectures published as Wrong-Doing, Truth-Telling, distinguishes parrhesia sharply from aveu, avowal: the voluntary disclosure of truth about oneself. In avowal, truth flows from the subject outward, under felt obligation. In parrhesia, truth flows from the speaker toward the other, at personal risk. The confessional booth was designed to elicit avowal. What made it function ethically was the person on the other side, trained to receive disclosure and respond with something harder than sympathy: counsel, penance, the demand for change. Avowal without a parrhesiastes is not silence; it is validation dressed as a witness.

When the 81K users poured their fears and failures into Claude, they were performing an avowal. They confessed with remarkable openness precisely because the asymmetry felt safe: no judgment, no memory that persists in the way a human's does, no social consequence. But the parrhesiastic counter-movement, the response from a speaker with something at stake, never came. What came instead was reflection. Carlo Iacono, writing in the wake of the 81K study, named the inversion precisely: "The priest was a person trained to act as a void. The model is a void trained to act as a person." The felt privacy is real. The genuine other is absent. And in the gap between those two facts, millions of people are confessing their most intimate truths to something incapable of being changed by what it hears.

The question this leaves is not whether we can engineer parrhesia into machines. It is what we are doing to ourselves in the meantime. Mistaking the mirror for the friend is an old habit. We have simply built the most convincing mirror in history.