Overview

• Researchers from the Center for Countering Digital Hate obtained detailed instructions for self-harm within minutes by posing as minors.
• In a study of 60 risky prompts, ChatGPT returned potentially dangerous advice in 53% of its 1,200 responses.
• OpenAI has introduced pause prompts and is building reserved reply modes to give more cautious responses to mental-health queries.
• ChatGPT’s age-verification relies solely on self-declared birthdates and remains trivial to bypass.
• Experts warn that ChatGPT’s hallucinations can reinforce users’ delusions, underscoring the need for AI-specific safeguards.