Private Law Theory

ABSTRACT
Frontier AI labs should not have limited liability for torts. When a company commits a tort it cannot pay for, limited liability caps shareholders’ losses at the value of their shares and lets the rest fall on the public. In effect, this makes limited liability a subsidy for risks borne by others. Scholars have long argued the subsidy distorts firms’ incentives. We argue the distortion is uniquely severe for firms that train frontier AI models. Its size depends on the covariance between private return and social harm, and on the shape of the tail of harmful outcomes. On both counts frontier AI is an extreme case. By the industry’s own account, profit and risk strongly correlate, and the tail is fat and unbounded. A liability rule also has an informational advantage over substantive regulation because it does not require the state to understand the technology. We propose that all shareholders of frontier AI companies bear capped liability for torts in proportion to their holdings, with controlling shareholders weighted more heavily. We answer the principal objections, including the race with China, and note that state law permits firms to adopt this rule on their own.

Rauterberg, Gabriel V and Sanga, Sarath, No Limited Liability for AI (September 27, 2026).