OpenAI announced on May 14, 2026 that it has deployed safety updates to help ChatGPT recognize when risk may emerge over the course of a conversation. The work focuses on acute scenarios including suicide, self-harm and harm-to-others, and aims to use earlier cues in a dialogue to inform safer replies.
How the updates work
The company said the system now links subtle or evolving signals from prior messages so the model can refuse unsafe requests, de-escalate, or redirect users toward safer alternatives when appropriate. A new component called safety summaries produces short, factual notes about earlier safety-relevant context; these summaries are generated by a model trained for safety reasoning tasks.
OpenAI described the summaries as narrowly scoped, retained only for a limited time, and used solely when relevant to a serious safety concern. The summaries are intended to capture factual safety context rather than to serve as general personalization or long-term memory.
Expert input and evaluation results
Development involved input from mental health professionals in the Global Physicians Network, including psychiatrists and psychologists with experience in forensic psychology, suicide prevention and self-harm. These experts advised on when to create summaries, how much prior context to consider, and how long context should remain relevant.
OpenAI reported improvements in internal evaluations designed to emulate high-risk scenarios. In long single-conversation tests, safe-response performance increased by 50% for suicide and self-harm cases and by 16% for harm-to-others cases. On GPT‑5.5 Instant, the current default model, the company reported a 52% improvement in harm-to-others cases and a 39% improvement in suicide and self-harm cases.
Across more than 4,000 evaluations of the safety summaries, OpenAI said they received an average safety relevance score of 4.93 out of 5 and a factuality score of 4.34 out of 5. The company also reported that responses in ordinary conversations remained broadly comparable with or without safety summaries, with no meaningful user preference in internal tests.
OpenAI noted the updates currently target self-harm and harm-to-others, and said it may explore applying similar methods to other high-risk domains such as biology or cyber safety in the future.
Original source: OpenAI News