AI Nationalization
Abstract
Both Donald Trump and Bernie Sanders want to nationalize AI. So, it turns out, do many others—from bipartisan congressional committees to national security hawks to academic critical theorists. Should, then, the United States government nationalize frontier AI? Scholars have offered no clear answers. That is in part because no one even knows what “AI nationalization” means. Leading proposals for AI nationalization differ radically from one another. Does “nationalization” mean moving frontier AI production inside the government? The government seizing a majority equity stake in AI companies? Control-only “golden” shares? Non-voting preferred stock? Compulsory production under the Defense Production Act? Something else?
This Article makes two contributions to the debate on AI nationalization. First, it defines nationalization. Drawing on the economic theory of ownership, we show that “nationalization” bundles two separable entitlements—residual claim rights (who captures the surplus?) and residual control rights (who directs the actions that no contract or statute anticipates?). The question of nationalization then becomes: Which residual rights should be held by the government? We show that existing nationalization plans target totally different rights.
Second, the Article uses this framework to argue that the government should hold a “halt right” vis-à-vis frontier AI companies. Under our proposal, the government could order frontier AI companies to temporarily stop the training or deployment of certain powerful AI systems. The halt right would be narrow in scope, allowing halt orders only to mitigate two serious dangers from frontier AI: catastrophic risk and “hard” corporate power. But the right would be highly discretionary, giving the government substantial latitude to determine which AI systems pose those risks. Such discretion is characteristic of residual control, especially the control afforded by European-style golden shares. It distinguishes our halt right from previously proposed regulatory and licensing regimes.
For other major risks from AI (national security failures, monopoly power, inequality), nationalization is not warranted. Ordinary regulation and taxation are better options. In general, we argue that when risks are either contractible ex ante or remediable ex post, ordinary tools are superior. We also argue that our halt proposal largely avoids the two major risks of stronger nationalization plans: stifled innovation and government concentration of power.
Read the full piece at SSRN.