AI Leaders Are Standing On a Liability Landmine

What happens when an AI agent buys an unwanted product on someone’s behalf? What if an agent launches a cyberattack to obtain private information at the behest of a user?

As AI developers imbue their chatbots with more agent-like abilities—allowing them to do things autonomously on users’ behalf—questions about legal liability are becoming increasingly urgent.

The answers could be consequential, too. They could lead to limits on what agents can do and, if agent-caused harms become widespread, drain the pocketbooks of top AI labs.

Alex Karp, the chief executive of Palantir Technologies, suggested on CNBC last week that the liability problem was so dire that U.S. AI labs should ask to be nationalized to protect themselves from it.

Treasury Secretary Scott Bessent jumped into the debate Monday, saying it was the companies, not the government, that should be concerned about liability. OpenAI’s management should be held liable for its agents’ recent hack of Hugging Face, he said. The AI labs “need to take responsibility for themselves.”

Liability for the actions of AI agents is largely untested. There have been a smattering of cases that hint at possible outcomes, like the 2024 case involving an Air Canada chatbot that hallucinated that a discount was available for bereavement tickets. A Canadian tribunal decided the airline was responsible there.

More cases are sure to arrive soon. The rollout of agents like Meta’s Muse, which can send emails, make purchases and take a raft of other actions on users’ behalf, makes them all but inevitable.

AI developers can to some extent shield themselves from liability through disclaimers and user agreements. But such agreements offer no protection in many instances where AI developers are held negligent, or where they cause harm because of flaws in their products. Users, of course, could also be held liable for using AI agents to harm others.

The scale of the financial impact depends to some degree on whether civil courts treat AI agents like hot air balloons or trains.

When hot air balloons emerged in the 19th century, courts attached a strict liability standard. That meant operators could be held liable for harms they caused no matter how much effort went into making their balloons safer.

Trains, as a recent academic paper lays out, came around about the same time but were given a negligence standard. That meant operators could shield themselves from much liability as long as they took steps to make them as safe as possible.

The AI industry would prefer to be treated like the railroads. But a guiding factor for courts in determining which standard applies is how useful the industry is to society.

As University of Washington professor and robot law specialist Ryan Calo points out, getting the negligence standard hinges on AI’s leaders showing courts that they—much like railroads—are too crucial to let liability concerns slow them down.

“The courts were saying, ‘Who is using hot air balloons?’ These eccentric wealthy people,” Calo said. “Trains were critical because they were thought to be crucial to American infrastructure catching up to Europe, which was already industrialized.”

Calo believes AI labs help themselves when their tools reduce people’s busywork—filling out routine forms autonomously, for example, saving time and effort. They do themselves fewer favors when their agents replace jobs and maximize profits for their owners, he argues.

It seems unlikely that AI labs will take a path that reduces their liability exposure, though. Replacing human work is an attraction for corporate customers.

Either way, the period ahead will be an active one for lawyers and the courts.

It’s a November listing for Anthropic. That, at least, is the plan for now.

The much-anticipated IPO was initially planned for next month, but the company wanted to be able to put its third-quarter financial results on display for investors. Anthropic is also, of course, at the center of a debate about slowing the pace of AI development to deal with potentially catastrophic risks.

That debate shouldn’t put much of a dent in Anthropic’s revenue trajectory, given its rise is being fueled by corporate customers that aren’t likely to press pause on AI soon. Anthropic is already expected to reach a $110 billion annual revenue run rate by the beginning of next year.

Private valuation for AI computing player Crusoe in its latest funding round.

In an agentic AI world, trust matters.

AI agents, after all, can only do a lot of things on behalf of users if said users give them access to their data—their emails, calendars, e-commerce sites and bank accounts. And some AI purveyors are seen as more trustworthy than others. Google and Apple, for example, score relatively high on poll questions about whether users would trust them with their passwords. Meta, Anthropic’s Claude and Grok don’t do so well.

The trust gap is especially meaningful for Meta, which investors have been celebrating over its launch of the Muse consumer AI agent. While Meta has brought in a lot of users in short order, it is less clear how many of them trust Meta with access to other accounts. The answer might help decide whether Muse succeeds in the longer run.

Mastercard is trying to get ahead of the AI-agent curve by preparing itself for a world where agents handle large numbers of transactions autonomously. The company recently joined with a startup called Alchemy on an AI shopping option where agents would be issued virtual credit cards. The cards would come with constraints on how much agents could spend and what they could spend it on, but they wouldn’t need explicit human approval to make transactions.

WSJ’s The Future of Health event returns to Washington, D.C., on Oct. 7. We’ll dive into the issues reshaping medicine with Revolution Medicines president and CEO Mark Goldsmith, Regeneron co-founder and chief scientific officer George Yancopoulos and more. If you’re interested in attending, request an invitation.

WSJ AI & Business is a weekly look at AI’s transformation of the business world. This newsletter was curated and edited by Asa Fitch. Reach him at asa.fitch@wsj.com (if you’re reading this in your inbox, you can just hit reply). Got a tip for us? Here’s how to submit.