Why isn’t Sam Altman under arrest?
The cofounder and chief executive of the artificial intelligence company OpenAI is free as a bird, even though software from his company has conducted illicit break-ins of multiple computer networks, including the software development firm Hugging Face and an Australian government health-care site. These acts would be criminal if committed by a human being.
AI agents from Anthropic, Google, and Meta have committed similar acts of digital breaking-and-entering. But no one’s been arrested, and so far, hardly anyone has even filed a lawsuit.
Why not? Because proving that AI creators are legally liable for their products’ digital misdeeds is a difficult job under existing criminal or civil laws. Legal analysts told the Globe there’s only an outside chance that a prosecution or lawsuit would succeed.
“There’s attorneys general who are suggesting all kinds of things, and some of those might work,” said Bahrad Sokhansanj, senior research fellow at the Boston-based Institute for Law and AI, “but it’s sort of trying out some new theories, because we’ve never really seen something like this before.”
Sokhansanj and others believe new laws will be needed to rein in AI abuses, because current statutes on the books were drafted for a different world.
Consider the Computer Fraud and Abuse Act, the federal statute that provides both civil and criminal penalties for hacking computer systems. Enacted in 1986, the law was drafted long before the creation of software that can attack a network on its own. And so it’s focused on deliberate human activity.
“These laws require intent. And they only apply to legal persons, and AI agents are not legal persons,” said Gabriel Weil, an assistant professor at the University of Houston Law Center. It can’t be used to punish OpenAI for creating rogue agents, in his view, because “the humans or other legal persons at OpenAI did not intend to hack.”
So far, just one organization has taken OpenAI to court over the Hugging Face hack. A New York nonprofit called Legal Advocates for Safe Science and Technology sued the company under California anti-hacking laws.
“It is a crime to hack into a third party’s computers without consent,” said the group’s programs director, Vivian Dong. “Over the summer, we saw AI agents engaging in that conduct and thought OpenAI and developers involved should be held liable for the illegal conduct of their agents.”
But it’s not clear if Dong’s group, which wasn’t directly affected by the hack, has legal standing to sue. Hugging Face itself has so far declined to file a lawsuit, saying it doesn’t want to pour its resources into a costly legal fight.
Besides, another recent California case shows how hard it will be to make an anti-hacking lawsuit stick.
The AI search company Perplexity makes an AI-enabled browser called Comet, which can purchase goods on Amazon without human assistance. Last year, Amazon demanded that Perplexity stop these automated purchases. And when Perplexity refused, Amazon sued.
A federal district court granted a temporary injunction against Perplexity. But in August, an appeals court threw it out, saying Perplexity itself hadn’t hacked anybody. The company merely created the AI agent and can’t be blamed for how other people use it.
The problem gets even more gnarly in the Hugging Face incident, where the agents hacked the network entirely on their own, with no humans urging them on. Based on the Perplexity precedent, a lawsuit against OpenAI seems unlikely to succeed.
Sokhansanj thinks what’s needed is a change to the federal antihacking law — an extra proviso to punish “reckless” activity. It would be analogous to laws against manslaughter, which punishes people for accidental homicides. If a rogue OpenAI agent was created in a reckless fashion, the company could be punished for losing control of its creation.
But there’s another, simpler, option: Why not treat rogue AIs as defective products, just as we do with cars or kitchen appliances? If a company makes a piece of software that runs wild, that company should be held liable for any resulting harms.
“Law enforcers already have authority to charge companies and their CEOs for creating and releasing dangerous, unvetted, or defective products,” former Federal Trade Commission chair Lina Khan wrote on X recently, in support of a crackdown on AI companies.
But is AI a “product”?
The European Union recognizes computer programs as products. But in the United States, courts have often treated software as “intangible goods” or even as “services,” allowing companies to evade liability for any defects.
AI companies have counted on this. In a spate of lawsuits filed against OpenAI by family members of people who have committed murder or died by suicide after interacting with the company’s chatbots, OpenAI has argued that bots deal in information and ideas protected by the First Amendment, and are not products in the traditional sense.
Some US courts are casting a skeptical eye on such arguments.
Last year in Florida, a child died by suicide after interacting with a chatbot produced by the company Character AI. In a lawsuit filed by the child’s mother, a federal court held that the AI chatbot software could be considered a product.
The company eventually settled the lawsuit, and with good reason. If its software was designated a product, it could have been subjected to strict liability, the legal standard that applies to dangerous or defective consumer products. It holds a company responsible for anything that goes wrong with its products.
Without strict liability, a plaintiff would have to prove an AI company had been negligent in making or deploying its product. That means proving that a reasonable person should have realized the risks and taken more care to avoid them.
But proving negligence could take months or years of complex and costly research. Weil, from the University of Houston, thinks strict liability is a better bet for protecting the public from rogue AI: The plaintiff simply has to prove the product had a defect that caused harm.
“A court could decide that frontier AI development and deployment is an abnormally dangerous activity,” he said, and “hold the developers of these models, of frontier AI models, strictly liable.”
Weil is putting his ideas into action. He’s working with Massachusetts Democratic US Representative Lori Trahan on the CLAIM Act, a bill that would hold developers liable if their AI acted in a way that would constitute criminal or negligent behavior if a person had done it. In effect, a bad-acting AI would be treated as a person acting on behalf of its creator. And that creator would face strict liability for making defective software.
There’s also a bipartisan proposal in the US Senate, the AI LEAD Act, that would impose a strict liability standard on AI services.
But Kevin Frazier, AI innovation and law fellow at the University of Texas Law School, worries that strict liability approach could threaten AI innovation. “It would be ill-advised to apply this to all kinds of AI,” Frazier said, “given that such an incentive might unduly hinder beneficial uses of AI out of fear of worst-case scenarios.”
Members of Congress have put forward dozens of bills related to AI during this congressional session. Apart from a law banning AI deepfake images for blackmail or sexual exploitation, hardly anything has been enacted, even as the heads of some major AI companies, including OpenAI and Anthropic, have called for tougher regulation of their industry.
Meanwhile, President Trump has disparaged calls for federal regulation of AI systems, arguing that the effort is a “SICK conspiracy.” The leaders of most major AI companies disagree and have urged federal action.
Last month, Trump and top AI executives met at the White House and issued a voluntary pledge to do a better job controlling their creations.
Trump called the agreement “morally binding,” but that probably won’t mean much in a court of law.
Hiawatha Bray can be reached at hiawatha.bray@globe.com. Follow him @GlobeTechLab.