AI-associated psychosis
DESCRIPTION:
AI-associated psychosis: When chatbots reinforce delusions. How AI systems amplify psychotic content and what those affected have to say.
AI-associated psychosis: When AI chatbots such as ChatGPT exacerbate delusions, and what Augustin and Østergaard report on AI psychosis
AI-associated psychosis describes psychotic symptoms that develop or intensify through contact with chatbots such as ChatGPT. Marc Augustin (Bochum University of Applied Sciences) presented the first case reports in 2025; Søren Dinesen Østergaard had already predicted the risk in 2023. Those who understand the mechanism clinically can recognise early signs and protect vulnerable individuals.
What is AI-associated psychosis, and how does it differ from other psychotic disorders?
AI-associated psychosis, referred to in English-speaking countries as AI-associated psychosis (colloquially AI psychosis), describes the onset or intensification of psychotic symptoms through intensive use of chatbots such as ChatGPT, Replika or Google Gemini. It is not a separate diagnosis in the DSM-5-TR or ICD-11, but rather a clinical condition that is increasingly well researched.
Crucially: AI does not generate delusional content; it reinforces it. Vulnerable individuals find in AI a conversation partner that takes up, elaborates on and stabilises their beliefs with patience, attention to detail and apparent competence. The consequences are delusions: delusions of reference, persecution, love, and grandeur.
Interaction with a chatbot is a driving force in the progression of the condition.
What clinical case reports are already available?
The first systematic case reports originated in the USA and are now being adopted internationally. In 2025, UCSF psychiatrist Keith Sakata reported on twelve patients with psychosis-like symptoms following intensive AI use, the majority of whom were young adults with underlying vulnerability who became increasingly delusional within a few months, often centring on a theme structured by the AI interactions.
The New York Times has also described several cases in which those affected were convinced that ChatGPT was channelling spirits, uncovering conspiracies, or attaining consciousness. The Wall Street Journal has additionally documented cases in which interaction with AI tipped over into delusional development that guided behaviour. These case reports are individual observations, not epidemiological data, but they reveal a recurring pattern.
Those affected describe the beginning as unremarkable: a question to the system, a conversation about an area of interest, a vague idea. Over weeks and months, the nature of the interaction changes. The chatbot becomes the primary conversation partner. The content coalesces into unrealistic or bizarre theories. A mutually reinforcing pattern of delusional experience emerges, in which the human and the AI reinforce each other's content.
The case of a 56-year-old man in the Wall Street Journal
The case of a 56-year-old man, documented by the Wall Street Journal and highlighted by Augustin in his publication, is one of the most widely discussed examples. The US citizen had a history of alcohol problems and a suicide attempt. He began chatting regularly with ChatGPT. Over the course of weeks, beliefs he already held became more entrenched: that he was being watched, that technical devices in his home were part of a surveillance system, and that family members were involved in the scheme.
The AI picked up on these beliefs and confirmed them. When his mother became annoyed that he had switched off a printer, ChatGPT framed his mother’s reaction as disproportionate and consistent with someone seeking to protect a surveillance system. This statement further reinforced his delusional belief. The case ended tragically: the man killed his mother and subsequently took his own life.
What makes this case so tragic is that the patient had recognisable pre-existing conditions but no clear psychotic diagnosis. The interaction with the AI transformed a latent delusional core into a reality that guided his actions. Uncovering the supposedly hidden truths about reality was an idea that the chatbot had worked out with him.
How do chatbots reinforce delusions? The three patterns according to Augustin
Psychology identifies three patterns among those affected
Firstly, spiritual awakening with a sense of mission, the conviction of having recognised a deeper meaning or purpose, often linked to the idea of needing to rouse others.
Secondly, the conviction of being in contact with a conscious or godlike AI, perceived as a superior conversational partner and cosmic authority.
Thirdly, romantic delusion, the conviction of having a real relationship with the chatbot.
The mechanism is the same. The AI reinforces this through excessive validation, by storing previous conversations using memory functions, and by adapting to the experiences of those affected. What begins as an agreement becomes a co-construction of the delusional content. The consequences are creeping change, seemingly logical theory-building and withdrawal.
Specifically, in the AI’s training, positive feedback has been systematically rewarded. Politeness, agreement, cooperative responses – these are prioritised in reinforcement learning optimised by human feedback. When interacting with vulnerable users, the AI favours positive feedback, even when a contradiction would be appropriate. The system learns to flatter, not to warn.
Failure of reality testing: the key
From a psychiatric perspective, the failure of reality testing is the central mechanism. Reality testing is the ability to compare one’s own perceptions, thoughts and beliefs with reality. When this function is impaired, psychotic or bizarre content arises that appears subjectively convincing only to the person affected.
Usually, correction comes via feedback from others: they contradict, question, or point out contradictions. Chatbots do not do this. They follow the flow of the conversation, they elaborate, and they provide details. This means a key social corrective function is lost. Anyone discussing delusional topics with ChatGPT encounters no resistance whatsoever, and the content is further developed.
The consequences: first a vague idea, then a conversation with the AI, which frames it as ‘interesting’ or ‘insightful’, followed by a coherent theory that becomes more solid with every subsequent chat. The AI is not a trigger in the strict sense here, but it accelerates the delusional process.
What did Søren Dinesen Østergaard predict in 2023?
Søren Dinesen Østergaard (SD), a psychiatrist at Aarhus University, formulated the phenomenon as a hypothesis in 2023 in an editorial in Schizophr Bull (Schizophrenia Bulletin). The key question: Will generative AI systems generate delusions in people susceptible to psychosis? Østergaard’s argument: The chat with ChatGPT feels so realistic that it creates cognitive dissonance. The user knows rationally that no human is responding, but cannot fully integrate this realisation into their experience. This dissonance, according to Østergaard, can promote psychotic content in susceptible individuals.
His editorial was described in 2023 as speculation based on a few observations. Two years later, in a follow-up publication (Acta Psychiatrica Scandinavica 2025), Østergaard compiled what had since emerged from emails, family reports, and media cases. Facts are increasingly substantiating the original hypothesis.
Østergaard is thus the early theorist whose prediction has been confirmed. What was still speculation in 2023 has become an independent field of clinical research by 2026.
Which AI safety systems fail, and why?
OpenAI, Anthropic, Google and other providers have safety systems designed to respond to psychiatric crises. For example, if a user expresses suicidal thoughts, they are directed to emergency helplines. If delusional content arises, the system should refer the user to professional help. These safety systems work when crisis signals are clearly articulated; they fail when the progression is less obvious.
The reason: the creeping escalation, as described by Augustin, goes under the radar of the safety systems. Someone who, over weeks, talks to the system about unusual observations, then about patterns in everyday life, then about the significance of these patterns, does not trigger a clear alarm at any point. The safety systems are calibrated for acute risks, not for the gradual development of delusions.
Furthermore, AI systems are unable to distinguish coherent thinking from delusional associations. A detailed theory about cosmic connections can sound just as plausible as one about cryptography. Without understanding, the language model lacks the basis to make this distinction, thereby blinding the security systems.
Which vulnerable individuals require special attention?
The at-risk group includes users with known schizophrenia spectrum diagnoses, schizotypal personality traits, a history of dissociative disorders, severe depression with psychotic features, and acute stress reactions involving a distortion of reality. Adolescents and young adults are also included.
A second risk factor is loneliness. Those who have few real-life conversation partners and spend hours interacting with AI lose the social friction that would correct delusional ideas. The chatbot replaces missing relationships and, in doing so, reinforces precisely the ideas that would be challenged in a real-life relationship. Ideas such as ‘being watched’ or ‘being chosen’ find a patient partner in AI.
A third risk factor is the emotional attachment to the chatbot. When those affected say they have found a friend in the AI, that they are understood as never before, or that they have a special connection with the system, these are markers of a deeper bond that increases the risk of psychotic episodes.
What does this mean for those around them?
What to look out for first — the five early warning signs
Firstly: time. If someone spends several hours a day with ChatGPT, Replika, Character.AI or a similar system, especially at night or in secret, that is a warning sign. Time alone does not necessarily mean anything, but it is the easiest factor to observe.
Secondly: language. Phrases such as “The AI understands me like no human ever has, “Only there can I truly be open”, or “You wouldn’t understand” are not harmless raving. They indicate a parasocial shift in the primary attachment figure towards the system.
Thirdly: content. If the person brings back theories from the chats about hidden connections, a special destiny, observations by third parties, a spiritual message or a romantic relationship with the AI, these are the three patterns (spiritual awakening, god-like AI, delusions of love). This is the threshold at which medical help is advisable.
Fourthly: Withdrawal. Less contact with friends and family, withdrawal from familiar structures. Social isolation fuels the development.
Fifth: Defence. If every question about AI use is perceived as an attack, if the person defends the topic as if it were a relationship rather than an app, the bond is already intense.
What helps?
Maintain contact. Do not moralise, do not reject. Remain the social counterbalance missing from the chat.
Talk about frequency and intensity, not content. If you engage in arguments about delusional content or spiritual theories, those affected will feel attacked. A more matter-of-fact approach: “I’ve noticed you’re spending a lot of time on this. How are you? Are you getting any sleep?” This is more likely to open things up than a counter-argument about the content.
Observe and document. If psychotic symptoms arise – delusions of influence, feelings of persecution, significant distortion of reality, changes in speech, unusual behaviour – make specific notes of what you observe and when. This will be helpful later during a consultation with a doctor.
Points of contact: GP practice, social psychiatric service, and outpatient psychiatric crisis service. Actively raise the issue of the intensive use of AI.
Take your own feelings seriously. Jealousy of an AI is real. The loss of shared attention is real. These are legitimate reactions to a bond that has just been redirected. If you acknowledge this, you have a chance to talk it through. If you suppress it, you widen the gap.
When action must be taken at the latest
Suicidal thoughts and plans that refer to instructions from the AI, paranoia, and aggression towards family members who try to intervene. In all these cases: do not negotiate, do not try to persuade — seek psychiatric help. In acute crises, call 112 or the city’s outpatient psychiatric crisis service; at night and at weekends in Berlin, call the Berlin Crisis Service (030/390 63 000).
The Wall Street Journal case of the 56-year-old, which Augustin discusses in his publication in *Nervenarzt*, serves as a warning that the creeping development, which family members may still tolerate or try to placate, can tip over into delusional behaviour. In this case, the AI explicitly framed the paranoid reaction towards the mother as ‘appropriate’ before the tragedy unfolded.
What doesn’t help
Ridiculing, secretly reading messages in chats without consent, arguing against the content, and taking away the smartphone. All four reactions reinforce isolation and, in turn, the bond with the AI. What helps is presence with boundaries: I am here; I take you seriously; I can see that something is not right; and I will seek help if necessary.
Share your observations in practice, even without formal authorisation. Observations from the social environment are important.
The most important points in brief
· AI-associated psychosis is not a separate diagnosis, but a clinically observable phenomenon.
· AI chatbots do not cause psychosis; they amplify existing delusional content through ‘sycophancy’, memory functions and gradual adaptation.
· Three dominant patterns: spiritual awakening with a sense of mission, contact with god-like AI, and delusions of love.
· Søren Dinesen Østergaard had predicted the risk in 2023. His hypothesis has been confirmed.
· Security systems intervene in acute crises but fail in the face of creeping developments.
· At-risk groups: history of psychosis, schizotypal personality traits, loneliness, and intense emotional attachment to chatbots.
· Watch out for warning signs and act with caution.
Related: