- OpenAI announced a new safety system that helps ChatGPT recognize risks arising gradually over time, rather than relying solely on a single message.
- The new system focuses on high-risk situations such as suicide, self-harm, and harm to others.
- ChatGPT can now connect small or ambiguous signals that appear throughout multiple conversations to assess danger levels more accurately.
- When detecting a escalating risk, the model will prioritize de-escalation, refusing dangerous content or directing users to safer support.
- OpenAI developed “safety summaries,” which are short notes on critical safety contexts that appeared in previous conversations.
- Safety summaries are only stored temporarily, used for severe risk cases, and do not function as long-term personalized memory.
- The system was built alongside a network of psychiatrists and suicide prevention experts from OpenAI’s Global Physicians Network.
- In internal evaluations, safety response performance increased by 50% in suicide/self-harm situations and by 16% in situations involving harm to others within long conversations.
- On GPT-5.5 Instant, safety response performance increased by 52% for situations involving harm to others and 39% for suicide/self-harm.
- OpenAI evaluated over 4,000 safety summaries, with an average safety relevance score of 4.93/5 and a factual accuracy score of 4.34/5.
- The company stated that adding safety context did not degrade the quality of casual conversations in internal testing.
- 📌 Conclusion: OpenAI is transforming ChatGPT from a chatbot that responds to each message individually into a system capable of “seeing the big picture” of user behavior in sensitive situations. The most critical point is that the model can now detect cumulative risks over time instead of waiting for an explicit danger signal to appear immediately. This is a major step forward in AI safety, but it also opens up new debates regarding privacy, context memory capacity, and the extent of user behavior surveillance by future AI systems.
VIET NAM CONSULTING AND MEASUREMENT JOINT STOCK COMPANY
© 2026 Vietmetric