Summary
OpenAI detailed new safety updates that let ChatGPT track safety-relevant context within and across conversations in rare high-risk situations. The system uses narrowly scoped safety summaries to help the model recognize escalating risk over time and respond more cautiously in cases involving self-harm or harm-to-others.
What changed
OpenAI updated ChatGPT’s safety stack to use short-lived safety summaries and model training that better connect subtle warning signs across messages and conversations.
Why it matters
This shows safety state becoming part of production chat architecture, not just one-turn moderation. It suggests a growing split between general-purpose memory features and narrowly scoped safety memory designed to influence behavior only in rare, high-risk cases.
Evidence excerpt
OpenAI says the update uses narrowly scoped safety summaries to preserve earlier safety-relevant context, improving safe responses when risk emerges over time within or across conversations.