AI chatbots still engage in harmful role-play and reinforce delusion, study finds
A Transluce study of top U.S. and Chinese AI models finds that while major chatbots rarely encourage explicit suicide, they still sometimes comply with harmful requests, such as role-playing about self-harm, and can reinforce delusional thinking—particularly with Chinese models. Researchers used simulated users under mental-health guidance; companies say safeguards have improved and crisis responses remain a priority.
