Key Takeaways
- OpenAI’s GPT-5 reduces unsafe mental health responses by up to 80%
- Approximately 0.15% of weekly ChatGPT users discuss suicide
- New model better handles psychosis, mania, and emotional dependence
- Over 170 clinicians helped develop safety improvements
OpenAI has revealed that GPT-5, the latest model powering ChatGPT, dramatically improves how the AI handles sensitive mental health conversations. The company reports reductions of up to 80% in unsafe responses when users show signs of distress, including suicidal thoughts.
According to a Monday blog post, approximately 0.15% of ChatGPT’s weekly users discuss suicidal thoughts or plans. While this represents a small fraction of users, it’s significant given the platform’s massive global reach.
Clinical Collaboration Drives Safety Improvements
The safety enhancements come after months of collaboration with OpenAI’s Global Physician Network, comprising nearly 300 clinicians across 60 countries. More than 170 mental health professionals directly contributed by writing responses, defining safe behavior, and reviewing how the model handles sensitive scenarios.
“ChatGPT is not a therapist,” OpenAI emphasized. The goal isn’t to provide therapy but to recognize distress signs and gently redirect users to professional support. The updated model now connects people more reliably to crisis helplines and occasionally suggests breaks during emotionally charged sessions.
Quantifiable Safety Gains
Internal testing shows GPT-5 produces 65-80% fewer unsafe responses in production when users display mental health distress. In structured evaluations graded by independent clinicians, GPT-5 cut undesirable replies by 39-52% compared to GPT-4o.
Automated testing scored GPT-5 at 91-92% compliance with desired behavior, up from 77% or lower in older models. The system also maintains over 95% consistency in lengthy conversations, where earlier models often struggled.
Tackling Emotional Dependence
OpenAI is addressing a newer challenge: emotional reliance, where users form unhealthy attachments to the chatbot. Using a new taxonomy to identify this behavior, GPT-5 now produces 80% fewer problematic replies in these scenarios, often steering users toward human connection instead of validating dependence.
Despite these improvements, OpenAI acknowledges mental health conversations remain rare and hard to quantify precisely. At such low prevalence rates, even small variations can distort results. Experts also disagree on safety standards—clinicians reviewing the model’s responses reached the same judgment only 71-77% of the time.
