ChatGPT safety upgrades spot risk over time

OpenAI's latest safety updates focus on a critical challenge: recognizing when risk emerges gradually across a conversation or even multiple separate conversations. The company emphasizes that context can matter as much as a single message, and that a request appearing ordinary on its own may carry a different meaning when viewed alongside earlier signs of distress. This work builds on over two years of collaboration with mental health and safety experts, including psychiatrists and psychologists from the Global Physicians Network.

The improvements involve training ChatGPT to use safety-relevant context more carefully, allowing it to de-escalate, refuse harmful details, or redirect toward safer alternatives in rare high-risk situations. The company also introduced safety summaries—short factual notes created by a model trained for safety reasoning—to help connect warning signs across separate conversations. These summaries are narrowly scoped and kept only for a limited time, designed not to serve as general personalization or long-term memory.

The stakes hinge on striking the right balance. While internal tests show significant gains in high-risk scenarios, ordinary conversations must remain unaffected. OpenAI reports that responses in everyday chats were broadly comparable, with no meaningful user preference between those with or without safety summaries. The future may see these methods applied to other high-risk areas like biology or cyber safety, but that remains a possibility rather than a commitment. For now, the focus stays on self-harm and harm-to-others scenarios, where getting it right is critically important.