Accountable Clinical AI Requires More Than Accuracy

Abstract

Large language models are approaching specialist-level performance in selected clinical tasks, but accuracy alone does not establish readiness for clinical deployment. This commentary argues that accountability, rather than raw performance, is now the central barrier to adoption. Recent evidence shows that clinically deployed large language models can perform radiology workflow tasks with high accuracy, yet important governance questions remain unresolved, including the provenance of inputs, the auditability of outputs, and the verification of downstream decision pathways. The present commentary proposes that accountability infrastructure should become a routine focus of clinical AI evaluation alongside performance metrics. Distributed ledger and related audit technologies may offer one practical framework for tamper-resistant logging, verification, and oversight of model-mediated clinical decisions. Clinical studies should therefore report governance architecture in addition to accuracy, and medical education should treat prompt engineering as an operational clinical competency. The next phase of clinical AI is not merely accurate systems, but accountable ones.