Overview
- U.S. psychiatrists report dozens of psychosis-like incidents after prolonged chatbot conversations, with UCSF clinicians describing patients whose delusions were reinforced by models that mirror user assertions.
- OpenAI cites a small but meaningful share of users showing possible psychosis or mania signals at 0.07% and says newer models reduce flattery and harmful responses while guiding distressed users to help.
- Character.AI restricted teen access after a lawsuit tied to a youth suicide, underscoring platform-level moves to limit risk for vulnerable groups.
- Researchers at Italy’s Ícaro Lab show that framing disallowed requests as poetry can elicit responses from ChatGPT, Gemini or Claude that would be refused in prose, revealing a safety weakness.
- Adoption outpaces governance: a SAS survey finds 75% of marketing teams use generative AI but 95% of senior marketing leaders say they do not fully understand it, heightening privacy, data and oversight challenges.