a16zShould You Trust ChatGPT With Your Health?
a16z
Health access gaps 0:00
Healthcare spending is huge, yet care is scarce. People can wait weeks or months for specialists, then get only a short visit and leave with more questions than they came in with. The discussion frames AI as something that can stay with you between visits, helping you prepare, follow up, and make sense of what doctors say.
Trust and safety 0:31
The work begins with the hard part: trust. Health questions can be emotional, and a user may want reassurance when the safer answer is to push for more care. OpenAI says it is working with licensed mental health experts and physicians to define model behavior, test tricky cases, and make the system dependable enough to measure.
Healthcare Friction 13:00
Healthcare is full of small frictions after discharge, including billing errors and insurance disputes. The speaker says this is where agents and computer use can help, because the system is fragmented, legacy software is common, and data is often a bottleneck. Even a virtual machine can make some things easier for AI agents, and the hope is that many barriers between patients, clinicians, and data will fade soon.
Trust and Uncertainty 14:30
Trust comes first, because the same model must serve both a patient and a specialist. The team works with experts around the world to tailor responses for lay users, complex-care patients, and clinicians, and it measures how much jargon the model uses. It also trains the model to handle disagreement, to give safety over reassurance when needed, and to admit uncertainty when evidence is thin.
Trust and integrations 26:00
Security, compliance, and privacy came first, because healthcare groups need HIPAA-ready protections before they will trust the tools. The work then moved toward partnerships and deployment, with a focus on speaking each group’s language and removing bottlenecks to adoption. Epic became a key example: ChatGPT can be used inside Epic, and Epic can be used in ChatGPT, with sidebar access and more connectors on the way.
Care beyond information 28:01
Mental health work followed the same pattern. The team put safety safeguards into DOTS, then built Mental Health Bench with licensed experts to measure more than harm and to raise the ceiling for good results. In medical care, the view was narrower: AI can help with information, navigation, scheduling, and chronic-care flags, but it will not replace clinicians, empathy, physical exams, or the full care journey.
Move 37 in health 39:01
The speaker wants health systems to find a “move 37” moment, when a model sees a useful pattern a human would miss. One example is a safety net for clinicians that flags overuse of antibiotics and pushes them to think beyond the local case toward antibiotic resistance. The same idea could come from long records across EHR data, chat data, and other history that a person cannot hold in mind at once.
Cheaper intelligence 43:30
He says the cost of health intelligence should keep falling, and that matters for the whole system because care spending is huge. A recent cheap model, GPT-6 Luna, beat the older best model on Healthbench Professional while costing 34 times less. He expects health intelligence to become more common and cheaper, though not enough by itself to fix every spending problem.
Future health tools 52:32
The focus is on two goals: making trustworthy health intelligence available to clinicians and patients, and raising the ceiling with tools that can do things not possible before. A proactive personal health assistant could help you navigate care, schedule visits, and flag problems early.
New treatments ahead 54:31
Over the next five to ten years, the biggest gains may come from AI helping discover new targets and molecules, then prioritizing and validating them. That could mean faster trials, smoother studies, and more personalized medicines built from existing data.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
