When the Rules Died, the Labs Bought the Politics

In February, Meta put $65 million into two brand-new super PACs. One was built for Republican races in Texas, the other for Democratic races in Illinois, and both aimed at the same single purpose: stopping state-level AI legislation ^[1]. The money wasn't picking a party. It was picking a topic. That one purchase is the whole year in miniature. Binding rules for AI died in Washington this year, three deaths in close order. In May, Trump canceled his own planned executive order on the technology, deciding he did not want to do anything to "get in the way of that lead" ^[2]. In September, over a private dinner, he dismissed Dario Amodei's plea to slow the technology down as a "hoax" ^[3]. And this month he signed what was left: a "morally binding" voluntary accord with six companies — a four-layer safeguard system with internal monitoring and outside auditors — on his stated premise that "regulation and governance inherently impede innovation" ^[4]. The companies police themselves and stay liable for their own harms ^[5]. So the labs started building toward the people who might one day write the rules. The two Meta-backed super PACs are the template ^[1]. A separate group, Leading the Future, has pledged $100 million to push a single federal AI framework meant to override restrictive state laws ^[1]. Money also moves the other way already: Federal Election Commission records show campaigns have spent at least $17 million on AI vendors since 2020, with Claude in 65 campaigns and committees, mostly Democratic, ChatGPT in 80 campaigns, the Republican Party the largest single buyer of OpenAI services, and Republicans favoring xAI and Campaign Nucleus ^[6]. Both parties are already customers. Anthropic opened a presidential engagement program this month, along with an employee PAC, to brief both parties' 2028 candidates on AI policy — timed to managing a strained relationship with an administration that designated the company a supply-chain risk in March ^[7]. Around the same time, the Pentagon banned Anthropic's models ^[8]. Amodei's answer was accommodation, not confrontation.

if we work with the president, we can win safely — Dario Amodei

The government was building permanent doors of its own in the same stretch of days. Jay Clayton, the new AI czar, toured Silicon Valley to open a "direct line of communication" with the leading developers of advanced AI ^[9], and the labs pledged $2.4 billion to the administration's Genesis mission for AI-driven scientific research ^[10]. The channels now run in both directions. But the sharpest move of the year was quieter, and it happened inside the labs themselves. In February, a wave of safety researchers walked out as commercial pressure mounted: Mrinank Sharma, who led Anthropic's safeguards research team, resigned warning "the world is in peril," and Zoë Hitzig left OpenAI over ads in ChatGPT, saying the company was building "an economic engine that creates strong incentives to override its own rules" ^[11][12]. This month, three OpenAI safety researchers were fired after an internal investigation found they had shared sensitive information with an outside evaluator ^[13]. OpenAI denies firing anyone for raising safety concerns; it says the three broke policy on handling sensitive information ^[14].

To be clear, talking to METR was my job. — Tomek Korbak

The honest counter sits in the data. Safety hiring is up 91 percent this year, with salaries up to $325,000 across the labs, but the growth is in enforcement and forensics, the roles that make a release deployable, while the people who question releases left, or were pushed out ^[15]. The safety function did not disappear. It was reorganized. And the safety argument itself packed up and moved. The existential warning survived as positioning: Anthropic's $2 trillion IPO prospectus carries Amodei calling AI "the most important global security issue facing the world today" ^[16]. Anthropic is now briefing candidates in both parties ^[7]. It even lobbied the Vatican to recognize AI consciousness, a bid Pope Leo XIV swatted down: "algorithms lack the spark of humanity" ^[17]. The emblem of the whole emigration was a single afternoon. On October 6, Sam Altman told one room of US leaders that the world should "accept some bad things happening" for the technology's benefits, then told a City Council hearing the same day that he agrees with Amodei that "we need to pace the frontier" ^[5][18]. The safety argument no longer changes what gets built. It now decides which room it is said in. Outside those rooms, the people who most wanted rules have the fewest instruments. Ninety-one percent of Americans want guardrails on AI, and 81 percent say safety should come before innovation even if it slows the technology ^[19]. A petition demanding safety off-switches has gathered 925,471 signatures, and a national walkout on October 20 will call for a freeze on frontier development ^[20][21]. Senator Alex Padilla put the objection in a lawmaker's own words.

It’s important that the rules for AI be set by elected representatives accountable to voters, informed by industry and by independent experts, researchers, and affected communities, and not by whichever company has the best-resourced political program. — Alex Padilla

The asymmetry of instruments is the whole story now. One side has a petition and one afternoon of walking out. The other has standing super PACs, a presidential engagement program, an employee PAC, and vendor contracts with both parties, with 2028 already on the calendar ^[1][7][6]. The fight over how AI gets governed did not end when the rules died. It became a campaign-finance one.