Helping ChatGPT better recognize context in sensitive conversations

People come to ChatGPT every day to talk about what matters to them—from everyday questions to more personal or complex conversations. Across hundreds of millions of interactions, some of these conversations include people who are struggling or experiencing distress. We design our systems to respond carefully in these moments, including by providing crisis resources and connecting people with someone they trust ⁠ when needed.

Today, we’re sharing new details about safety updates that help ChatGPT better recognize when risk may be emerging over time by identifying subtle or evolving cues, and using that context to inform safe responses. This helps ChatGPT distinguish between the hundreds of millions of safe interactions people have every day and the much rarer cases where added caution is needed, so it can respond more carefully—for example, by de-escalating, refusing harmful details, or redirecting toward safer alternatives.

These improvements build on years of extensive work ⁠ across model training, evaluations, monitoring systems, and more than two years of collaboration with mental health and safety experts.

These experts helped inform decisions around when safety summaries should be created, how much prior context may be relevant, and how long the model should consider that context when responding. Their input helped ground this work in real-world expertise and support more appropriate responses in sensitive situations.

In sensitive conversations, context can matter as much as a single message. A request that appears ordinary or ambiguous on its own may carry a very different meaning when viewed alongside earlier signs of distress or possible harmful intent. To respond appropriately, we train ChatGPT to recognize the potential harmful intent from the surrounding context so that it can refuse the request, de-escalate, and guide the user toward support.

These cases are uncommon, but critically important to get right. Our goal is to help ChatGPT connect relevant signals when they matter without overreacting in ordinary conversations.

We focused this work on acute scenarios including suicide, self-harm, and harm-to-others. Working with mental health experts, we updated our model policies and training to improve ChatGPT’s ability to recognize warning signs that emerge over the course of a conversation and use that context to inform more careful responses.

In these rare, high-risk situations, ChatGPT can better distinguish between benign requests and those that may signal a higher risk of harm. This builds on our safe completion approach ⁠ , which is designed to refuse unsafe parts of a user request, and respond cautiously where it can safely do so. The goal is to help the model respond more appropriately to context, escalating caution when signals of harm emerge within conversations, while continuing to respond helpfully in benign situations.