Improving health intelligence in ChatGPT
Summary
OpenAI built health intelligence into ChatGPT by assembling 100+ practicing physicians who evaluate conversations in real-time, focusing on how responses are communicated as much as accuracy. One doctor reaches 15 patients daily in clinic but impacts millions through AI—demonstrating medicine's scalability paradox.
Key Takeaways
- Assemble domain experts as continuous evaluators, not one-time consultants. OpenAI's team of hundreds of practicing physicians evaluates ChatGPT responses daily, ensuring real-world medical context informs model improvements iteratively.
- Tone and framing matter as much as factual accuracy in high-stakes domains. Evaluating 'impact that a response is going to have on a person' alongside accuracy prevents technically correct but harmful outputs.
- Use AI to democratize specialized knowledge geographically. A hematologist and transfusion medicine specialist noted AI can take super-specialty expertise to underserved regions where no human specialist exists.
- Measure leverage through reach multiplier effect. A pediatrician seeing 15 patients daily can impact millions through AI work—a 60,000x+ leverage ratio that justifies shifting time allocation.
- Barrier reduction is the core value proposition in regulated domains. Focus product strategy on accessibility and lowering information barriers rather than replacing professional judgment.
Related topics
Transcript Excerpt
We live in a very complex healthcare system. This is a technology that can hugely [music] enhance the way we live. And there's so much promise if we can get this right. >> [music] >> We all went into medicine because we want to take care of people. >> As a 13-year-old, I saw my dad change in ways that I never expected. That really lit the fire in me to be a physician. >> My grandfather [music] and my uncle were both diagnosed with schizophrenia. It made me choose [music] psychiatry as my specialty. >> I decided to be a pediatrician to show some compassion. My current contribution [music] to the AI team is a continuation of that desire. >> [music] [music] >> I've been a pediatrician for 16 years. I lead a team of hundreds of physicians, and we evaluate >> OpenAI's [music] models for health,…