The accuracy–utility gap: Why AI accuracy does not guarantee clinical benefit in digital pathology

Abstract

Artificial intelligence (AI) in digital pathology has repeatedly achieved expert-level diagnostic accuracy under controlled validation conditions. However, its clinical translation—reflected in regulatory approval, routine adoption, changes in diagnostic decisions, or demonstrated patient benefit—remains inconsistent. This viewpoint argues that this discrepancy arises because diagnostic accuracy is a property of a model under specified conditions, whereas clinical utility emerges from the entire sociotechnical pathway in which the model is deployed. Building on recent work on standardization, system selection, institutional validation, and trial governance, we integrate these complementary layers into a continuous tissue-to-outcome pathway comprising six stages: tissue, image, model, workflow, decision, and outcome, with each stage explicitly linked to the claim it can support. Variability introduced at any stage may remain undetected and propagate downstream, potentially affecting clinical decisions and patient care. Accordingly, evidence generated at one stage cannot substitute for that required to support claims at subsequent stages. We therefore propose a stage-linked evidence map that identifies the principal risk and evidence required at each stage, while making explicit which stronger downstream claims remain unsupported. This framework enables reviewers and institutions considering adoption to identify where a study has addressed the accuracy–utility gap and where uncertainty remains. Evaluating AI across the complete pathway reframes the evidentiary requirements for clinical translation: high diagnostic accuracy is necessary, but insufficient, to establish meaningful clinical utility.