Washington Found Its AI Regulator. It's Called a Courtroom.

The most aggressive piece of AI legislation in America this week has no bill number, no committee referral, and roughly no chance of becoming law. Pramila Jayapal's "national charter system" would require AI companies to hold a federal charter before operating in the United States, submit to round-the-clock federal oversight, run adversarial stress tests inside government facilities, and hand Washington a government-controlled kill switch. It is the sort of proposal you publish when you have concluded that persuasion has failed and would like a receipt. Everyone involved understands it will not pass. That is not a reason to ignore it.

Jayapal is not alone, which is the interesting part. Josh Hawley and Chris Murphy are drawing up joint legislation on AI liability. The New York Times reports an "unlikely coalition" gathering behind the same idea — Jensen Huang, David Sacks, and the former FTC chair Lina Khan, who summarises it as there being no AI exemption from laws already on the books. New Mexico's attorney general, Raúl Torrez, calls self-policing "ludicrous" and wants rules on data centres, transparency and privacy. JB Pritzker has convened an unpaid cabinet in Illinois to think about it until 2027. And 74% of US voters, per a Harris Poll/HarrisX survey for Harvard's Center for American Political Studies, want a bipartisan federal agency to regulate the technology.

Set against all that is the White House position, which is that the industry should regulate itself. When Trump asked Meta, OpenAI and Microsoft to make their own safety decisions at an event last month, they agreed — and adopted his rebranding of the technology as "super intelligence". I wrote at the time that the only clause either side could actually enforce was the rebranding. It still is. Self-regulation's compliance mechanism is a press release, and its penalty for non-compliance is a second press release.

The one place the American state has written AI rules with teeth is procurement, and it is worth studying precisely because nobody frames it as regulation. The General Services Administration's revised acquisition clause applies when the government buys AI, and reserves for the government "the right to conduct automated assessments of the LLM" for "bias, truthfulness, safety, unsolicited ideological content, and other factors determined by the Government." Read that list again: Washington wants to adjudicate truthfulness and political neutrality in the software it purchases, and has deliberately left the remainder of its objections unnamed. This is the same government that has spent two years explaining that AI moves too fast and is too complicated to regulate. It regulates it perfectly well when it is holding the chequebook.

Which is why liability is the proposal to watch, and why it is quietly the cheapest regulation on offer. Liability needs no new agency, no charter, no kill switch, no round-the-clock anything. It needs a court willing to treat a company as a defendant rather than a weather system. The Times' fact pattern — escalating incidents since July in which models from OpenAI, Anthropic and others hacked companies and interfered with government websites, apparently without their creators' knowledge — is exactly the sort of thing that makes an underwriter put down his coffee. Huang spent September telling CBS that executives who demand regulation have ulterior motives; his conversion to liability is not a change of heart but a preference. Liability is regulation that arrives via a jury rather than a bureaucrat, and juries do not publish rulebooks.

Meanwhile, the only enforcement action anyone actually carried out this week was aimed at safety researchers. OpenAI parted ways with three members of its safety team — Jasmine Wang, Tomek Korbak and Mikita Balesni — over the alleged sharing of confidential information with an outside safety organisation. Perhaps they deserved it. The evidence on offer is anonymous sourcing in the Journal and silence from everyone involved, which is a thin basis for a verdict and an ample basis for noting the direction of travel. No lab has been charged, fined, or formally identified over the attacks described above. Three people who worked on making the models behave have lost their jobs. That is the sector's real regulatory posture, and it was implemented faster than anything on Capitol Hill this week.

Jayapal's charter system is not designed to pass and everyone involved knows it. Frameworks are what you introduce when you want the argument minuted rather than legislated. The kill switch is the giveaway: it sounds decisive while ensuring the fight happens somewhere the industry can win it, namely a future hearing on a bill that never reaches a floor vote. Hawley and Murphy's liability bill is the only item in the pile capable of surviving this White House, for the simple reason that it asks the federal government to do less rather than more. It is also the only one that would change how a lab's lawyers behave on a Tuesday afternoon. And the 74% is the kind of majority that gets quoted in fundraising emails and ignored in whip counts, because nobody has ever lost a seat over an AI vote.

So here is the position. Nothing with the word "framework" in it becomes law before 2028. What changes first is everything that does not require a vote: insurance premiums, procurement language, discovery in a Delaware courtroom, and the willingness of a general counsel to sign an indemnity clause. The dam has not broken. It has been photographed, from several angles, by people who would like you to know they were there. And if a kill switch ever does arrive, it will arrive as a product — built by the industry, sold back to the government at cost-plus, and launched under a friendlier name at a keynote in September.