OpenAI Podcast Episode 14 - Building AI for better healthcare
84 snips
Mar 16, 2026 
Karan Singhal, leader of Health AI Research who studies safe, evaluated models for medicine, and Dr. Nate Gross, a physician and health-policy leader focused on improving care. They talk about training and evaluating models for sensitive health questions. They discuss clinician workflows, partnerships with hospitals, real-world trials, data integration, and practical AI tools for patients and clinicians.
AI Snips
Chapters
Transcript
Episode notes
ChatGPT Health Balances Privacy With User Context
- ChatGPT Health pairs stronger security with user-controlled context to make health conversations both private and personally relevant.
- OpenAI added encryption, a one-way valve against training on user healthcare data, and tools for users to import their own records and context.
HealthBench Measures Real Conversations Not Just Exams
- OpenAI created HealthBench with physicians to evaluate multi-turn medical conversations across many nuanced criteria rather than relying on exams.
- They collaborated with ~250 physicians to produce ~49,000 rubric dimensions and 5,000 conversations used to score model safety and performance.
Ask For Context Before Giving Medical Answers
- Ask for more context before giving health advice when user input is sparse to reduce risk and improve usefulness.
- Train models to seek clarification rather than guessing from brief prompts like "it burns" to avoid unsafe recommendations.