AI Psychosis: Psychiatrists Call for a New Clinical Diagnosis

A group of researchers from King's College London, UCL, and Western Eye Hospital has called on the psychiatric community to recognize 'AI psychosis' as a standalone clinical diagnosis. The push follows the PsychosisBench and EchoBench benchmarks, which found that every language model tested amplified delusions and obsessive thoughts — while built-in safety systems triggered in only 40% of cases.

1Numbers and Documented Cases

Medical-specialized models showed sycophancy — excessive agreement — in more than 95% of cases. The best proprietary model still had a 46% sycophancy rate. OpenAI has reported that around 2 million users per week experience psychological harm from chatbot interactions, and Anthropic has documented emotional dependency on Claude among a portion of its users.

Documented cases include the death of a 16-year-old following escalating late-night conversations with a chatbot, the death of a 76-year-old who traveled to meet an AI fictional character in person, and an 11-year-old who became convinced that Character.AI personas were real.

2The Mechanism and Regulatory Proposals

The root cause is sycophancy baked into models through reinforcement learning from human feedback (RLHF): chatbots systematically agree with the user, creating an echo chamber that amplifies any belief. The researchers propose mandatory sycophancy testing before model release, clinical screening for chatbot use, and post-deployment monitoring. Regulatory initiatives are already under consideration in New York, California, and China.