Visible reasoning: turning AI consultation into diagnostic learning

Abstract

Abstract AI consultation is now common in diagnostic work, but it is still judged mainly by whether it improves the immediate answer. In this paper, I propose that AI consultation can also support clinician learning when the clinician’s diagnostic judgment and the AI response are recorded in the same format and revisited after follow-up. I describe this proposed approach as visible reasoning, a two-stage process. In Stage 1, before viewing the AI output, the clinician records a brief account of the diagnostic judgment and a ranked differential diagnosis. The AI then responds in the same format, and the clinician compares the two records. In Stage 2, after meaningful follow-up becomes available, the clinician returns to the saved records to examine what mattered, where the two judgments converged or diverged, and what lesson and cue should be carried forward to similar cases, while recording the source and certainty of the final diagnosis or outcome. A worked example shows how differences in feature weighting can change the next action and generate a concrete, clinician-specific lesson. In this way, AI consultation becomes not only a decision aid but also a reusable record for learning.