OpenAI accepts AI rules from any venue that stops short of gating a model release

Anyone who ships a model, or ships a product on somebody else's, wants one thing out of a week of policy noise: a straight answer on whether any of it moves a launch date. On the record so far, it does not. OpenAI's limit is the same in Sacramento and in Washington. It accepts evaluation, reporting, standards and embedded reviewers, and it stops at a gate on release ^[5]. Sam Altman has told Congress to fund AI testing and not to require approval before a model ships ^[4].

Ashley Gold argued in Semafor that the industry is applying the lesson of the social media backlash, and that shaping regulation beats insisting you are harmless ^[11]. The industry stops agreeing with itself on venue. Gold reports the reasoning behind Lehane's phrase: going to statehouses is easier than waiting for Congress, and the companies can live with a series of similar state rules ^[17]. The super PAC money is buying the other outcome ^[2]. The Next Web argued that both positions can be sincere and are simply not aimed at the same result ^[12].

Congress has turned preemption down twice, and the Senate's version of no was 99 votes to one ^[6]. The pressure then moved to instruments that need no floor vote. An order signed last December set up a litigation task force to challenge state AI laws and directed Commerce to look at withholding broadband money from states with rules it considered onerous ^[7]. The litigation task force and the broadband review are both administrative levers ^[16].

The objections to all this do not require believing the safety concern is fake. Vance's Trojan horse remark is an argument about incumbent cost: a large firm absorbs compliance that a smaller rival cannot, so rules written with incumbents can entrench them ^[10][19]. Trump called the case for slowing AI down a hoax ^[9].

There is also evidence against reading the safety talk as theater. An Anthropic researcher resigned in public warning that the leading labs are gambling with people's lives, and two disclosure failures at OpenAI preceded the calls for pacing ^[14]. Anthropic committed to embedded external evaluators without waiting for anyone to require it, and OpenAI said it would match the commitment ^[15].

A team writing its own release process ends up sorting every proposal into two columns. The first holds obligations you satisfy after launch with paper: evals, incident reports, published standards, an outside reviewer with a desk in the building. The second holds obligations that make a date contingent on somebody else's sign-off. The first column costs headcount and delays nothing. The second is the only one that moves a ship date, and it is the one the money is pointed at ^[2][4].