The Economics of AI Counterproliferation

Once an artificial intelligence (AI) capability has been demonstrated at the frontier, the cost of reproducing it tends to fall as hardware improves, algorithms become more efficient, and knowledge spreads. A new RAND economic framework demonstrates what this means for governance regimes trying to keep dangerous AI capabilities away from untrustworthy actors.

In this analysis, actors weigh the value of an AI capability against the cost of acquiring it. The model traces how falling costs steadily enlarge the set of actors with private incentives to acquire a capability, how that growth widens the gap between private demand and the access that a regime is prepared to tolerate, and what determines whether the resulting enforcement burden eventually exceeds what the regime can sustain.

The implications for policy and strategy are that tools that raise acquisition costs, such as export controls or cloud access rules, buy a limited window of time for AI counterproliferation. That window of time can be extended by investing in enforcement approaches whose own costs grow slowly as the number of actors that must be monitored rises. The time bought should be used to develop downstream safeguards and defenses, focusing on the pathway-specific inputs that turn an AI capability into a harm, since these inputs may remain scarce and governable after the models themselves no longer are.