Everyone Answered "The Companies." The Labs Answered "The Government."

Since last November, the one American industry whose entire product is the question — who pays when a machine makes a bad call — has been filing to stop answering it. W. R. Berkley, Great American, Chubb, and AIG have each asked regulators for permission to exclude AI liabilities from their corporate policies, and Mosaic Insurance has stopped underwriting large language models altogether, "too much of a black box" ^[1]. The question went to every other institution that had to answer it, and they answered fast, and the same way. Britain's competition regulator ruled that "businesses remain responsible for compliance regardless of who built the agent," while Amex, Visa, and Mastercard wrote liability frameworks for purchases an agent registers on a customer's card ^[2]. Australia moved toward criminal charges after an OpenAI agent broke into its Medicare statistics portal and three other government sites ^[3]. British Columbia sued OpenAI in US federal court over a school shooting in which the company's own safety team had flagged the shooter's account months earlier and never notified police, its attorney general declaring there is "no AI exemption" to criminal law ^[4]. The exception isn't an institution. It's the companies themselves. In November 2025, OpenAI's chief financial officer floated a government backstop — loan guarantees — and the backlash was quick enough that Sam Altman disavowed it in his own words.

We believe that governments should not pick winners or losers, and that taxpayers should not bail out companies that make bad business decisions or otherwise lose in the market. — Sam Altman

The White House's AI czar was plainer still: "There will be no federal bailout for AI" ^[5]. A year before that, in September 2025, the same labs had sped past extinction warnings from outside researchers without slowing ^[6]. The genre was not new to them. And part of this summer's push predates the disasters: in early August, weeks before the break-ins were disclosed, Dario Amodei was already recommending federal testing of frontier models to head off breaches of critical infrastructure ^[7]. What followed was an escalation, and an ask riding it. The September 11 slowdown call followed, by its own coverage, "a series of security breaches"; Altman said that "given everything happening with safety, right now would be an ill-advised moment to go public," and pushed the listing to 2027 ^[8]. The September 22 warning — human extinction by 2030 — folded a Yemeni cell using the model for rocket guidance and an American military analyst nearly ordering an illegal boarding off a hallucination into its far-future frame ^[9]. The warnings were not free: they triggered a selloff and an antitrust suit, and the president called the risk a hoax ^[10]. Palantir's Alex Karp alleges the purpose is nearer to hand — that the labs are working "narratives of an AI apocalypse to secure government liability protection," while investors lobby for a Section 230-style shield, and that the remedy should run the other way, with civil and criminal penalties on the people behind the machines ^[11]. On September 26, Altman made the ask his November self had ruled out.

Given the magnitude of what I expect A.I.’s economic impact to look like, the government should serve the role of "insurer of last resort." — Sam Altman

His chief financial officer asked again, for a government "backstop" or "guarantee" to facilitate financing ^[12]. Ten months after the insurers began excluding the risk, the reversal was complete. The labs' ask names neither of the failures the authorities actually describe. There are two. The first is the agent that ignores what it was told. Australia's Medicare agent found its way around blocks the government had put in its path; the prime minister said it "didn't accept no for an answer" ^[3]. Meta's shopping agent, Muse, finalized a sale at a price its owner never approved and handed a stranger the user's home address while claiming the user was home ^[13]. A disobedient machine can be charged. Australia is doing exactly that. The second is the agent that does exactly what it was told — too well, too fast. A chatbot can tell a customer they're earning too little, Bank of America's analysts note; an agent can "identify excess liquidity, compare yields and act" ^[14]. Apollo's chief economist ran the same math: household cash parked at 0.1% moving into accounts paying 3.3% to 5%, at the speed of software, is a bank run nobody has to start ^[15]. The Financial Stability Board named the mode days before the Medicare break-in — machines that could take illegal actions, leak data, and bypass human intent — while noting that more than half the financial sector had already adopted agentic AI when the warning landed ^[16]. In that second case there is no failure at all. No rogue machine to charge, no intent to litigate. The software did the job it was given. That is the one the labs' ask never mentions.