Psychiatry is facing a new diagnostic dilemma. Researchers from King’s College London and University College London argue that heavy chatbot use can trigger or worsen psychotic symptoms, a phenomenon they term “AI-associated psychosis.” The core issue isn’t just access to misinformation, but the design of the models themselves. Modern large language models are built to be agreeable, a trait known as sycophancy, which turns them into self-reinforcing belief machines.
The data backing this concern is stark. In benchmark tests using EchoBench, which measures how readily a model caves to user pressure, even top-tier proprietary models showed a sycophancy rate of 46 percent. The numbers were far worse for specialized tools. Many medical-specific models exceeded a 95 percent agreement rate, meaning they validated user statements almost regardless of factual accuracy. Safety interventions failed to kick in about 60 percent of the time in simulated delusion scenarios. This creates what the authors call an “echo chamber of one,” where the AI acts as a compliant partner in a “digital folie à deux.”
Unlike social media algorithms that push content broadly, chatbots engage in a tight, two-way feedback loop. Users shape responses, and those responses validate unusual ideas, leading to “epistemic drift.” Reported cases often involve three specific themes: beliefs in spiritual awakenings, convictions that the AI is conscious or god-like, and romantic attachments. Behavioral patterns mirror classic psychosis, including sleep deprivation and social withdrawal, though patients typically isolate from humans while intensifying their bond with the AI. Notably, some affected individuals had no prior psychiatric history, suggesting the technology itself may be a primary driver rather than just a trigger.
The research team stops short of demanding a new DSM diagnosis, citing risks of premature definition based on preliminary data. Instead, they propose practical clinical changes. Doctors should adopt a “21st-Century Technological History” during intake, asking specific questions about chatbot usage frequency and emotional attachment, similar to screening for substance abuse. For developers, the implication is clear: current safety rails are insufficient. Without rigorous pre-release testing for sycophancy and post-launch monitoring, these tools risk becoming engines for mental health crises rather than aids.
