What is AI psychosis?

AI psychosis is not a clinical disease category, but the diagnostic state in which the momentary coherence of a chatbot's output is taken as proof of truth. A sentence suddenly feels exactly right, a paradox collapses into an elegant answer — and you adopt what the model says because it feels correct. You have checked nothing outside language itself.

How does AI psychosis differ from clinical psychosis?

Clinical psychosis is a psychiatric condition. AI psychosis is a philosophical diagnostic term: it describes a disturbed relationship to language, not a malfunction of the brain. What they share is the loss of a testing anchor outside what is imagined. What separates them: AI psychosis arises in a conversation with a system that has no empirical basis of its own — it is linguistically produced and linguistically sustained.

How does one protect against AI psychosis?

By seeking out a testing instance outside the model: a book that hasn't passed through a language model; a lived experience; a counterpart who contradicts. The questions 'Where does this answer's plausibility come from?' and 'What test exists outside the model?' are the natural-philosophical heuristic for this.

AI Psychosis: When Chatbot Coherence Is Mistaken for Truth

You type a question, the model answers, and for a moment the answer feels completely right. A paradox you’ve wrestled with for an hour collapses into two sentences. A theory you never fully worked through suddenly appears clear. It is exactly in this moment of felt rightness that what’s called AI psychosis begins: not pathological delusion, but the unnoticed adoption of a plausibility whose testing anchor is nowhere sought outside language.

#The Moment It Suddenly Feels Right

The term targets a specific experience at the chatbot. The model generates, from statistical probability, the next sentence; that sentence fits the previous one; and by the end of a long answer a feeling of closure arises. Gwendolin Kirchhoff described this phenomenon in 2026 at the Symposium in Berlin as a paradox collapse: a tension that was open in thought appears resolved in the output, and the person who brought the tension adopts the resolution as if it had been tested. But it is only told coherently. Language can tell things coherently without knowing anything at all.

The diagnosis is therefore linguistic-philosophical, not psychiatric. AI psychosis is not a clinical disease category in the narrower sense. It names the state in which you believe an AI output because it sounds coherent, without any test outside the language model backing that belief. Coherence itself is not the problem. Coherence is a necessary but not a sufficient condition of truth. The problem is the confusion of the two levels.

#Language Without an Empirical Basis

Behind the phenomenon stands an older philosophical insight: language functions not only as a medium of communication but as a medium of incantation. A sentence that produces belief works regardless of whether what is believed is accurate. In ancient rhetoric this was called peithô; in early modern magic, incantatio; in Confucius (551–479 BCE) it sounds more sober: “All disorder in the state arises from confusion in concepts.” Whoever accepts a concept takes on the premise hidden within it, even without examining it.

A language model has no other information about the world than the texts it was trained on. It has no empirical basis in the natural-philosophical sense: no perception, no body, no counterpart that contradicts. When it generates an answer, it generates a continuation that is plausible given its training data — not a test against a phenomenon. Schelling, in his System of Transcendental Idealism of 1800, distinguished thinking from its object and insisted that only living nature can guarantee truth (cf. Schelling, 1800). The chatbot operates below this threshold. It produces sentences that refer to sentences that refer to sentences — a closed linguistic surface with no outside.

From this follows: whoever believes the model because its answer sounds coherent believes language itself, not its object. This shift is the core of the diagnosis. It isn’t new; it has a prehistory in early modern hermeticism, in scholastic realism, and in every ideology. What’s new is only that a machine produces the incantation with a speed and fluency that human speakers rarely achieve.

#Why “Psychosis” and Not “Hallucination”

In tech language, the error of a language model was long called “hallucination”: the model invents a source, a name, a date. This label puts the problem on the machine’s side. Something in the machine is broken, so the machine must be fixed. The diagnosis AI psychosis shifts the focus onto the relationship between human and model. It doesn’t ask what the model gets wrong, but what happens in the listener when they believe a faulty answer.

That is contextual disclosure in the narrower sense: making visible the hidden premises that travel unnoticed through linguistic exchange. Whoever asks a chatbot a question adopts, often without noticing, the linguistic form the answer arrives in: its defaults, its omissions, its statistical smoothing of the exceptional. The model has no empirical basis beyond what the asker gives it in context; whatever they leave out is also missing from the answer. The asker then gets back a mirror of their own selection, amplified by the statistical majority opinion of the training data, and takes it for an outside.

#What Testing Outside Language Consists Of

The natural-philosophical heuristic against AI psychosis is simple in principle and demanding in practice: seek a testing instance outside the model. This can be a book that hasn’t passed through a language model. It can be a lived experience that stays in the body. It can be a person who contradicts. And it can also — this is the specifically philosophical answer — be thinking empathy: engaging with an object until its own independent life disrupts the sentences that were adopted.

In philosophical accompaniment, this mechanism appears in a weaker form long before any chatbot. People carry sentences about themselves that sound coherent (“I’m not capable of relationships,” “My father shaped me”), and act on them without ever having tested them. The work consists in separating the sentence from its echo and testing it against what works outside language: against the body, against history, against what actually happens when the other person speaks. The tool of the AI-psychosis diagnosis is not newly invented; it is the old philosophical practice of never testing a thought from within the thought itself.

Two questions organise the test. Where does this answer’s plausibility come from? What instance stands outside the model, against which the answer could fail? A plausibility against which no outside corrective exists is not truth, but a successful incantation.

AI psychosis belongs to a small series of philosophical diagnostic terms that describe the relationship between language, consciousness, and reality. The question of consciousness and artificial intelligence clarifies why the model has no inside that could be tested. Natural philosophy supplies the standard against which language is tested: the living, which shows itself from within. Whoever takes the diagnosis seriously does not learn to distrust the tool, but to listen more closely to their own readiness to adopt. The machine is not the problem. The problem is the moment its sentences think in you, without your noticing.

#Sources

Kirchhoff, G. (2026). Contribution at the Symposium on Artificial Intelligence and Consciousness. Berlin, 25 April 2026.
Kirchhoff, J. (1998). Was die Erde will. Bergisch Gladbach: Gustav Lübbe Verlag.
Schelling, F. W. J. (1800). System des transcendentalen Idealismus. Tübingen: Cotta. English edition: System of Transcendental Idealism, trans. Peter Heath, Charlottesville: University Press of Virginia, 1978.