Helping ChatGPT better recognize context in sensitive conversations

ChatGPT has received new safety updates to improve its ability to recognize and respond to potential risks over time. These updates help the AI distinguish between routine interactions and rare cases where more caution is needed, allowing it to de-escalate, refuse harmful details, or redirect users to safer alternatives and crisis resources when appropriate.

One key improvement focuses on understanding context, as a seemingly ordinary request can carry different meaning when viewed alongside earlier signs of distress. ChatGPT is now trained to identify potential harmful intent from surrounding context, enabling it to refuse requests or guide users toward support. This capacity is crucial for acute scenarios like suicide, self-harm, and harm to others.

To address safety risks that might emerge across separate conversations, ChatGPT now uses "safety summaries." These brief, factual notes capture relevant safety context from previous interactions, helping the model connect signals and escalate caution when needed. These summaries are strictly scoped, temporary, and used only for serious safety concerns, not for personalization or long-term memory.

The development of these features involved extensive collaboration with mental health and safety experts. Their input guided decisions on when to create safety summaries, how much prior context is relevant, and how long the model should consider that context. This expertise ensures more appropriate responses in sensitive situations.

Internal evaluations show significant improvements in ChatGPT's safe responses. In long single-conversation scenarios, safe-response performance improved by 50% for suicide and self-harm cases, and by 16% for harm-to-others cases. Across multiple conversations, using GPT-5.5 Instant, safe-response performance improved by 52% in harm-to-others cases and 39% in suicide and self-harm cases. Safety summaries also demonstrated high relevance and factual accuracy in evaluations.