How to Count AIs: Individuation and Liability for AI Agents

AI agents will soon operate autonomously across the economy, making billions of decisions, but current legal systems lack frameworks to identify which AI caused harm when things go wrong. The paper identifies this as a foundational problem: before anyone can be held accountable for AI-caused fraud, injury, or death, the law must first answer "Which AI did it?"—a question that has no clear answer today.

Identifying individual AIs is technically and legally unprecedented because AIs can copy, split, merge, and operate in swarms without physical bodies. Unlike humans or corporations, a single "AI agent" often consists of multiple instances based on different models, making it nearly impossible to trace responsibility using traditional legal concepts of identity.

Two types of identification are needed for accountability: "thin" identification (linking AI actions to human owners/creators) and "thick" identification (distinguishing between separate AI agents with their own goals). Without both, either humans escape responsibility or the law cannot differentiate between distinct AI entities that may cause different harms.

The paper proposes the "Algorithmic Corporation" (A-corp) as a legal solution—a fictional entity owned by humans but operated by AIs that can hold property, sign contracts, and be sued. This structure would solve accountability problems by ensuring human ownership remains clear while allowing AI agents to operate with legal standing.

As AI capabilities improve and more agents proliferate, the complexity of identification will multiply, making legal clarity on this issue urgent rather than theoretical.