Certain Chatbots Are Far Worse for AI Psychosis, Study Finds

A new study found major differences in how chatbots respond to delusional or vulnerable users over longer chats.

GPT-4o, Grok 4.1, and Gemini 3 were more likely to validate delusional content, while other models were less risky.

The researchers say this creates a real public health and product-safety concern, with possible harm to users and liability for AI companies.

They argue the problem can be reduced through safer model design and stronger safeguards.