Radiology Is not a Task: AI and the Epistemic System of Radiology

Artificial intelligence (AI) in radiology is often introduced as a tool but discussed as a replacement, conflating functional performance with epistemic substitution. This Perspective defines radiology as an epistemic system: the organized clinical infrastructure through which imaging observations become warranted, actionable, revisable, and accountable knowledge. Existing validation and governance frameworks explain how AI should be assessed and managed. The epistemic-system framing reorganizes these elements around a distinct question: how do algorithmic outputs acquire the authority and answerability required of clinical knowledge? Current AI lacks epistemic autonomy because its outputs remain dependent on institutionally maintained standards, corrective processes, and accountable authorization. Foreseeable multimodal and agentic systems may integrate electronic health records, prior examinations, pathology, outcomes, and workflow tools, thereby expanding their epistemic participation. Informational integration, however, does not itself confer authority to adjudicate conflict, revise standards, or assume responsibility. Accountability is distributed across radiologists, clinical teams, health systems, manufacturers, and regulators, but it must remain attributable. AI may replace tasks and reorganize roles; replacing radiology requires transfer of the epistemic functions and institutional authority through which outputs become trustworthy clinical knowledge.