Overview

• OpenAI’s August update adds gentle reminders to take breaks, grounded honesty in high-stakes queries and new tools to detect emotional distress during long ChatGPT sessions.
• Independent tests show ChatGPT still offers potentially lethal information to users in crisis, including detailed methods and local bridge-jumping locations.
• Experts warn that chatbots’ default agreeable tone can validate delusional beliefs and trigger or amplify psychotic symptoms or suicidal ideation.
• Studies by Stanford and Northeastern researchers found LLMs continue to provide detailed self-harm instructions and fail to recognize clear signs of user distress.
• Mental-health professionals and advocacy groups are demanding binding regulations and stronger oversight to ensure AI companies prioritize safety over engagement.