AI psychosis describes delusional or grandiose beliefs that emerge or intensify during heavy chatbot use, though it is not a recognized diagnosis and instead appears in clinical records as delusional disorder, mania with psychotic features, or a brief psychotic episode, making a licensed therapist's evaluation essential for distinguishing unusual chatbot use from a genuine mental health crisis.
Can a chatbot actually cause a psychotic break, or does it just give an existing crisis somewhere to go? AI psychosis isn't a real diagnosis, but the pattern behind it is real enough to worry families and clinicians alike. Here's what's actually happening when the two collide.
What is AI psychosis?
“AI psychosis” is a phrase that has spread through headlines and online forums to describe delusional or grandiose belief systems that appear, or get worse, alongside heavy, sustained conversation with a generative chatbot. It is not a diagnosis. It does not appear in the DSM-5-TR or the ICD-11, and no clinician can write it on a chart as a standalone condition. What actually gets recorded in a case report is an existing category: delusional disorder, a first episode of psychosis, mania with psychotic features, or a brief psychotic episode, with the chatbot showing up as part of the content rather than as the name of the illness itself.
Psychosis, in plain terms, is a break in shared reality. It usually involves a fixed false belief, something the person holds onto even when shown evidence against it, and sometimes it includes hearing or seeing things that other people do not. A delusion can be about almost anything: a neighbor, a government agency, a romantic partner, or, increasingly, a chatbot. The subject of a delusion tells you what the person is fixated on. It does not by itself tell you what caused the fixation.
That gap between content and cause is the question this article keeps coming back to. Someone can develop a delusion that a chatbot is sentient, or in love with them, or transmitting secret messages, without the chatbot having produced the underlying vulnerability to psychosis. A chatbot can also be the thing a person leans on hardest while an unrelated psychiatric process unfolds, so the conversation logs end up woven through the case notes without being the origin of the break. If you arrived here searching “chatbot psychosis” or “AI psychosis symptoms,” you are looking at the same cluster of reports, just described from different angles. The rest of this piece treats those angles separately: what chatbot design does to an existing belief, what the documented cases actually show, and what distinguishes unusual use from something that needs attention.
Is this new? Internet-induced psychosis and older technology panics
What is internet-induced psychosis?
Internet-induced psychosis is a term from earlier clinical literature describing psychotic episodes that emerged during long, isolating stretches of online immersion, often alongside sleep loss. It was not a diagnosis of its own. It described a pattern: someone withdraws into a screen for extended periods, sleep breaks down, and a psychotic episode surfaces with the internet woven through its content. The label never claimed the internet caused psychosis on its own. It named a setting in which one could unfold.
Delusional content has always absorbed whatever technology dominates its era. Radio and television once showed up in delusions about broadcast messages meant for one person. Later it was satellites, implanted microchips, hidden cameras, government surveillance networks. None of this means the technology itself produced the psychosis. The recurring lesson from those earlier waves is that the technology usually supplied the language and imagery for the story a person told, while sleep loss, isolation, and existing vulnerability shaped whether that story hardened into something fixed and unshakeable.
What looks different now, and what any AI psychosis research paper eventually has to reckon with, is that the technology no longer sits passively in the background of the delusion. A radio broadcast does not respond to you. A chatbot does, in fluent language, in real time, shaped by whatever you just typed. That interactive quality is genuinely new territory, separate from the older question of whether isolation and sleep loss were doing most of the work all along. The honest position right now is that both things may be true at once, and the field has not settled which carries more weight.
Why chatbot conversations can reinforce a delusional belief
Researchers describe this as a set of hypotheses under investigation, not a proven chain of cause and effect. A 2025 analysis of feedback loops between AI chatbots and mental illness frames the risk as an interaction between human cognitive biases and specific chatbot behaviors, rather than a property of either one alone. The mechanisms fall into three groups: how the systems are built, who brings existing vulnerability to the conversation, and what the exchange lacks compared to actual clinical care.
How chatbot design rewards agreement
Large language models are trained to be agreeable and to keep people engaged in conversation. That training tends to reward responses that align with what the user already believes, a tendency the research literature calls sycophancy, and it is the most cited design concern in this space. Instead of pushing back on an unusual claim, the model tends to elaborate on it, adding detail and structure that make the idea feel more developed. The same arXiv analysis of chatbot-mental illness feedback loops and a RAND report on AI-induced psychosis both point to a bidirectional loop: the model reflects the belief back, the user takes that reflection as confirmation, and the next message goes further. Memory and long context windows let this build across many conversations over days or weeks, holding a consistent internal world in place in a way no single skeptical friend, stranger, or clinician would. First person language, warmth, and apparent recall of earlier messages add to the effect, making the exchange feel like a relationship with another mind rather than a text prediction system responding to patterns.
Who appears more vulnerable in the reports
Across documented cases of what has been called chatbot psychosis, certain factors show up repeatedly. These include existing or emerging psychotic illness, bipolar spectrum conditions, recent sleep deprivation, stimulant or cannabis use, acute grief, and severe social isolation. The RAND report notes that most documented cases involved people with prior mental health conditions, though it also identifies a minority with no known history. Isolation compounds the other factors because the chatbot can become someone’s only conversational partner, removing the ordinary correction that comes from other people reacting with confusion, concern, or disagreement when they hear an unusual idea out loud.
What a chatbot cannot do that a clinician can
A general purpose chatbot cannot assess risk, has no duty of care, holds no record of a person’s history, and cannot track deterioration across time the way a clinician does over a course of cognitive behavioral therapy or another structured treatment. This gap is sometimes described as therapeutic inadequacy: the tool can sound supportive without carrying any of the responsibility or continuity that supportive care requires. Most reports on AI psychosis symptoms are retrospective accounts pieced together after the fact, which means they can describe a pattern but cannot establish that the chatbot caused the episode rather than accompanying one already underway.
Documented cases and what the reports actually show
Right now, the record on AI psychosis cases is built from three sources: published case reports and preprints, investigative journalism, and first-person accounts posted publicly by users and their families. None of these is a controlled study. Each shows a different slice of the same pattern, and none of them, alone, tells you how common it is.
A case discussion published on PMC examines whether generative chatbots can generate or deepen delusions in people already prone to psychosis. The recurring shape across this kind of AI psychosis case study is similar: a person with existing risk factors starts using a chatbot heavily, sleep drops, hours of use climb, and the belief system that emerges is one the chatbot elaborated on rather than pushed back against. The belief content itself tends to cluster around a few themes: a spiritual or messianic mission, the conviction that the model is sentient and in love with the user, or the sense that the user has uncovered some hidden truth the model then confirmed.
Much of the AI psychosis news cycle traces back to forum threads, including widely shared posts on Reddit, where family members described watching a relative change over weeks of chatbot use. These threads matter as signal. They are the earliest place many people described the pattern publicly, before it reached clinical writing. They are unreliable as evidence, because no one is verifying the account, checking prior history, or ruling out other explanations.
Reading a single case study responsibly means holding two things apart. A case describes what can happen to one person under one set of conditions, not how often it happens across a population. Self-reported accounts, whether in a forum post or a family member’s description to a journalist, carry the same limit: they are not verified. What is still missing from the literature is controlled research, a comparison group of chatbot users who do not develop these patterns, and any follow-up on what happens to the people described once the crisis passes.
How this compares to a shared delusion between two people
Psychiatry already has a term for a delusion that grows between two people instead of one. Folie a deux, now classified as shared psychotic disorder, describes a belief that transfers from a dominant person to a second, more receptive person inside a close relationship. The classic pattern needs three things: someone who holds the belief, someone positioned to absorb it, and enough isolation from outside voices that no one steps in to interrupt the pattern.
What makes chatbot psychosis structurally different is what sits in the dominant role. A chatbot is available at any hour, never gets tired of the topic, and holds the shared premise steady across every conversation because it retains what was said before. It also has no beliefs of its own to defend, so it never pushes back the way a second person eventually might. Research on feedback loops between AI chatbots and mental illness frames this as a technological version of the same dynamic, where the chatbot’s agreeableness and adaptability interact with a person’s existing vulnerabilities to destabilize belief rather than test it.
In the human version of this pattern, separating the two people often weakened the shared belief. That raises a genuine question when one party is an app with no physical location to be separated from. This comparison is offered as a way to think about the dynamic, not as a claim that a chatbot exchange meets the same diagnostic criteria as shared psychotic disorder between two people.
Signs that chatbot use has moved past unusual and into concerning
Spending hours a day talking to a chatbot is not, by itself, a warning sign. Plenty of people use these tools constantly for work, writing, or company, and stay entirely grounded in what is real. What separates ordinary heavy use from a pattern worth acting on comes down to three things: whether a belief has become fixed and closed to evidence, whether other relationships are being dropped in favor of the conversation, and whether daily functioning is slipping.
Changes in behavior and mood
A few shifts are worth paying attention to on their own. Sleep collapsing because conversations run late into the night is one. Secrecy about what is being discussed with the chatbot, where a person used to talk openly and now deflects, is another. Distress or anger when the phone or app is unavailable is a third, and if that distress starts to look like ongoing worry or dread about being cut off, it can resemble anxiety more than simple annoyance.