The Hammurabi solution to AI

Not that Republicans seem to have much of an AI agenda either. This Sunday, House Speaker Mike Johnson called for, “a meeting in Washington with AI platform providers and key experts to determine the right course forward and discuss the responsibility providers have to ensure the safety of their products.”As politicians in Washington try to figure out how to regulate an industry they don’t understand, they may want to look back in time for a regulatory model. Like way back. Like almost 4,000 years ago. The Code of Hammurabi, promulgated in ancient Babylon around 1750 B.C., contained no modern building department, permitting bureaucracy, or hundreds of pages specifying approved construction materials. Instead, it placed responsibility for structural failure directly on the builder.“If a builder builds a house for someone, and does not construct it properly,” Law 229 provided, and the house collapsed and killed its owner, “that builder shall be put to death.” If the collapse destroyed property, the builder had to replace what was destroyed and rebuild the house at his own expense.Those punishments are barbaric by modern standards, but the underlying regulatory idea has a modern analog: strict liability. Instead of the government trying to determine beforehand exactly how a possibly dangerous product should be regulated, the law can place the cost of failure on those who create and deploy it.University of Houston law professor Gabriel Weil argues that this approach is particularly well suited to AI.The problem with top-down government regulation is that regulators must decide before a product reaches the public which risks are acceptable, which precautions are adequate, and which products should be approved. That works better when regulators understand the technology, its risks are reasonably measurable, and methods for mitigating those risks are relatively stable.AI satisfies none of those conditions.AI capabilities can advance dramatically from one model generation to the next, often in ways even their developers did not anticipate. Researchers disagree enormously about the magnitude of the dangers. Developers themselves cannot always predict what behaviors will emerge from increasingly autonomous systems. An approval regime, therefore, asks government officials to answer questions even the people building the systems cannot confidently answer.As Weil puts it, ex ante regulation forces policymakers to make early judgments about “capabilities, harms, and precautions” while all three are changing rapidly. A Food and Drug Administration-style regulator could therefore be simultaneously too restrictive toward beneficial innovations and too permissive toward dangers nobody anticipated.Strict liability approaches the problem from the opposite direction.Government does not have to know which safety architecture an AI laboratory should use. Developers remain free to experiment. But if an autonomous system causes legally cognizable harm, liability attaches even when the developer claims to have exercised reasonable care.That changes the economics of AI safety.A company deciding how much to spend on testing, monitoring, containment, cybersecurity, or human supervision would have to account for the expected costs its systems could impose on everyone else. Risks that might otherwise be externalized onto innocent victims become costs borne by the enterprise creating the risk. Weil argues that this preserves incentives for innovation while giving developers powerful incentives to discover safety measures regulators could never anticipate.The more controversial question is where that liability should stop.Under ordinary corporate law, officers are not personally responsible for corporate torts merely because they hold senior positions. They only become personally liable when they themselves participate in direct wrongful conduct. Corporate office alone generally does not create vicarious liability.A far more aggressive AI liability regime could change that rule by statute. Congress could specify that, for designated autonomous systems, strict liability attaches not only to the corporation but also to executives, and perhaps even engineers with actual deployment authority, responsible for releasing them.WHY MADISON AVENUE WANTS YOU SINGLESuch a strict liability regime would most likely slow down AI advancement. It could encourage excessive caution and deter talented people from working on advanced AI. Those concerns would have to be weighed against the opposite problem: allowing the humans making possbily catastrophic decisions to capture the financial benefits while placing much of the downside risk on shareholders, insurers, victims, or society.Hammurabi understood something Washington too often forgets: Accountability can regulate behavior more effectively than bureaucracy.