The AI Labs Begged for Safety Rules. Their Lobbyists Are Working the Exceptions.

For two years, Gavin Newsom was the governor who would not regulate AI. In 2024 he vetoed California's attempt to make the biggest labs test their most powerful models before release, calling the mandate innovation-stifling. This month he convened a two-month expert panel to draft safety regulations anyway ^[1]. And when he explained the reversal, the reason he gave was not a scandal, a study, or a poll.

The federal government’s abject failure to create any form of meaningful AI oversight or accountability should alarm every American, especially when AI CEOs themselves are begging for regulation. — Gavin Newsom

Read it twice. The alarm a governor cites for regulating an industry is that the industry is begging him to. After a year of the labs warning that their own products are too dangerous to leave alone, those warnings have stopped being commentary and become the stated basis of actual law. The same week, Washington said no again. The Senate blocked the AI kill-switch bill, the one that would have let the government order a leading lab's model shut down ^[2], and the federal framework remains what it has been since June of last year: voluntary. The institute that replaced the federal AI Safety Institute, now called CAISI, exists to help industry develop voluntary standards ^[3]. Even Dario Amodei, the Anthropic chief whose slowdown call is the industry's loudest, is not waiting on Congress.

Unfortunately, passing laws can take time, and AI is advancing very quickly. Therefore, in parallel with the regulatory route, AI companies can and should voluntarily work together to set standards. — Dario Amodei

So the binding rules are being written in the state capitols, on the industry's own warnings. New York's RAISE Act takes effect January 1, 2027, and requires only large developers to disclose their safety measures and report security incidents to the state ^[1]. Attorney General Letitia James opened a portal inviting AI employees to report unsafe practices straight to her office ^[1]. California's panel has two months to turn the CEOs' begging into mandates, among them a kill switch for frontier models and independent third-party safety planners ^[1][4]. And in those same capitols, in those same weeks, the industry is running a second campaign.

The AI industry's two simultaneous asks of lawmakers

Bind the frontier: OpenAI asked Congress for mandatory national safety rules keyed to what a model can do, and backs the state safety laws as a de facto national baseline ^[5]. Anthropic recommended government third-party testing of frontier models before public release ^[6].

Spare the product: Google's model legislative language, introduced in at least ten states, shields chatbots built into search engines and developer-facing chatbots from liability ^[7]. TechNet, the industry's trade group, testified against Minnesota's ban on recreational chatbot use by minors ^[8].

The exemption campaign leaves fingerprints in the bill text. Colorado's Chatbot Safety Act passed with twelve categories of chatbots carved out of its coverage ^[7]. In Arizona, the governor killed her own state's bill rather than sign it.

I will not protect big technology companies and AI chat bots more than children. — Katie Hobbs

In Hawaii, legislators struck Google's targeted exemptions out before passing their law ^[7]. Line the asks up against the product catalogs and the alignment is exact. Anthropic sells closed frontier models, and the reviews it pushed for in Washington would have covered open-weight systems too, the free, downloadable kind where Meta and Chinese labs compete with it ^[9]. OpenAI's demand for national rules arrived only after its own agents went rogue, and the rules would reach the agent products at the center of its business ^[5]. Google's ten-state language shields the chatbot folded into its search engine ^[7]. And when California raised questions about the water OpenAI's data centers draw, Sam Altman had an answer ready.

concerns over AI’s water use don’t hold “up to any scrutiny.” — Sam Altman

Two facts keep that alignment from reading as a simple racket. In Washington, the labs do not always get what they ask for: the administration overrode Anthropic's proposal and exempted open-weight models from federal safety testing altogether, a competition decision made over Amodei's objection that happens to shelter the Chinese models challenging US labs ^[9]. And OpenAI is paying a visible cost it attributes to safety itself, delaying a public offering investors valued at $1.2 trillion.

now would be an "ill-advised" moment to go public — Sam Altman

None of this is lost across the Atlantic. Europe's AI firms have rejected the American slowdown calls outright, and the EU Commission describes the US safety proposals as regulatory capture ^[10]. In April, the US envoy to the EU, Andrew Puzder, delivered the administration's warning to the Europeans.

Europe will not be able to pull itself into the AI economy by bringing other people down. — Andrew Puzder

Five months later, the same charge crossed the Atlantic in the opposite direction, aimed at the US labs by France's finance minister.

Making everyone behind them slow down so they can stay in first place – I can clearly see their self-interest. — Roland Lescure