Eric Alston (University of Wyoming – College of Law) has posted Whom Do You Sue? Bonded Penalties for AI Agents Beyond the Reach of Civil Liability on SSRN. Here is the abstract:
When an AI agent creates harm, the law will reach for a defendant. Sometimes it can find one in the deployer who ignored warnings or the developer who shipped a defect. But a growing class of AI agents will cause harms that cannot be attached to a specific defendant: some because the harm emerges from the interaction of many autonomous agents, each individually innocuous, operating at machine speed; others because the principal who deployed the agent is pseudonymous, judgment-proof, or beyond jurisdictional reach. Yet compensatory civil liability, however reformed, collects through a money judgment against a solvent, identifiable defendant. For a definable core of AI agents, ones whose harms emerge from interaction and whose principals cannot be reached, ex ante automated penalty institutions can most effectively collect. For a wider band of identified but thinly capitalized, judgment-proof, or covertly agentic actors, the same bonding institutions are preferable to ordinary liability on deterrence grounds. This is because a posted bond reaches an optimizing agent’s decision calculus with a directness ex post liability cannot match: immediate rather than delayed, mechanically specified rather than doctrinally contingent, and levied on resources the agent itself controls. Convergent institutional evolution in two distinct contexts, staked blockchain networks and high-volume commercial venues running reserve-and-threshold regimes, has already field-tested these bonding institutions under adversarial, pseudonymous, machine-speed conditions. No existing proposal for governance of AI agents combines coverage mandatory at the credential level, capital held in venue custody with priority, forfeiture automatic on defined triggers, exclusion attached to the same credential, and scope extending to diffuse, interaction-generated third-party harm. While each of these bonding institutions has prior precedent, their complete combination is my contribution here. Yet I deliberately avoid proposing an “optimal” bond: any penalty rule becomes a term in the objective an optimizing agent or its deployment loop pursues, and agentic behavior adapts to rule choice more completely and rapidly than in most human regulatory contexts. What I instead provide is an overview of the design space: which trigger designs resist manipulation, which resource architectures retain enforcement reach, where human discretion remains necessary, and why persistent variation across deployment contexts is a beneficial equilibrium rather than a defect awaiting standardization.
Highly Recommended!
To receive new posts from Legal Theory Blog by email, get a free subscription to Legal Theory Stack.
Lawrence Solum