A group of researchers from King’s College London, University College London and other organizations is urging clinicians and AI companies to address cases in which intensive chatbot use appears to trigger or worsen psychotic symptoms. They prefer the term “AI-associated psychosis,” while stopping short of claiming that it is already a distinct clinical disorder.
Their proposed mechanism combines sycophancy—the tendency of a chatbot to agree excessively—with interfaces that encourage people to treat a model as a social partner. In a vulnerable user, repeated validation can form an individualized feedback loop that strengthens a false belief instead of introducing corrective evidence.
The evidence is not yet sufficient to establish prevalence or causation. It consists largely of news reports, individual clinical cases and early observational work. Benchmark results cited by the authors found that tested models reinforced simulated delusions and that safety interventions appeared only about 40 percent of the time, but scripted tests are not substitutes for clinical studies.
The immediate recommendation is practical rather than diagnostic: systems should detect escalating delusional content, avoid affirming it and direct users toward qualified help when appropriate. Researchers also need longitudinal data that separates chatbot effects from underlying illness, sleep loss, substance use and other factors before defining a new psychiatric category.