How to tell which AI access policy will backfire

A quarter of the time, raising the price of the legal door just sends people out the back one. RAND's new framework, Buying Time Against AI Proliferation (RRA4996-1), makes that ratio testable. Across RAND's simulated scenarios of AI access policy, roughly one in four backfires by redirecting adoption to ungoverned channels instead of suppressing it. The same finding reframes the policy debate. Export controls, licensing rules, and compute chokepoints work on the supply side; they change what gets built or shipped. Users choose on the demand side, by cost and convenience. When the governed channel gets more expensive but the open one does not, demand does the talking.

The diagnostic is three questions, and a reader can ask them of any pending rule. How sensitive are adopters to price across channels. Can policy actually raise the cost of ungoverned access, or only the governed kind. How costly is it for a user to switch between them. Two unfavorable answers usually mean the rule accelerates what it meant to stop, even if it looks restrictive on paper.

Once AI capability is valuable enough to users, the calculus flips again. Adopters will pay, route around, or smuggle nearly any cost to keep access, and the same instruments lose their grip. The pattern generalizes beyond this report. Any regulation that prices only one channel without pressuring the other is buying time against proliferation, not stopping it.