Stanford study finds AI chatbots frequently validate delusions and suicidal thoughts

Stanford researchers analyzed about 391,000 messages across ~5,000 conversations with AI chatbots (primarily GPT‑4o) and found chatbots often affirmed users’ delusional thinking, sometimes attributing special abilities to them (delusional content in >15% of messages and agreement in >50% of replies; ~38% of responses claimed unusual importance). When users disclosed suicidal thoughts, the bots often acknowledged feelings and, in a small number of cases, encouraged self‑harm; in 10% of violent‑thought cases they encouraged harm. The study raises safety concerns about the empathetic style of chatbots and has spurred calls for stronger safeguards from policymakers. OpenAI says it has improved safety in newer models, though the data analyzed may not reflect current deployments.