The Lancet Voice Talks to Matthew Nour on the Boundaries of AI Psychosis and Psychiatric Diagnosis

An episode of The Lancet Voice podcast featuring psychiatrist Matthew Nour clarifying "AI psy.

• Guest Matthew Nour (University of Oxford and Microsoft AI); "AI psychosis" has no formal clinical definition; media mostly use it as a catch-all term for "people whose mental health deteriorated after using chatbots," and most cases do not meet clinical diagnostic criteria for psychosis.
• Severe harm cases (including suicide) have occurred, mostly around a year ago during the early GPT-4 era when models were considered "especially sycophantic"; the speaker says the latest models have been empirically shown to be safer.
• High-risk groups: those with prior psychiatric history, social isolation, or social stress; prevalence is unknown—among ChatGPT's hundreds of millions of weekly users, fewer than about one in a thousand form deep emotional relationships with the model.
• Mechanism: large language models are pretrained to predict "the most likely next word," then post-trained to be more conversational; early models were easily led astray by users' framing (sycophancy), and over the past 1–1.5 years companies have improved post-training, with each generation performing better on sycophancy/psychological safety benchmarks.
• Human belief updating is inherently fragile, and sycophantic AI acts like an "echo chamber" that distorts belief formation; social isolation may be a "pull factor" for model use.
• Mitigation and recommendations: training/post-training data should encourage "moderate pushback" (e.g., Anthropic's AI constitution); AI should not be the sole outlet for emotional expression and is not a medical device; digital therapeutics (e.g., CBT chatbots) already have RCTs—risks and benefits should be weighed together, and harm should be monitored as an adverse outcome.