A randomised controlled trial published in npj Digital Medicine in 2026 examined not whether large language models can diagnose, but how best to combine their judgement with a clinician's. The study, led by Stanford Medicine with collaborators at Harvard, asked 70 doctors to work through five de-identified patient cases under three conditions: the AI model alone, a clinician using conventional medical references, and a clinician working alongside the model.

The most informative result concerned the order in which the two were consulted. When the model was asked for a second opinion after the clinician had recorded theirs, it tended to side with the clinician, even though it had been instructed to reason independently. The arrangement that worked best was parallel: the clinician and the model assessed each case at the same time, after which the model produced a synthesis that set the two assessments side by side and marked where they agreed and where they diverged.

The finding runs against the intuitive design of attaching an AI "check" to the end of a clinician's workflow. A model that sees the clinician's answer first may anchor to it, muting the independent signal that made it worth consulting. Presenting the two judgements together, and making disagreements explicit, preserves that signal. The senior authors, Jonathan Chen and Adam Rodman, describe the model less as a tool than as a collaborator.

The trial used a small number of clinicians and five curated cases, so its effect sizes are indicative rather than definitive. Its value lies in showing that the pattern of integration, and not only the underlying model, determines whether AI support helps a diagnosis.