Helping ChatGPT better recognize context in sensitive conversations (openai.com)

OpenAI has announced significant updates to ChatGPT’s safety protocols, enhancing its ability to recognize and respond to potentially harmful conversations. By leveraging advanced contextual understanding, the model can now identify subtle cues of distress over the course of interactions, which is crucial for distilling benign requests from those indicating higher risk for self-harm or harm to others. These updates are based on extensive collaboration with mental health professionals and incorporate years of model training focused on safety. The improvements include the introduction of "safety summaries," short contextual notes to help the model understand related risks across conversations. This enables ChatGPT to more effectively assess the evolving nature of a user's intent, leading to a remarkable 50% improvement in safe-response performance in suicide and self-harm scenarios, and a 16% increase for harm-to-others cases, as demonstrated in rigorous internal evaluations. The ability to connect relevant signals over time is fundamental in managing sensitive topics, ensuring that the AI can provide appropriate help while maintaining responses in regular conversations. This initiative reflects a broader commitment to developing responsible AI systems in high-risk situations and may pave the way for similar approaches in other critical areas.