The Burden Shift Is Not Enough: Rebuttal Technology and Deterrence in Artificial Intelligence Liability

Abstract

This working paper formalizes the evidentiary margin in artificial-intelligence liability. It models a two-stage game in which a developer chooses unobservable safety investment and harm realizes probabilistically, comparing claimant-burden strict liability, negligence, and strict liability with a rebuttable presumption of defect. The paper shows that shifting the burden of proof alone is exactly neutral when rebuttal is care-insensitive. Deterrence is restored only when the rebuttal standard responds to safety work actually performed. The record includes a corrected PDF and companion scripts reproducing the numerical tables and figures. This record is part of the Law as Extended Phenotype research program.