Executive Summary
OpenAI has deployed new safety updates for ChatGPT designed to better recognize and respond to sensitive conversations where risk may emerge over time. The system now analyzes context both within a single conversation and across multiple sessions to identify subtle cues related to potential self-harm or harm-to-others. A key feature is the use of temporary, model-generated "safety summaries" that help the AI maintain safety-relevant context, enabling it to respond more appropriately by de-escalating or redirecting users without impacting performance in ordinary conversations.
Key Takeaways
* Contextual Risk Detection: The updates improve ChatGPT's ability to understand that a seemingly benign request may be harmful when viewed in the context of previous messages in the conversation.
* "Safety Summaries" for Cross-Conversation Awareness: A new system creates short, factual notes about safety-relevant context from earlier conversations. These temporary summaries are used only in rare, high-risk situations to inform safer responses in later interactions.
* Developed with Experts: The system was developed with input from mental health professionals, including psychiatrists and psychologists from OpenAI's Global Physicians Network, to inform policy and model training.
* Quantified Performance Improvements: Internal testing showed significant improvements in providing safe responses. In long, single-conversation tests, safe responses improved by 50% for suicide/self-harm cases. Across multiple conversations on GPT-5.5 Instant, performance improved by 39% for suicide/self-harm and 52% for harm-to-others cases.
* Availability: The updates have been implemented and tested on current models, including GPT-5.5 Instant, the default model in ChatGPT.
Strategic Importance
This update addresses a critical AI safety challenge by enabling stateful risk assessment, bolstering user trust and mitigating potential real-world harm from sensitive conversations. Proactively managing these complex safety scenarios is crucial for OpenAI's public image and long-term regulatory standing.