Helping ChatGPT better recognize context in sensitive conversations

Key Findings

• — In long single-conversation scenarios, safe-response performance improved by 50% in suicide and self-harm cases and by 16% in harm-to-others cases
• — On GPT-5.5 Instant, then the default model in ChatGPT, safe-response performance improved by 52% in harm-to-others cases and by 39% in suicide and self-harm cases
• — Across more than 4,000 evaluations the safety summaries scored an average 4.93 out of 5 for safety relevance and 4.34 out of 5 for factuality
• — Internal testing found responses in everyday chats broadly comparable, with no meaningful user preference between responses with and without safety summaries
• — Safety summaries are described as distinct from general personalization or long-term memory: created by a safety-reasoning model, narrowly scoped, time-limited, and used only for serious safety concerns
• — Design decisions on when summaries are created, how much prior context is relevant and how long it is considered were informed by psychiatrists and psychologists in OpenAI's Global Physicians Network with expertise in forensic psychology, suicide prevention and self-harm

Methodology Notes

Vendor-published post, not peer reviewed; all figures are self-reported and the evaluation design, scenario construction and scoring rubric are OpenAI's own and not independently audited. Percentage improvements are relative and no baseline rates are given, so the absolute safe-response rate cannot be recovered from the post. openai.com returns 403 to automated fetchers and the Internet Archive was globally offline on the verification date; the full text was retrieved through the r.jina.ai text proxy and the publication date confirmed independently against the pubDate in OpenAI's own news RSS feed (openai.com/news/rss.xml), which is served without the block. All quoted figures were read from the retrieved text.

Sources