AI in radiology, five years on: what actually changed

In 2021, a widely-shared essay predicted that radiologists would be automated out of existence within a decade. Five years on, the number of practising radiologists in the United States and United Kingdom has grown, not shrunk. Their productivity, however, has changed materially.

The lesson from radiology is not that clinical AI failed. It is that clinical AI diffuses as a productivity layer inside existing professional workflows before, and often instead of, replacing them. This is a well-understood pattern in labour economics; it was underweighted in the initial wave of AI-in-medicine forecasts.

Three concrete changes are visible in modern radiology practice. Triage models flag suspected large-vessel occlusions and pulmonary emboli within seconds of image acquisition, shortening time-to-treatment in stroke and PE care. Reporting assistants pre-populate structured report fields. And workflow orchestration tools route studies to the best-matched sub-specialist across a network.

None of these individually is a paradigm shift. Together they represent a genuine, measurable improvement in throughput at a moment when radiology demand is growing faster than the workforce.

The implication for the next wave of clinical AI — in pathology, cardiology, dermatology, and ophthalmology — is that the automation-first framing continues to over-promise and under-diffuse. The productivity-layer framing is a better predictor of where the technology actually lands.