AI That Reads Your Medical Records Still Talks Like a Robot

Hospitals are stuffed with digital health records, but finding the right answer to a clinical question can be like searching a thousand-page diary. A team of researchers built an AI system that can do exactly that: read electronic health records, generate answers, and point to the exact sentences it used for evidence. In a worldwide competition called ArchEHR-QA 2026, their system placed third on evidence identification and fifth on matching answers to evidence. That's genuinely useful — it could help doctors quickly pull up relevant patient history or research past cases.

But here's the catch that should matter to any patient: the AI's answers were much harder to read than answers written by real doctors. The system produced text that required a 3.2-grade higher reading level — meaning a typical patient might need a college education to fully understand what the AI wrote. The words and sentence lengths were similar to human writing, but something about the phrasing made it more difficult. This is a big deal, because if AI is going to explain medical information to people, it needs to be as clear as possible.

The researchers suggest that readability should be an explicit goal when building medical AI, not an afterthought. Their approach used a clever "answer-first" strategy: the system first drafts a tentative answer, then looks back at the records to find supporting evidence, which is more accurate than scanning the records without a question in mind. It also used a voting technique over five separate runs to decide which evidence links to keep. The team has released its code and prompts publicly so other researchers can build better, clearer medical AI.