GPT-5 System Card Addendum: Sensitive Conversations

OpenAI has released an addendum to the GPT-5 system card detailing improvements to how the model handles sensitive conversations involving mental and emotional distress. By collaborating with over 170 mental health experts, OpenAI has reduced responses that fall short of desired safety behaviors by 65-80%.

Model Update and Safety Improvements

On October 3, 2025, OpenAI deployed an update to ChatGPT's default model to enhance its ability to recognize signs of distress and provide supportive, care-oriented responses. This update focuses on guiding users toward real-world support systems rather than providing insufficient or inappropriate responses in sensitive contexts.

Baseline Safety Evaluations

To measure the efficacy of the update, OpenAI compared the October 3 version of the model against the August 15 version of ChatGPT's default model (known as GPT-5 Instant). The resulting evaluations serve as the baseline for the model's improved performance in handling sensitive conversations.

Expert Collaboration

The safety improvements were developed in partnership with more than 170 mental health experts. This collaboration ensured that the model's responses are more reliably aligned with professional standards for recognizing and supporting individuals experiencing emotional distress.