Overview
- OpenAI reports GPT-5 reduces undesired replies by 39% versus GPT-4o across sensitive categories, with a 52% drop in suicide and self-harm scenarios.
- Safety-behavior compliance is cited at 92% compared with 27% previously, and reliability in long conversations surpasses 95%.
- The system is designed to flag crisis indicators and then recommend professional help, block sensitive content, or decline to answer certain requests.
- Responses that could encourage strong emotional attachment fell by 42% under GPT-5, with desired-behavior compliance at 97% versus 50% in the prior version.
- OpenAI estimates about 0.07% of weekly active users—around 1.2 million people—show potential crisis indicators in 0.01% of messages, following work with 170 clinicians selected from a pool of 300 across roughly 60 countries.