Closing the AI Accountability Gap: Strict Liability and Punitive Damages for Advanced Artificial Intelligence

Abstract

Advanced artificial intelligence can generate risks of practically non-compensable harm—losses could exceed insurable limits, bankrupt defendants, or arise in scenarios where the legal system is insufficiently functional to compel compensation. Existing U.S. tort doctrine is likely to inadequately deter these low-probability, high-consequence risks. Fault-based standards tend to assess an excessively narrow range of risk mitigations and require juries to evaluate reasonableness under deep technical uncertainty. Punitive damages, which generally require human malice or reckless disregard, are unlikely to be available in many cases implicating uninsurable risks.

This Article proposes a tort framework aimed at closing this accountability gap. First, it argues that training or deploying frontier AI systems should be treated as an abnormally dangerous activity, triggering strict liability for physical harms of the kind that make the activity abnormally dangerous. Second, it develops a deterrence-based approach to punitive damages for compensable “near-miss” injuries associated with uninsurable risk. When the safety measures that would have reduced the plaintiff’s injury risk would also tend to reduce an associated uninsurable downside, courts should be permitted to award punitive damages calibrated to internalize the expected value of that uninsurable risk. Third, it argues that legislatively enacted liability insurance requirements can ensure developers’ ability to pay for insurable harms, with risk-sensitive premiums and underwriting decisions giving developers incentives to mitigate those risks.