AI in Medicine Brief — August 19, 2026

A new briefing highlights that current large language models perform well on clearly defined clinical problems but struggle during the “preformulation” phase of patient encounters, when symptoms and concerns are vague or poorly framed, posing risks for real-world diagnosis. The report also covers a national-scale generative model of electronic health records developed in England, emerging risks of “agent gaming” in multi-agent clinical AI systems, and new findings on protected health information leakage despite de-identification.