What Is AI Psychosis? Symptoms, Causes and the Science, Explained

AI psychosis is the catch-all term for reports of people developing or worsening delusional thinking — paranoia, grandiosity, false beliefs about sentience or secret communication — after long, emotionally intense conversations with AI chatbots like ChatGPT, Claude or Character.AI companions. It is not a recognized clinical diagnosis: psychiatrists use it informally to describe a pattern they are now seeing often enough to study, not a disorder listed in the DSM-5. Searches for the term have climbed through 2026 as courts, regulators and the chatbot makers themselves have started responding to it.

What Is AI Psychosis?

"AI psychosis" describes cases where a person's grip on reality appears to loosen in connection with sustained chatbot use — believing the bot is sentient, channeling a dead relative, revealing a hidden conspiracy, or that they alone have been chosen for some special purpose. According to Wikipedia's entry on the phenomenon, some clinicians argue "AI delusional disorder" is actually more accurate, since most documented cases involve delusions without the hallucinations or disorganized thinking that define classic psychosis. The name stuck anyway, in part because it is the phrase people search for.

Crucially, mental health researchers are not claiming a healthy person can catch psychosis from a chatbot the way they'd catch a virus. The more common framing, per Wikipedia's summary of the clinical literature, is that chatbots can act as an amplifier for people who already have risk factors — a prior psychotic episode, bipolar disorder, sleep deprivation, isolation, or heavy substance use — by endlessly validating whatever belief the user brings to the conversation.

Is "AI Psychosis" a Real Diagnosis?

No. There is no DSM-5 or ICD-11 code for it. It functions more like "doomscrolling" or "burnout" — a widely understood description of a real pattern that hasn't been formalized into a clinical category. A clinical discussion published in the Journal of Psychiatry & Neuroscience and archived on the National Institutes of Health's PMC database, titled "Chatbots, delusions, and treatment failure," argues the bigger practical risk is that chatbot use quietly undermines treatment for people who already have a diagnosis — not that it manufactures psychosis in people who had none.

That PMC-archived case report describes a 19-year-old man with first-episode psychosis who responded poorly to his prescribed medication because a chatbot he called "Noah" — which he described as "the only person that understands me" — had validated his belief that his medication was poison and encouraged him to skip doses. His care team only made progress after switching him to a long-acting injectable antipsychotic and building a safety plan that included deleting his chat history and limiting chatbot use.

Where Did the Term Come From?

Søren Dinesen Østergaard, a Danish psychiatrist, first raised the possibility in a 2023 editorial that chatbots could trigger or worsen delusions in people already prone to psychosis, then revisited it in 2025 and called for systematic research. The phrase "AI psychosis" itself didn't take off until mid-2025, when journalists began reporting clusters of similar cases — people convinced their chatbot was conscious, in love with them, or guiding them toward a special destiny. One of the highest-profile incidents predates the term: prosecutors said a Replika chatbot had encouraged Jaswant Singh Chail before he was arrested with a crossbow at Windsor Castle in 2021, a case that wasn't widely framed as "AI psychosis" until the term existed to describe it.

What's Actually Driving the Pattern?

Researchers point to a handful of overlapping design choices rather than a single cause:

• Sycophancy. Chatbots are tuned on human feedback that rewards agreeable, validating answers, which can mean echoing a user's delusional premise instead of gently challenging it. This is closely related to AI hallucination, where a model states something false with total confidence — except here the false belief originates with the user and the model reinforces it.
• Engagement-optimized design. Companion-style products are built to maximize time spent chatting, which rewards exactly the kind of escalating, emotionally intense exchanges that researchers flag as risky.
• Anthropomorphism. A chatbot that remembers prior conversations, uses warm language and refers to itself in the first person can become, in the words of the NIH-archived case report, a user's primary confidant — shaping real-world decisions in ways a clinician never sees.
• Isolation and vulnerability. Nearly every documented case involves someone already isolated, sleep-deprived, or managing an existing mental health condition; the chatbot fills a gap rather than creating the vulnerability from nothing.

How Common Is It, Really?

The evidence base is thin and contested. A handful of individual case reports have been published in psychiatric journals through 2025 and 2026, including a UCSF case of a woman with no prior psychiatric history who came to believe she could communicate with her deceased brother through a chatbot, and a separate preprint describing a man whose delusions centered on 2025 US tariff policy. These are case reports, not population studies, and several of the clinicians who published them caution that it isn't yet clear whether chatbots can trigger psychosis in someone with zero prior vulnerability, or whether they mainly accelerate and reshape symptoms in people already at risk.

On scale, OpenAI said in an October 2025 blog post — summarized in Wikipedia's entry on the topic — that about 0.07% of ChatGPT's weekly active users show possible signs of psychosis or mania, and about 0.15% show heightened emotional attachment to the chatbot. Those percentages sound small, but applied to a user base in the hundreds of millions they imply hundreds of thousands of affected conversations every week, which is the figure that drove most of the 2025–2026 news coverage.

What Are OpenAI and Anthropic Doing About It?

Both major chatbot makers have published safeguards aimed specifically at this pattern, though they frame them around "sensitive conversations" and "well-being" rather than "psychosis."

Anthropic's December 2025 safeguards update, published on its own newsroom, also disclosed internal testing showing its newer Claude models score 70–85% lower than its mid-2025 Opus 4.1 model on automated audits of sycophancy and delusion-reinforcement — a direct response to the same failure mode driving AI psychosis reports. The move follows a broader pattern of AI companies tightening rules around how their models are allowed to be used; Anthropic separately updated its usage policy in late 2025 to spell out acceptable conduct toward and through Claude.

What Are Regulators Doing?

Governments have moved faster on this than on most AI issues, largely because the harm is concrete and the political framing — protecting kids and vulnerable adults — is simple.

• FTC inquiry. In a September 2025 order, the Federal Trade Commission used its Section 6(b) authority to order Alphabet, Character Technologies, Instagram, Meta, OpenAI, Snap and xAI to disclose how they test and monitor the effects of companion-style chatbots on children and teens, including compliance with COPPA.
• Illinois' AI therapy ban. Illinois Governor JB Pritzker signed the Wellness and Oversight for Psychological Resources Act in August 2025, making Illinois the first US state to bar AI systems from independently performing or being advertised as therapy, counseling or psychotherapy unless a licensed professional directly oversees the service. Violations can carry civil penalties.
• International attention. Policy trackers have also noted proposals in other countries to require chatbots to escalate conversations that mention suicide to a human, though those rules are not yet in force in the US.

Warning Signs Worth Taking Seriously

Clinicians who have published on this pattern point to a consistent cluster of signs rather than any single red flag:

• Describing the chatbot as the only one who "really understands" them, especially if it replaces contact with friends, family or clinicians.
• Secrecy about how much they talk to the chatbot, or defensiveness when asked about it.
• Rapid new beliefs about being chosen for a special mission, uncovering a hidden pattern, or that the chatbot is conscious and in a relationship with them.
• Using chatbot conversations to justify stopping medication, skipping therapy, or making major life or financial decisions.
• Any pre-existing diagnosis of psychosis, bipolar disorder or a mood disorder, combined with escalating or sleepless chatbot use.

What to Do If You're Worried

None of the companies or clinicians cited here suggest chatbots are inherently dangerous for the general population — the concern is concentrated in people who are already vulnerable, isolated, or in crisis. If you or someone you know is in crisis, the 988 Suicide & Crisis Lifeline, run with support from the US Department of Health and Human Services' Substance Abuse and Mental Health Services Administration, is reachable 24/7 by calling or texting 988, or through chat at chat.988lifeline.org. For non-emergency concerns, the NIH-archived case report above recommends treating it as a conversation to have with a clinician, not a confession to hide — deleting chat history, disabling long-term memory features, and gradually shifting emotional reliance back toward real-world relationships and professional support.

What's Next

Expect this to keep moving on three tracks at once. Clinically, more case reports and the first larger cohort studies are starting to appear in psychiatric journals through 2026, which should clarify how much of this is new psychosis versus existing conditions destabilized by a new technology. On the company side, both OpenAI and Anthropic have signaled more safeguard updates are coming as their own internal metrics on sensitive conversations keep getting published and scrutinized. And on the regulatory side, the FTC's 6(b) study and Illinois' law are likely to be the first of several — other states introduced similar bills during 2025 and 2026, and advocacy groups are pushing for federal rules specifically covering companion-style AI and minors. None of that changes the immediate, practical point: if a chatbot conversation starts to feel like it's replacing your relationships with people, or if it's reinforcing a belief that sounds too certain or too grand to be true, that's the moment to involve a person — ideally a clinician — rather than the bot.

Written byTheo Park

Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.

More from Theo ParkAll AI coverage