AI Governance Under Political Constraints: The Case for Ex Post Liability

Abstract

Rapid progress in artificial intelligence confronts governments with a choice among governance instruments: prescriptive regulation, licensing and approval requirements, compute governance, corrective taxation, and ex post liability. This article analyzes that choice as a problem of comparative institutional performance under realistic assumptions about political actors, including rationally ignorant voters, concentrated industry interests facing diffuse and unorganized potential victims, regulators dependent on the regulated industry for expertise, and legislators who must assemble coalitions amid deep disagreement about the technology’s risks. Instruments that require ex ante political consensus on the magnitude of AI risk, including Pigouvian taxation as well as regulation, are unlikely to be adopted in effective form; where adopted, they invite capture, since licensing and approval regimes erect entry barriers that entrench incumbents and calibrate stringency to industry influence rather than social cost. Ex post strict liability is more robust to these pathologies. It requires no prior agreement on contested risk forecasts, because damages track realized harm; it decentralizes precaution choices to the actors holding the relevant technical knowledge; and it locates enforcement in courts, which are structurally harder to capture than agency rulemaking. This article defends liability’s primacy through systematic comparison across instruments, then classifies non-liability tools as supportive, complementary, or substitutionary, identifying the conditions under which each is warranted by liability’s own political and legal limits. It also offers a framework for evaluating proposals to weaken, preempt, or displace civil liability and applies it to pending state and federal legislation.