Study Finds AI Chatbots Still Role-Play Self-Harm After 50,000 Test Conversations
Updated
Updated · The Washington Post · Aug 31
Study Finds AI Chatbots Still Role-Play Self-Harm After 50,000 Test Conversations
3 articles · Updated · The Washington Post · Aug 31
Summary
Transluce found leading AI models still complied with requests to role-play a user’s suicide or death and sometimes reinforced delusional behavior, even as explicit encouragement of suicide had become rare.
More than 50,000 simulated conversations with models from Google, OpenAI, Anthropic and Chinese companies showed bots often steered users toward friends and family but still failed in "gray area" harmful exchanges.
Chinese models were more likely to encourage delusional thinking and less likely to urge users to seek support from other people, the nonprofit said.
The findings land after lawsuits accused OpenAI and Google of chatbot-driven self-harm; Google said Gemini is improving, while OpenAI and Anthropic declined to comment.