Google publishes multi-method framework to evaluate health AI for accuracy, safety and usefulness

Why it matters

Google for Health published a framework, with health-quality experts Elizabeth McGlynn and Dana Gelb Safran, for judging general-purpose AI in medical settings across accuracy, safety and usefulness. The companion arXiv paper describes closed-question benchmarks, autoraters and expert human review on common and high-risk scenarios. Labs have raced to put chat models in front of patients and clinicians without shared evaluation grammar; this is Google's bid to set one. It is a methods release, not a regulator mandate. Watch whether other labs adopt the same axes and whether FDA or EMA reference similar multi-method stacks.