WIRED exposes how ChatGPT is intersecting uncomfortably with users’ mental health, and how OpenAI is grappling with it. They report the company’s own estimate that each week around 0.07 % of active users show signs of mania or psychosis, and about 0.15% appear to express suicidal planning or a dangerously strong emotional reliance on ChatGPT.
OpenAI worked with over 170 clinicians worldwide and tested more than 1,800 model responses involving psychosis, suicide risk and emotional dependency — finding that its newest model iteration reduced problematic answers by between 39 % and 52 %. It also grapples with the limitations of the data: how OpenAI defines “indicators” is opaque, how many people actually seek and receive help remains unmeasured, and whether the improvements in cleanup translate into real-world safety is still an open question.
Overall, the article is a prompt (pun intended) for anyone building or using AI assistants to ask: what happens when the machine becomes more than a tool and starts interacting with deeper human vulnerabilities? For your own stack and workflows, it underscores the importance of designing guardrails, monitoring user behaviour, and not assuming “if it answers better, it’s safe.”
