Paper proposes six-layer framework for governing agentic AI

A new paper by Shruti Rajagopalan examines regulatory challenges posed by AI agents that transact, publish, and act on external systems without contemporaneous human approval. The paper argues that legal personhood for AI is neither necessary nor sufficient for governance, shifting the focus from status to enforcement. It notes that for two millennia, nonhuman legal personality—from the Roman universitas to the corporation, Hindu idol, waqf, and river—has operated through human officeholders the law can locate, question, prosecute, and replace. Agentic AI inverts that design, exercising practical agency without legal status, sometimes with no identifiable human in a responsibility-bearing role. The paper sorts AI deployments into three categories: one firm builds and deploys the agent; developer and deployer are separate but known; and no identifiable developer or deployer. It stress tests each category against five liability doctrines—agency law, products liability, enterprise liability, negligence, and strict liability—finding each fails at different points in the third category due to the absent responsibility-bearer. The paper assembles an alternative from regimes governing aircraft, ships, drones, driverless cars, and motor carriers, proposing a six-layer stack of registration, identification, verification, financial responsibility, lifecycle traceability, and suspension to place a human back at the end of the chain.

What’s reported

Key figures

Sources: marginalrevolution.com