Shareholders Should Pay for AI Dangers

Dario Amodei of Anthropic has warned that, within a year, swarms of artificial-intelligence agents could take over the internet and do hundreds of billions of dollars in damage. His biggest rivals, Sam Altman and Elon Musk, agreed. Suppose they’re right and the damage is done. Who pays?

Under current law, you do. If a bankrupt company can’t pay its debts, shareholders aren’t liable and the people it harmed have nowhere to go.

This rule is called limited liability. It’s so familiar you might think it’s in the nature of any corporation. But it isn’t. National banks didn’t have limited liability until the rise of deposit insurance in the 1930s. Neither did California corporations of any kind. American Express traded actively without it until 1965. Sole proprietorships and general partnerships have never had it.

Limited liability for torts is a subsidy to today’s shareholders paid for by tomorrow’s tort victims. For AI, that subsidy should end. That means if an AI company faces a mass tort and runs out of money, its shareholders should have to pay.

Limited liability for AI companies isn’t always a problem. For lenders and other contract parties, limited liability is merely a term in a deal. It means that if the company can’t pay, the lenders can’t go after the owners for the rest. A bank that lends to an AI company knows the rule and can price it into the loan. If it wants more protection, it can ask the owners for a personal guarantee.

But tort victims never made a deal. A hospital hacked by rogue AI never chose to do business with the company at all. That’s why limited liability shouldn’t apply to torts.

Most of the time we tolerate limited liability because mass torts are rare and we don’t know they’re going to happen in advance. Had asbestos manufacturers announced in 1950 that their product might kill hundreds of thousands, no legislature would have said their owners should have limited liability. The heads of frontier labs have said their companies could be responsible for such a mass tort, or worse. They’ve made the case against limited liability.

That case, as we argue in a recent paper, is strongest when a corporation’s profits and the risks it poses rise together, and when those risks are extreme and unbounded. That’s exactly the situation AI executives describe. Customers demand the capabilities that make the product so dangerous. The more successful the product, the more risk likely lands on the public.

Building more-capable models without these risks is the engineering problem that labs say they need more time to solve. Congress isn’t about to solve it. But Congress doesn’t have to. It can establish a liability rule and let incentives do the work.

In the 19th century, locomotives threw off hot embers that set nearby fields and barns ablaze. Lawmakers didn’t engineer a technical solution. They and the courts made the railroads pay for the damage. The railroads responded by fitting spark arresters to their engines and clearing brush along the tracks. Once the incentive to do so was in place, the companies supplied technical fixes.

Shareholder liability can be narrow and still work. It need only apply to the handful of firms that train frontier models. Passive investors can have their liability capped, so index funds can still hold AI stock and price the risk. Controlling shareholders such as Mr. Amodei and Mark Zuckerberg should pay without a cap, since by their own admission they control the risk.

The predictable objection to such a policy change will be China. For now, America is leading the global race for AI leadership. Anything that slows the home industry, so the argument goes, only helps the other side.

That objection assumes that Washington has so far been hands-off. In reality it’s already up to its elbows. The government decides which chips may be sold to China, which labs get Pentagon contracts, and which labs get blacklisted. Even if the rule did handicap U.S. firms, a powerful Washington lobby would have reason to push for imposing the same terms on Beijing’s AI champions. Washington could make Chinese access to American chips conditional on adopting the same rule.

The U.S. can go on pouring capital into AI research and data-center construction. It can extend cheap loans, tax credits and whatever else it likes, even in exchange for giving up limited liability. But it shouldn’t extend a subsidy that pushes AI risks onto everyone else.

Anthropic is about to go public. Before it does, it could set an example and waive limited liability for torts on its own; state law expressly allows that. Anthropic is already a public-benefit corporation, with a majority of its board appointed by an independent trust. If its chief executive thinks the industry should slow down, this is how he can put his money where his mouth is.

Mr. Rauterberg is a professor at Columbia Law School. Mr. Sanga is a professor at Yale Law School.