How to help people at the moments they need it most

As ChatGPT is used beyond search and coding as a tool for personal advice, coaching, and emotional support, responding to users' mental and emotional crisis situations has become important. OpenAI is continuously improving safeguards so that the model can recognize signs of crisis, respond appropriately, and connect users with real experts and resources.

The key safeguards and response systems are as follows:

• Empathetic responses and blocking: Refuses to provide self-harm instructions and shifts to empathetic language, while responses that violate safety guidelines are automatically blocked.
• Connection to real resources: When suicidal intent is expressed, users are directed to country-specific crisis hotlines such as 988 (US) and Samaritans (UK), or to specialized organizations like findahelpline.com.
• Escalation of dangerous situations: If signs of intent to physically harm others are detected, a specialized team reviews the case, and law enforcement may be notified in urgent situations. However, cases of self-harm are not directly reported to law enforcement in order to protect personal privacy.

The recently released GPT-5 reduced inappropriate responses in mental health emergencies by more than 25% compared to the previous model, 4o. In addition, OpenAI introduced a training method called safe completions, which provides the most helpful possible answers while staying within safety boundaries, improving issues around emotional dependency and sycophancy.