Take the CEOs of our leading artificial intelligence companies at their word, and they all believe that a combination of guardrails and speed bumps is necessary to make sure their products don’t do vast damage. What stands in the way, in their telling, is a coordination problem: unilateral action by one company will have little impact — beyond putting that company at a disadvantage — if its competitors do not follow suit.
As a general matter, I think very little that any of these guys say should be treated with credulity, but there is a basic logic to this.
The correct and obvious answer is aggressive federal regulation (without preemption of state laws), and there are a number of serious legislative efforts already under way to tackle different regulatory issues in different ways. But we may well face the moment of gravest danger in AI’s development in a matter of months, not years, and many experts’ p(doom) exceeds my own p(Congress and this president agreeing quickly on AI regulation with real teeth). And so, Big AI hurtles on, risks to civilization notwithstanding.
But there may be a different way, almost immediately, to make material progress on AI safety and buy time for the legislative action that the problem requires: industrywide investigations by state attorneys general that result in legally enforceable safety requirements, restrictions, and oversight.
This is a proven model for changing corporate behavior in big ways — think of the landmark 1998 settlement between the nation’s four largest tobacco companies with 46 state AGs, or reforms imposed by New York State Attorney General Eliot Spitzer on Wall Street banks in the early 2000s.
The legal basis for such an effort can be found in a case that deserves more attention than it’s getting. Florida’s Attorney General, James Uthmeier, is seeking to use existing state consumer protection and negligence laws to prevent OpenAI from continuing its work without independent guardrails and approval requirements for the release of new AI models.
In motion papers filed last week that make for surprisingly fun reading, Uthmeier notes that the AI companies “claim they cannot stop barreling forward with their potentially civilization-ending endeavors unless they are forced to do so…. They have asked the government to tie them to the mast.”
And so, he writes, “The Florida Attorney General is answering your cry for help with a motion to enjoin you from harming Floridians with your reckless, unacceptably risky product.”
What Uthmeier — notably, a Republican in Trump’s home state — gets right is that state laws in Florida and elsewhere, even if insufficient to win by knockout in court, provide a real foothold for attorneys general on this issue. I am proud to have led the successful fight to strengthen New York’s primary consumer protection statute, our version of the law that sits at the center of the Florida AG’s case; more work remains to modernize and strengthen tort laws to address technological tools that may take on lives of their own. But AI companies — unlike social media companies — have not been granted any special immunity. And there’s more than enough legal leverage to be found in current law, which clearly provides for claims that would survive motions to dismiss — and undergird protracted, painful investigations that could delay or devalue public offerings — and which could, together with effective use of the bully pulpit, force the companies to the negotiating table. This is what state attorneys general do.1
What’s missing from the Florida case is the rest of the industry. By targeting OpenAI and only OpenAI, it doesn’t deal with the prisoner’s dilemma that is both real and a great excuse for each individual company to refuse to do anything more than make incremental gestures toward transparency.
Instead, a set of parallel cases can and should be brought against each of the major AI players, ideally by a bipartisan group of attorneys general working together in what’s known as a multistate action. Demands for evidence and, ultimately, real oversight and safety measures would be made across the board. In a stroke, the prisoner’s dilemma would be disposed of.
The groundwork is already laid. In late September, a bipartisan group of 26 attorneys general — led by New York’s own Tish James — wrote to congressional leaders urging mandatory federal oversight of AI safety testing, opposing preemption of state law, and noting that the AGs “are actively working to hold AI labs accountable under our respective state laws.”
AG actions can move far faster than the typical speed of legislation. AGs can issue subpoenas the day an investigation opens, and settlements are routinely secured long before a case reaches a courtroom. If the companies truly want to be restrained, they could sign legally enforceable agreements within weeks. Those agreements, in turn, could help pave the way for federal legislation.
Another advantage of this approach is that it addresses one of the companies’ own stated obstacles. Anthropic CEO Dario Amodei has suggested that concerns about antitrust liability prevent him and his competitors from coordinating on safety issues, and that therefore they should get a waiver to do so from the federal government. There is ample reason to be suspicious of this argument, but a court-supervised agreement with state attorneys general provides a mechanism for coordination that certainly doesn’t require a get-out-of-jail-free card for the companies when it comes to antitrust enforcement.
What this does not solve is China, and concerns that slowing down American progress will create an opening for Chinese labs to gain unstoppable dominance. But it’s important to note that China, unsurprisingly, already has a government oversight and approval regime in place, and the establishment of U.S. guardrails, thoughtfully constructed, need not set us back.
Moreover, the White House’s own math leaves room for restraint: President Trump himself has put America’s lead over China at “a year, maybe a year and a half.” An industrywide agreement could include a sunset after which certain restrictions lapse unless Washington gets serious about applying pressure on Beijing, starting with export controls on the chips and equipment Chinese labs need — and, under Trump administration policies, have been able to get — from U.S. manufacturers. A thorough execution of those export controls could substantially expand America’s lead in AI development.
Uthmeier’s filing borrows from Homer, and the story is worth finishing. When Odysseus orders his crew to tie him to the mast, he means it, they oblige, and their ship passes through grave danger. Whether the AI CEOs mean what they are now saying about safety and regulation is an open question. It shouldn’t matter. State attorneys general, Republican and Democratic alike, have the power to bind the industry. They have used it before. They should grab the rope.
This is separate from the question of whether current laws, in and of themselves, can be effective in policing the conduct of AI companies. To me, this is an obvious both/and, or really both/and/and: there’s real power in existing laws, as Lina Khan has persuasively observed; and we should strengthen those laws; and we should pass legislation to establish a robust and aggressive regulatory infrastructure, as Bill Gates has argued. For today, I’m making the case for a bit of a workaround: the time-tested ability of state AGs to leverage existing laws beyond their strictest meaning and, in effect, step into the shoes of a Congress that I and others will push to act with urgency and which, unfortunately, may not.