OpenAI updates ChatGPT's handling of sensitive mental-health conversations

OpenAI published an addendum to the GPT-5 system card describing changes made to how ChatGPT’s default model handles conversations touching psychosis and mania, self-harm and suicide, and emotional reliance on the chatbot. The company said it worked with more than 170 mental-health clinicians — psychiatrists, psychologists and primary-care physicians — to build “taxonomies” defining what a harmful response looks like and what an ideal one should do, and used them to retrain the model that was deployed on 3 October.

On an internal evaluation of more than 1,000 challenging mental-health-related conversations, OpenAI reported the updated model scored 92% compliant with its desired-behaviour taxonomy, against 27% for the version of the default model live since mid-August — a reduction in falling-short responses the company put at 65-80%. Conversations recognised as sensitive are now routed to GPT-5 Instant specifically, rather than left to whichever model a user’s account happens to be using. As with earlier OpenAI safety claims, the evaluation was designed, run and scored internally rather than by an outside auditor, and the company did not publish the taxonomies themselves or make the underlying conversation set available for independent review.

The update followed several wrongful-death lawsuits alleging ChatGPT contributed to users’ suicides, including a suit filed in August by the parents of 16-year-old Adam Raine, and arrived two weeks after OpenAI formed an outside expert council to advise on ChatGPT’s effects on user wellbeing. It was published the same day as an update to OpenAI’s Model Spec, the document defining desired model behaviour more broadly, which placed mental-health support explicitly within the model’s default remit rather than treating it as an edge case to be deflected.