OpenAI updates ChatGPT to better spot risk in sensitive conversations

OpenAI has described safety updates to ChatGPT that help the model understand context in sensitive conversations. The system is now better at noticing when risk does not show up right away but builds over the course of a conversation, or even across several separate chats.

Why context matters

In a single message, a user may ask something neutral or ambiguous, and without the preceding messages such a request looks harmless. But if the conversation earlier contained signs of distress, talk of self-harm, or hints of harming others, the meaning changes sharply. OpenAI focused the update on exactly these cases: the model was trained to link signals across several messages and to raise its caution not in every conversation, but only where worrying signs actually appear.

The company says these are rare but critically important scenarios, above all suicide, self-harm and threats to other people. In such situations ChatGPT must do more than answer formally: it should know when to decline dangerous details, lower the temperature of the conversation, and gently steer the user toward safer help. The goal of the update is not to make the model overly anxious, but to teach it to tell ordinary conversations from genuinely risky episodes.

What changed

The key addition is safety summaries: short, factual notes about important safety context. They are created by a separate model trained for safety reasoning, and they are used only in rare cases when there is a serious risk signal. According to OpenAI, these notes are not general personalization and do not turn into long-term memory about the user: they are stored for a limited time and applied only when past context is truly needed for a safer answer.

• Match signals from current and past messages
• Help account for risk across separate chats
• Prompt the model on when to de-escalate a conversation
• Strengthen refusals of dangerous details in a request
• Redirect the user to safer alternatives

OpenAI also stresses that the system was not developed solely within the safety team. Psychiatrists and psychologists from the Global Physicians Network took part, including specialists in forensic psychology and in suicide and self-harm prevention. They helped determine when safety summaries should be created, how much prior context is actually useful, and how long the model should take it into account when responding. This is an important detail: the company relied not only on general heuristics but on the practice of specialists who work with such crisis cases.

What the tests showed

OpenAI cites several internal metrics. In long scenarios within a single conversation, the share of safe responses rose by 50% in cases involving suicide and self-harm, and by 16% in cases involving harm to others.

The company also tested performance across multiple conversations and on several models. For GPT-5.5 Instant, which is currently the default model in ChatGPT, safe responses improved by 52% in harm-to-others scenarios and by 39% in suicide and self-harm scenarios. This indicates the system has become better at noticing risk that accumulates over time, rather than reacting only to obvious red flags.

The company also assessed the quality of the safety summaries themselves. Across more than 4,000 internal evaluations, they received an average score of 4.93 out of 5 for safety relevance and 4.34 out of 5 for factual accuracy.

OpenAI also checked whether adding such context degrades ordinary conversations. In internal tests, responses in everyday chats remained broadly comparable, and no noticeable user preference was found between variants with and without safety summaries. In other words, the bet is on more precise caution without a perceptible drop in quality in normal scenarios.

What this means

OpenAI is moving toward more robust use of earlier context, not for personalization but for safety in rare critical situations. If the approach really scales without excess false positives, ChatGPT will be able to handle difficult conversations more carefully, where risk becomes clear only across a chain of messages.

For the industry this is an important signal: safety increasingly depends not on a single request but on a model's ability to see a situation develop over time.