tech
Helping ChatGPT better recognize context in sensitive conversations
New safety updates help ChatGPT respond safely when risk emerges over time.
TL;DR
- ChatGPT has received safety updates to better recognize and respond to emerging risks in conversations.
- The updates focus on identifying subtle or evolving cues that may indicate distress or harmful intent over time.
- Context within conversations is crucial for understanding requests, especially in sensitive scenarios.
- The system is trained to de-escalate, refuse harmful details, or redirect users toward safer alternatives.
- Improvements were made in collaboration with mental health experts, focusing on acute scenarios like suicide, self-harm, and harm-to-others.
- Safety summaries, short notes about prior safety-relevant context, are used to inform responses in rare, high-risk situations.
- Internal evaluations show performance improvements of up to 52% in recognizing and responding safely to harm-to-others cases and 39% in suicide and self-harm cases.
- Testing indicates that these safety updates do not meaningfully reduce the quality of responses in ordinary conversations.