AI diagnostic models score 90%+ on benchmarks, then produce no statistically significant improvement when placed beside actual physicians. That gap is the story.
RareBench measures whether findings appear in a ranked list for rare genetic disease cases. LabBench tests experimental lab reasoning. Combining them into a clinical diagnosis claim mixes different tasks entirely.
The MASAI trial found 29% higher cancer detection and 44% less screen-reading workload. The ACCESS trial got eye exam completion to 100% vs 22% with standard referral. Narrow tasks, strong results. General diagnosis: much messier.
The economic unit is a completed episode, not a model inference. 10 minutes of physician review can cost more than the entire hypothetical service fee. Token prices hide most of the real cost.
Subscribe to www.onhealthcare.tech for free and paid articles, podcasts, and more.








