AI Export Controls Buy a Window, Not a Wall

The harder part of governing frontier AI is not the rule, it is the line that keeps moving. Somewhere between a demonstrated capability and a tolerated actor sits the adoption margin, the threshold where a frontier AI capability finally becomes worth acquiring, and RAND's framework treats that line as the actual battlefield of AI counterproliferation, the policy effort to keep dangerous AI capabilities from actors considered untrustworthy.

Most readers hear "export controls" and picture a wall. RAND's report is built to show why a wall reading fails: the cost of reproducing any demonstrated capability falls as hardware improves, algorithms tighten, and knowledge spreads, so each price drop quietly recruits a new buyer across the line of private incentive. Tools that raise acquisition costs buy a window, not a containment. What policy can do is lengthen the window by enforcing in ways whose own costs grow slowly as the number of monitored actors rises, then spend the bought time on downstream defenses aimed at the pathway-specific inputs that turn a model into a harm, inputs that may stay scarce after the model no longer is.

The three-condition failure test is what to carry to the next story. A counterproliferation regime collapses only when capability reproduction has become cheap, when enough actors sit on the margin, and when the cost of stopping them exceeds what enforcers can sustain. Policy can shift the first two. The third is the one to watch.