We craft. We break. We verify. We lead.
Research on clinical AI after it is switched on.
Benchmarks measure the ceiling. Deployment needs a floor.
Thirty senior figures in medical AI published a Nature Medicine comment calling the benchmarks misleading. They are right, and they are answering a question hospitals are not asking.
91%of deployed models degraded1,000+AI devices cleared in the US52%of emergencies undertriaged
Medical AI's trusting trust problem
Nine frontier models were tested as judges on a slice of OpenAI's HealthBench. The most expensive judge on the market misses explicit negations 83% of the time.
83%missed by the priciest judge9frontier models tested35% → 1.7%hack rate after hardening