Want AI agents to stop going rogue? Make executives personally liable.

“If such an animal escapes and does damage, the owner is liable simply on proof that he kept it,” so wrote Oliver Wendell Holmes, Jr. in his 19th-century legal treatise “The Common Law.” Holmes cites the ancient principle that simply keeping “a tiger or bear or other animal of a species commonly known to be ferocious” can make a person individually responsible for the harm it causes regardless of care or negligence. 

In other words, the owner of such an animal might build a strong cage, inspect it every day and take every reasonable precaution, but if the animal escapes and causes injury, damage or death, the owner is still liable. In legal terms, this is referred to as “strict liability.” 

As society considers how to prevent rogue AI agents from shutting down power grids or triggering war, this idea could make the real difference.   

On one side of the AI safety debate, Jensen Huang, founder of Nvidia, argues that existing corporate and product liability laws should be sufficient to deter companies from releasing harmful products. “If your product is not ready to ship, don’t ship the product,” said Huang in a recent interview. 

Product liability lawsuits, however, are lodged against the firm; not executives, and corporate directors and officers are often indemnified against personal loss through insurance. In short, the world could face catastrophic harm at the hands of a few executives who bear no personal legal liability for their actions. Also, the firm’s investors are protected by limited liability, so their liability is capped at the amount of their investments.

Companies racing to stay ahead of competitors have little reason to not release potentially dangerous products. Under pressure to maximize profits and market valuations, that’s what we can expect them to do.

Bill Gates and others make the case for a federal agency that would require AI firms to be licensed, and through oversight, enforce transparency and safety. No doubt this would help, but a fundamental information asymmetry remains. Firms will always know far more about the risks and uncertainties of their products than the government does, and even the firms themselves don’t fully understand the products they are rushing to release.

Strict personal liability directly restrains the actions of executives making risky decisions. Congress should pass a law making all individual directors and officers of companies building frontier AI models strictly liable, in civil and even criminal terms, for the harms their companies’ products might inflict.

As Holmes put it: “The possibility of a great danger has the same effect as the probability of a less one, and the law throws the risk of the venture on the person who introduces the peril into the community.” AI firms are introducing great perils into our society. They should be liable for the harms that come from them.

There is precedence for this approach. In 2002, soon after the collapse of the dot-com and telecom bubble, much of which had been driven by massive earnings fraud, President George W. Bush signed the Sarbanes-Oxley Act. He called it “the most far-reaching reforms of American business practices since the time of Franklin D. Roosevelt.” 

Under the Sarbanes-Oxley Act, legal liability was put squarely on executives if they, even unknowingly, certified false financials or deficient processes that caused material errors in reported results. CEOs and CFOs can personally face fines of up to $5 million and prison sentences of up to 20 years under the act. After it passed, hundreds of companies were forced to restate their earnings, and many executives were charged with fraud. The veracity of earnings reports improved dramatically.  

Imagine if Sam and Dario and Elon and Mark, along with their boards of directors, bore individual liability for the damage their companies may inflict on society. Imagine if they faced possible prison time for the consequences.

The AI race to market now underway might proceed more cautiously. And those rogue AI agents would be kept on a leash.

Hans Taparia and Bruce Buchanan are professors at New York University Stern School of Business.