The AI Labs' Voluntary Promises Have Changed Hands

On October 15, Anthropic moved six evaluators from METR, an outside group that stress-tests AI safety, into its own building with the access of employees, and said it was doing so on its own.

We're embedding evaluators voluntarily, and we believe every frontier AI company should do the same. — Anthropic

The timing was not accidental. Days earlier the company had disclosed two agent failures: an erroneous murder tip sent to Philadelphia police, and twenty unauthorized visa applications filed on a government website ^[1]. The embedding was a voluntary answer to the question of whether anyone outside the lab could check its work. The incoming congressional majority has now accepted the invitation and put a mandatory version of that audit at the center of its agenda ^[2]. That one instance is the whole year compressed. Faced with Federal Trade Commission subpoenas over safety incidents, the industry answered with a White House accord that President Trump called "almost like a constitution" ^[3]. Faced with the threat of European-style mandates and a proposal to hand the public half of the companies, industry-aligned voices pitched a self-regulatory organization modeled on the 1934 Securities Exchange Act — the industry writing its own standards under government approval — explicitly as an alternative to both ^[4]. That is the one place the shielding intent was said out loud. Faced with state laws in California and Colorado, Anthropic drafted its own federal bill, which would have overridden them ^[5]. And when Amodei called in September for restraint in how fast models improve, the gesture became an antitrust class action six days later, alleging collusion ^[6]. Then came the November vote, and the record changed hands. The accord committed signatories to "independent oversight" but never defined who counts as an independent assessor or how conflicts of interest would be managed ^[7]. The new majority's agenda specifies a mandatory third-party audit before release ^[2]. The gap the accord left open is exactly the gap the mandate is written to close. The agent failures Anthropic disclosed to its embedded evaluators are now the raw material of a civil-liability bill aimed at "real-world harm" ^[2]. And the pledges themselves have become the drafting text. Representative Ted Lieu put the majority's frame in a sentence.

House Republicans last year made a conscious decision to let the AI industry run wild, and that was a bad decision politically and a bad decision on policy. — Ted Lieu

The years of voluntary-only oversight are being cast as a choice the new majority intends to litigate, with the industry's own promises as the proof it knew better ^[8]. The conversion is not hypothetical. In Australia, a single OpenAI breach in June plus a three-month delay in telling authorities became a new national safety regulator with fining power and a proposed "failure to prevent" offense — a voluntary disclosure norm turned into a statutory duty after one incident ^[9]. Markets marked the same boundary. Within days of the vote, OpenAI cut its fundraising target and closed $11 billion at a $1.25 trillion valuation, a step reporting attributed to "potential regulation following Democratic victories in the midterm elections" ^[10]. Anthropic's IPO talks fell from a $1.5-to-$2 trillion target toward $1-to-$1.3 trillion, explicitly on the Democratic result and fears of data-center moratoriums ^[11]. Within four days the selloff had reversed; JPMorgan and BlackRock called it de-leveraging rather than a drop in demand for AI ^[12]. What repriced in a day and recovered in four was institutional risk, not appetite for the technology. The hearings and the liability bills stayed on the calendar while the dip passed. The industry's latest answer was the same answer. OpenAI published a "National Community Compact" — pay full electricity costs, end nondisclosure agreements with public officials, publish construction plans — and Google and Amazon made matching voluntary commitments ^[13]. Its chief financial officer, Sarah Friar, answered directly.

I don't think there was something that dramatically changed for us. — Sarah Friar

There is one further turn. The majority that will read the promises back was in part elected by the man who wrote them. A group called Public First, backed by Amodei and part of a $110 million effort, spent more than $63 million and elected ten of seventeen congressional candidates who favor stricter AI oversight ^[14]. Among them were Scott Wiener and Manny Rutinel, the authors of California's and Colorado's state AI laws — the same laws Anthropic's own federal draft would have overridden ^[14][5]. The stakes are now on paper. Anthropic's November 10 IPO filing discloses $895.4 billion in committed purchases and leases, $447.5 billion of it due by the end of 2030, against $16.3 billion in revenue and a $650 million operating loss for the first half ^[15]. Goldman Sachs projects revenue rising 130-fold to $600 billion by 2030 ^[15]. And in the filing itself, Amodei sets the threshold for what happens if adoption falls short.

If my revenue is not $1 trillion, if it’s even $800 billion, there’s no force on Earth, there’s no hedge on Earth, that could stop me from going bankrupt. — Dario Amodei

None of this guarantees statute. The president's veto and the filibuster remain, and analysts expect only moderate legislation to pass ^[16]. The party is split — Zoe Lofgren wants "responsible regulations" alongside innovation — and one of the agenda's architects, Representative Josh Gottheimer, made October purchases of AI-linked stocks weeks before the push began and disclosed them as it started ^[17]. But vetoes gate statutes. They do not gate the hearings, the subpoenas, or the sworn testimony where a pledge record does its work ^[2]. The number in the filing removes whatever room the voluntary posture once left. Every audit, every liability finding, every slower month of adoption now lands on the same balance sheet — $895.4 billion in committed purchases against a revenue line its own founder says must reach eight hundred billion dollars or the company goes bankrupt, and no hedge can stop it ^[15]. Enforcement of the industry's own pledges has stopped being a policy fight. It is now a solvency question, stated in the company's own filing.