OpenAI Podcast Episode 14 - Building AI for better healthcare
Mar 16, 2026
Karan Singhal, who leads Health AI Research and focuses on model evaluation and safety; Dr. Nate Gross, a physician leading Health efforts with experience in policy and clinical care. They discuss training models for sensitive health questions, HealthBench and physician-guided evaluation, tailoring responses to patients and clinicians, integrating records and devices, and real-world deployments and measurement challenges.
AI Snips
Chapters
Transcript
Episode notes
ChatGPT Health Adds Security And Personal Context
- ChatGPT Health combines extra security with context-aware features so health conversations are both private and personalized.
- Nate Gross describes an encrypted one-way valve and tools that let users bring chosen context so responses are tailored to their situation.
HealthBench Uses Clinician‑Driven Multi dimensional Eval
- OpenAI built HealthBench with about 250 physicians to evaluate multi-turn health conversations across many dimensions of safety and performance.
- Karan Singhal says HealthBench measured ~49,000 performance dimensions using 5,000+ conversations and physician-created rubrics.
Ask For Context Before Making Clinical Judgments
- Ask for more context before answering ambiguous health queries rather than offering premature conclusions.
- Karan notes users often type minimal prompts like "it burns" and models should prompt for details to be safest and most helpful.