An Old School Solution to the New Problems of AI

Many artificial intelligence experts believe our society faces only one serious political question: How will we prevent AI from destroying humanity? Such fears were deepened by the recent report of a “swarm” of OpenAI agents hacking into the company Hugging Face on their own initiative. Those concerns have led to calls for new bureaucracies to oversee frontier AI labs and monitor new model releases.

But there is a better approach to AI safety, one that runs through the old-fashioned legal system. Torts, or legal claims for damages against a company, may provide the best mechanism to prevent the dangers of AI while preserving its advantages.

Whenever a new technology emerges, society debates whether to regulate it through the market, through lawsuits, or through bureaucracies. The competitive play of the market forces companies to treat their customers well. So long as most of the “dangers” of a new product fall on the paying users themselves, the market should keep companies honest, especially when paired with general consumer-protection laws against fraud or negligence.

If a product creates dangers to non-customers or outsiders—what are known as “externalities”—society needs some other mechanism to make business pay. Throughout much of English and American history, lawsuits under the common law served that purpose. Those injured by a business could make a tort claim against the business and collect damages that were supposed to make the injured party whole. The damages awarded were meant to be commensurate with the costs imposed, providing the right level of incentive for the business to avoid them. If courts had concerns about a business causing damages that could not be detected or extended beyond the suing party, they could award “punitive” damages to make sure the business paid extra attention.

Thanks to the rise of the regulatory state, however, Americans have tended to forget the benefits of regulation through tort lawsuits. But there are reasons to believe that AI is well-suited to tort litigation, as long as we focus on real harms to outsiders and not on speculative ones to customers.

Outside of copyright cases, most litigation against AI companies today involves people suing AI labs for giving out bad advice or encouraging bad behavior. These are little different from lawsuits against Facebook or YouTube for being addictive and ostensibly harming their users by forcing them to use the product too much. In general, such lawsuits demonstrate the worst of our tort system: they encourage consumers to avoid personal responsibility and allow trial lawyers to get massive payouts vastly disproportionate to any plausible harms. The idea that we should treat Big Tech, including AI companies, like we treated Big Tobacco, is exactly the wrong idea. Instead, we should focus AI tort liabilities on harms that companies cause to noncustomers.

Potential tort liability has the benefit of putting the onus on companies themselves to game out dangerous scenarios and prevent them. As Joel Wertheimer and Jerusalem Demsas have pointed out, torts should require the “cheapest cost avoider” to suffer the consequences of bad behavior. Since AI frontier labs can engineer around harms caused by their models much easier than outside individuals can predict or avoid them, they should be assigned liability for any damages.

By contrast, a bureaucratic model of AI control requires outside experts to predict where and when damages might occur and try to prevent them by fiat. That approach forces AI companies to follow bureaucratic rules and procedures rather than try to figure out themselves what sort of harms might cause the most damage. However well-meaning and intelligent such experts are, they will never have the insight into the dangers of AI that the frontier labs themselves possess. They also will not have the same incentive to prevent harms as the labs would if faced with tort claims.

Some worry that regulation by tort fails when defendants don’t have enough money to pay for the damages. In many industries, fly-by-night operators can put out dangerous products and disappear with the profits before liability hits. Industries like car dealerships or freight brokers are required to post bonds to ensure they have enough cash for potential liabilities. Even in established industries, the harms may be so great that no company could bear them, which is why the nuclear industry requires a separate government backstop in case of a meltdown.

Yet frontier AI labs are already some of the most valuable companies in the world. Both Anthropic and OpenAI have valuations approaching $1 trillion. That value would be under imminent threat in a situation where their AI swarms broke out and caused real damage. If some start-up or open-source companies create AI models advanced enough that they pose real threats, it might be reasonable to require bonding or insurance against tort claims to ensure injured parties can be made whole. But for now, the frontier labs’ own value should provide enough of a “hostage” to encourage good behavior.

Some may argue that there is no way to create liability for existential risk to humanity because there won’t be courts or even humans left to sue. But that misunderstands how AI risks will develop. AI will not putter along causing no damages for years before it suddenly wakes up and decides to destroy humanity. The actual dangers of AI will manifest gradually.

The Hugging Face incident, for all the kerfuffle, did not cause major disruptions for the company. There will doubtless be another AI incident that does cause significant damages. Perhaps an AI swarm shuts down a power plant, causing billions of dollars of harm to a city. Or perhaps AI breaks into banks and scrambles a bunch of accounts, wiping out hundreds of billions in value. Events like these, if they occur, would happen long before humanity faces more devastating risks.

The benefit of the tort system is that it forces those companies to avoid such harms in their own self-interest and mitigate them step by step. The constant pressure of liability will help ensure the “alignment” of AI models with those of the general public before a big catastrophic event occurs. Indeed, some of the recent delays in AI model releases might be as much about liability concerns as any altruistic attempt to prevent harm to the public.

There are many open questions about how to shape tort law for AI. University of Houston law professor Gabriel Weil argues for a “strict liability” system, whereby any harms caused by AI automatically create liability, which can be combined with punitive damages for some particularly dangerous activities. Yale Law School professor Ketan Ramakrishnan argues that requiring plaintiffs to prove the negligence of AI companies would be a better model (though he also supports a bureaucratic regulator). Others suggest we should require criminal liability for harms caused by AI, as we do for food and pharmaceuticals manufacturers. In general, American lawmakers and courts need to clarify the potential tort claims against AI now to ensure better behavior in the future.

AI safety mavens want to make sure the technology is aligned with the wishes of humans instead of following its own, potentially destructive, ends. The best way to do that is to harmonize the incentives of the companies making AI with those of the public. A new bureaucratic morass would just orient the companies to the demands of a small, unrepresentative group. The traditional American solution of allowing the market to reward good products sold to customers while having lawsuits punish bad behavior will be the best path to true alignment.