Liability for AI Harms: How Ancient Law Can Govern Frontier Technology Risk, with Prof Gabriel Weil

Episode Summary

Executive Summary: The episode examines Gabriel Weil’s proposal to use liability law—especially negligence, products liability, abnormally dangerous activities, and punitive damages—to make frontier AI developers internalize third-party risks without heavy-handed upfront regulation. Weil argues tort law can adapt faster than legislation, but should focus mainly on misalignment and external harms, with complementary insurance and state-level rules to handle catastrophic tail risks.

Main Topics: Primer on liability law (Priority: 5/5): Weil explains the main tort doctrines relevant to AI: negligence, products liability, vicarious liability, and abnormally dangerous activities. He emphasizes how these doctrines allocate responsibility when harm occurs and how they might map onto AI development and deployment. Why liability may fit AI better than prescriptive regulation (Priority: 5/5): Because AI capabilities and risks are evolving quickly and uncertainty is high, Weil argues liability can flexibly incentivize safer behavior without requiring policymakers to predict exact technical standards in advance. Third-party harms vs user harms (Priority: 5/5): A central distinction is between harms to users/customers and harms to non-users or the public. Weil sees liability as most valuable for externalities—pedestrians, bystanders, and society-wide risks—where market feedback is weakest. Punitive damages and near-miss catastrophe deterrence (Priority: 5/5): Weil’s most novel idea is that punitive damages can punish companies for catastrophic risk they exposed society to, even if the actual harm was smaller, provided the incident could easily have escalated into an uninsurable disaster. Open-source, closed-source, and chain-of-responsibility questions (Priority: 4/5): The conversation digs into who should be liable across the AI supply chain—base model developers, scaffolders, deployers, and users—especially in open-weight systems where contractual privity is absent. State legislation and policy entrepreneurship (Priority: 4/5): Weil discusses bills in Rhode Island and New York that would make model developers strictly liable when an AI does what would be a tort if a human did it, absent intent or reasonable foreseeability by the user or intermediaries. Comparison with private regulatory markets (Priority: 4/5): The episode contrasts liability-based governance with private certification/MRO models, debating race-to-the-bottom risks, third-party protection, and whether insurance or market standards can substitute for direct liability.

Key Arguments: Negligence law can already apply to AI, but courts will likely ask narrow questions about specific precautions rather than whether frontier AI itself was a risky undertaking. Products liability is less clean for software because software is often treated as a service, and in practice its design-defect analysis often ends up looking similar to negligence. Strict liability is most defensible where the activity is unusually dangerous and the harm is the kind of harm that made it dangerous in the first place. Punitive damages can extend liability beyond actual harm when a small incident reveals that the defendant exposed society to much larger, uninsurable risk. Liability is especially suited to AI because harms are often easier to attribute after the fact than they are to measure ex ante, unlike climate change where taxes are more natural. Third-party harms create externalities that companies and their customers may not internalize; liability can make developers treat public risk like bottom-line risk. Weil argues that misuse and misalignment should be treated differently: misuse is often better handled with ordinary negligence or contractual tools, while misalignment justifies stronger developer liability. In cases where a model does something the user neither intended nor could reasonably foresee, Weil thinks the model developer should often bear responsibility. He does not want AI liability to overcorrect and block socially beneficial uses such as medical diagnosis, self-driving cars, or other cases where AIs may outperform humans. Liability should scale with the social benefits and harms of the underlying activity; if an AI activity creates large positive externalities, a blanket strict-liability regime could be counterproductive. Insurance can play a complementary role by pricing risk, demanding safeguards, and rewarding companies that implement cost-effective safety measures. The proposed state bills are intentionally narrower than broad AI regulation: they target tort-like harms from model behavior and carve out misuse and malicious modification in some forms. Private governance and liability-based governance are not mutually exclusive, but the strongest version of private certification becomes problematic if it shields third-party harms from suit.

Data Points: Negligence elements: 5 elements - Plaintiff must prove duty, breach, reasonable care failure, causation, and actual harm in negligence claims. Common-carrier analogy: Higher duty of care - Airlines/trains/buses are subject to negligence plus a heightened common-carrier standard. Constitutional classifier overhead: Mid-single-digit compute overhead - The host cites Anthropic’s classifier approach as costing a small extra percentage of compute while materially reducing bio-risk outputs. Risk reduction from classifier: About one order of magnitude - The host says Anthropic’s classifier approach appears to reduce bio-risky outputs by roughly an order of magnitude or more. Warning-shot calibration example: 10x - Weil explains that to internalize a $10T risk when only $1T is insurable, warning shots need to occur about 10 times more often than actual catastrophes. Threshold for illustrative catastrophic risk: $10 trillion - Used as a hypothetical high-end harm level to explain punitive-damages internalization. Maximum insurable risk example: $1 trillion - Used as the notional upper bound beyond which compensatory damages become inadequate. Bill coverage test: Tort if a human did it - The proposed state legislation would make developers liable if an AI does something that would be a tort if done by a human, and the user/intermediary didn’t intend or foresee it. State bill jurisdictions: 2 states - Weil says he is working with legislators in Rhode Island and New York. Certification regulator example: California Attorney General - In the private-regulatory-market discussion, the California AG was the proposed approver of multi-stakeholder regulatory organizations. Public survey framing: Broad support for doing something - The host notes surveys suggest public appetite for AI action despite disagreement over which policy tool is best. Frontier capability posture: 1 company planning to wait until superintelligence - The host references a company with a stated plan not to deploy broadly until superintelligence is achieved. Model ecosystem: Open weights and closed source - Liability allocation differs depending on whether there is contractual privity in the chain. Industry reference point: Waymo safer than human drivers - Used to illustrate a case where society may demand a higher safety standard for AI systems than for humans.

Pivotal Quotes: "“Why not use liability law to incentivize AI developers to properly consider and account for the risk that their development and deployment decisions are imposing on the rest of society?”" — Host/narration: The opening framing for the episode’s core thesis: liability as AI governance. "“I want them to treat risks to the public like risks to their bottom line and act accordingly.”" — Gabriel Weil: Weil’s concise statement of the incentive mechanism he wants liability to create. "“If an AI system does something that would be a tort for a human, then someone should be liable.”" — Gabriel Weil: Weil summarizes the core principle behind his state-level legislative proposal.

Implications: If Weil’s framework gains traction, AI companies would face stronger incentives to build, test, and deploy more cautiously—especially where non-users are exposed to serious harm. Liability may become a key governance layer alongside insurance, standards, and targeted regulation.