ChatGPT Gets Smarter on Sensitive Chats

OpenAI is rolling out new safety updates for ChatGPT designed to improve its ability to recognize subtle, evolving cues of distress or harmful intent within conversations. These enhancements aim to help the AI respond more cautiously and appropriately in sensitive situations, distinguishing them from the vast majority of benign interactions.

The core of the update focuses on context. A seemingly innocuous request can take on a different meaning when viewed alongside earlier messages indicating distress or potential harm. OpenAI has trained ChatGPT to analyze this surrounding context, enabling it to refuse dangerous requests, de-escalate tense exchanges, or guide users toward safer alternatives.

Context is Key in Sensitive Conversations

This is particularly crucial for acute scenarios like suicide, self-harm, and harm to others. By working with mental health experts, OpenAI has refined its model policies and training data to better identify warning signs that emerge over time. This allows ChatGPT to differentiate between harmless queries and those signaling higher risk.

The system builds upon OpenAI's existing 'safe completion' approach, which aims to refuse unsafe parts of a user's prompt while still responding cautiously where possible. The goal is to escalate caution only when harm signals appear, without overreacting to everyday conversations.

Cross-Conversation Safety Monitoring

Some risks can span multiple interactions. A subtle sign in one chat might only become concerning when combined with a subsequent request in another. To address this, OpenAI has developed 'safety summaries', brief, factual notes about prior safety-relevant context.

These summaries are generated by a specialized safety model, are narrowly scoped, time-limited, and only used when a serious safety concern is detected. They are not intended for general personalization or long-term memory, but rather to capture critical safety context.

Expert Input and Performance Gains

These systems were developed with input from OpenAI's Global Physicians Network, including psychiatrists and psychologists specializing in areas like suicide prevention and forensic psychology. Their expertise helped shape decisions on when to create summaries, relevant context duration, and how the model should consider this information.

Internal evaluations show significant improvements. In long single-conversation scenarios, safe-response performance increased by 50% for suicide and self-harm cases and 16% for harm-to-others scenarios. For the default GPT‑5.5 Instant model, these updates led to a 52% improvement in harm-to-others cases and 39% in suicide and self-harm cases.

Testing also confirmed these safety measures do not negatively impact the quality of ordinary conversations. OpenAI plans to continue refining these capabilities, potentially expanding them to other high-risk areas in the future.