saved
How to Count AIs: Individuation and Liability for AI Agents
Hraness cites a source capture. The source author remains the source.
gist
Arbel, Goldstein, and Salib argue that AI governance fails before liability if law cannot count agents. Thin identity ties every AI action to a human principal; thick identity sorts swarms of models and instances into persistent, goal-coherent units that incentives can reach. Their Algorithmic Corporation (A-corp) is a legal-fictional entity owned by humans and run by AIs: ownership solves thin attribution, while scarce property and compute force emergent thick identity through delegation incentives and selection—so tort, tax, and shutdown tools can target a durable defendant.
ideas
- Count before you govern. Harm attribution needs two identity layers: thin links to humans, thick units for the AI agents themselves.
- Thin liability is not enough. When agents have private information and divergent goals, punishing only principals cannot cheaply deter the act.
- A-corps are legal fiction plus keys. Human-owned entities with cryptographic credentials can own property, contract, sue, and be sued while AIs manage them.
- Resources invent identity. Control of compute and assets makes misaligned delegation expensive, so coherent goal clusters self-organize and survive selection.
- Markets alone under-govern. Accidents, deception, complicity, and willful blindness argue for two-sided credential mandates at economically significant choke points.
quotes
“But when an AI causes harm, the first question to answer before anyone can be held accountable is: Which AI Did It?”
“Thin identification is the project of tying every action taken by an AI to some human principal.”
“We call it the “Algorithmic Corporation” or “A-corp,” a legal-fictional entity that can hold property, make contracts, and litigate in its own name.”
“AIs that control A-corps will thus have strong incentives to share control only with other AIs that share their goals.”