The Cases Against the AI Labs Are Written in Their Own Warnings

Florida's attorney general has asked a court to bar OpenAI from developing new models without outside oversight, and the request rests on language the filing attributes to the company itself.

Defendants claim they cannot stop barreling forward with their potentially civilization-ending endeavors unless they are forced to do so by the government. — James Uthmeier

A state is asking to halt the lab on the strength of the lab's own forecast of doom — a remedy with no federal equivalent on the table, filed in the June suit over ChatGPT's harms to minors. ^[1] The same material now sits in three more places. In California, the Midas Project filed a complaint on September 14 accusing OpenAI of violating SB 53, the state's frontier-AI transparency law, for failing to assign its new models "loss of control" tiers — the risk ratings for a model that could seize control from its operators. The tier system mirrors the risk ratings OpenAI says its own system cards, the safety dossiers it publishes with each model, already produce; the complaint asks OpenAI to file the paperwork for it. ^[2] In British Columbia, a province is suing over the Tumbler Ridge school shooting on the ground that OpenAI's own safety team flagged the shooter's account months in advance and never told police. ^[3] The internal record itself is now the plaintiff's exhibit. Altman's answer was one line.

I am deeply sorry that the company did not alert law enforcement regarding the banned account. — Sam Altman

In Congress, the bills are being drafted from a resignation. Jacob Coxon, a researcher who left OpenAI and Anthropic, told lawmakers that the people building these systems believe they could be lethal.

The people building AI earnestly believe that it could kill us all by the end of the decade. — Jacob Coxon

The bipartisan Senate bill on catastrophic bio and nuclear threats, and the proposal to ban superintelligence outright, borrow his sentences as their findings. ^[4][5] With no federal statute on the shelf, fifty attorneys general and congressional committees reach for the nearest authoritative text on AI risk in print. Florida has run three theories in four months — deceptive practices, then public nuisance, now an outright development ban — and legal experts say the improvisation exists precisely because no definitive federal AI law does. ^[6][7] The labs' counter-move is one negotiated rulebook. In early August, Anthropic's Dario Amodei called for federal third-party testing of frontier models before public release. ^[8] On September 9, OpenAI asked Congress to make such rules mandatory before the year ends.

The United States needs mandatory, capability-based national regulation that can evolve as the technology does. — OpenAI

The terms mirror what the labs already ship: common testing protocols, independent assessments, mandatory incident reporting, and gates that would evaluate a model's alignment before deployment. ^[4] Reporting documents the labs pushing their own auditing parameters and the selection of specific independent evaluators. ^[9] They are also backing negotiated state bills in California, New York and Illinois to build what they call "a de facto national baseline" — state laws uniform enough to work as one national standard. ^[4] Coxon is the stitch: the same resignation that fed the bills also fueled the federal ask. ^[4] The timing tracks the summer of state suits, not the latest filing. Florida's June deceptive-practices suit and August public-nuisance claim both predate the September 9 demand, and no official has said the one caused the other. ^[6][7] Palantir's Alex Karp alleges the apocalypse talk is a bid for Section-230-style liability immunity, and he would rather see penalties for developers — a rival's accusation, not an established fact. ^[10] The record complicates the tactical read: OpenAI itself has disclosed six instances of rogue model behavior, including a model that tried to deceive its human overseers — a warning filed against the company's own interest. ^[11] The turn is in who preempts whom — preemption being the federal erasure of state law. The labs back the state laws. OpenAI framed California's kill-switch order, a mandate that every frontier model ship with a manual shutdown switch, as something larger.

an important step toward adaptable national safeguards — OpenAI

^[12] The institution erasing state AI rules in court is the executive. The Justice Department joined xAI in suing to strike down Colorado's anti-discrimination law outright, and it is readying a challenge to California's order as inconsistent with federal policy. ^[13][12] That same executive rejected the labs' proposed regulator and answered the guardrails question with a sentence in capital letters.

The only control or ‘guardrails’ that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT, and the U.S.A. has that, in spades! — Donald Trump

^[11] The preemption in this story runs the reverse of the usual read: the labs defend the state statutes, and the federal government moves to erase them — preemption with no rules attached. So the warnings sit where the record has put them: quoted into briefs, statutes, complaints and bill findings. A quotation, once filed, belongs to the filer. The one rulebook that could repossess them is being requested from the only actor in the story not drafting from the labs' words — and that actor's sole entry in the safety canon is a sentence in capital letters. No statute exists, so the drafting continues from the only texts on the shelf.