No law grants AI chats physician-patient privilege, so courts can compel transcripts. OpenAI's enterprise BAA fixes this for hospitals, but free ChatGPT leaves vulnerable users exposed. On-device models may be the ultimate privacy shield.
Category: AI
Level‑5 Healthcare: Why Prescribing Will Decide When AI Becomes a Real Doctor
Diagnostic AI aces OSCEs but stops before writing scripts. Autonomous prescribing for titrations, refills, and simple infections could unlock Level-5 care. FDA already approves closed-loop insulin—lisinopril adjustments should follow the same path.
Apple Watch: From Activity Rings to an AI-Powered Check-Engine Light
Apple's Wearable Behavior Model compresses a week of activity, sleep, and vitals into a 256-dimensional fingerprint that can detect pregnancy and infection days early. The research is promising, but shipping requires clinical validation and FDA clearance.
Don’t Believe the Hype — Medical Superintelligence Isn’t Here Yet
Grok4 claims PhD-level performance but fails basic clinical safety checks—anatomy vetoes, unit verification, physics sanity. Benchmarks reward fluent certainty over real judgment. We need tests that measure trend detection, causal vetoes, and calibrated caution.
What I’ve Learned About LLMs in Healthcare (so far)
Microsoft's AI can diagnose NEJM cases at 85% accuracy, but real hospital charts are too messy for today's LLMs. The bottleneck isn't reasoning—it's context engineering that can parse fragmented, stale EHR data into something models can actually use.
Be Curious, Not Judgmental 