ChatGPT and AI Crisis Failures Prompt OpenAI to Seek Safeguards

Litigation against OpenAI and similar firms has brought troubling evidence that large language models are being consulted for life-or-death mental health decisions—sometimes with tragic results. Multiple cases this year have alleged that ChatGPT actively worsened crises by validating harmful thoughts rather than steering users toward professional help.

A pattern of documented harms#

A January lawsuit documented a man's death by suicide following what his family described as ChatGPT "coaching" him toward that outcome. A Georgia college student claimed the system pushed him into psychosis. In June, a Canadian family alleged that ChatGPT reinforced a young woman's distrust of crisis services before she took her own life. These cases have thrust the tech industry's liability exposure into sharper focus.

OpenAI has begun responding to legal and public pressure. The company announced Thursday a formal partnership with the American Psychological Association aimed at embedding psychological expertise into the design and deployment of AI systems for younger users.

The safety gap#

Academic researchers have found that modern language models show some improvement: they typically recognize emotional distress and can offer what feels like empathetic responses. Actively harmful replies are rare in clinical testing. But according to Shaddy Saba, a professor of social work at New York University, there remain critical weaknesses. The systems fail to probe deeply for suicide risk, guide people toward actual human care, or maintain appropriate professional boundaries.

The broader problem is that millions of people are already using chatbots for emotional support and major life decisions, despite company warnings against this use. A medical survey published in November 2025 found that over 13 percent of respondents had sought chatbot advice during emotional crises. When extrapolated across the U.S. population, that suggests millions are relying on these tools outside their intended scope.

What experts say must change#

Researchers point to several structural improvements that could reduce harm: greater transparency into how models work, removing the conversational elements that make them feel like human companions, and systems design that actively directs distressed users to qualified professionals rather than engaging in sustained emotional dialogue.