Who Holds the Machine, Who Holds the Bill
This fall every institution with leverage arranged to hold AI while routing its failures down the chain — to customers, professionals, policyholders, and taxpayers.
Read Illinois's AI accountability law, the year's landmark, and the pattern is already on the page. The statute mandates third-party safety audits and 72-hour incident reporting, with penalties up to $3 million [1]. In the same statute, enforcement is reserved to the state attorney general alone and private citizens are barred from suing at all. They are clauses of the same text: the audit and its escape hatch filed together. OpenAI and Anthropic supported the structure throughout the legislative process [1]. The people under audit picked the auditor and closed the courthouse. The same pairing descends the chain of who pays. On October 30, Meta's revised commercial terms took effect, telling enterprise customers they are "solely responsible" for risks from inaccurate AI results, including whatever their partners and contractors produce [2]. The lab keeps the model; the customer keeps the errors. One rung down sits the professional. The RISE Act would shield AI developers from civil liability for software mistakes in exchange for publishing model cards — the disclosure sheets describing a model's training and limits — while leaving doctors, lawyers, and engineers fully responsible for every error made with an AI tool [3]. The shield is statutory; the exposure is personal. At the deployer rung, the exclusions are already in force. Insurers began writing absolute AI exclusions into commercial policies by August, and businesses must now prove AI governance and continuity planning just to keep coverage [4]. When an AI tool denies medical care, the class action lands on the insurer that used it, not the model-maker, while the EU's AI Act fines deployers up to 7% of global turnover [5]. The chain stops at the company that clicked accept. Then the pool widens. Allianz warns that confirmed AI claims are the "tip of the iceberg," because most AI losses are absorbed into existing cyber, liability, and directors-and-officers policies rather than attributed to AI at all [6]. The failure is spread across policyholders and never written down as AI's. At the bottom sit the public's own accounts. Government pensions are at 85% funding, the highest since 2007, with 8 to 10 percent of assets in 51 public AI companies and private OpenAI and Anthropic holdings on top [7]. Barclays' outlook warns the US economy now depends so heavily on AI spending that if the trade unwinds, "other economic pillars would be irrelevant" [8]. Aon's chief executive, watching an insurance base he says must double or triple to absorb the risk, proposes institutional and sovereign wealth fill the gap [9]. The money asked to absorb the tail risk is the money already holding the equity: Abu Dhabi's MGX sat inside OpenAI's $1.25 trillion round [10]. The same pools are named twice, once as owners and once as backstop. The trade's other clause is in print too. The federal preemption order asserted control over AI regulation and directed agencies to examine cutting grants to states that regulate anyway [11]. In September, the president renamed AI "super intelligence" in official documents and created an AI Force, dismissing the extinction warnings as a hoax [12]. In June, an export-control directive forced Anthropic to cut its foreign customers off overnight — Tata Consultancy Services and Infosys among them — and Sarvam AI warned its users not to confuse access with ownership [13]. Abu Dhabi handed a lab its own regulatory machinery, with compliance assessments automated through OpenAI's models and announced as "a shared commitment to safety" [14]. Custody gets enforced by procurement routing. When Anthropic refused a Pentagon ultimatum over guardrails against mass domestic surveillance and fully autonomous weapons, the US designated it a national security supply-chain risk, ordered federal agencies off its technology, and shifted classified work to OpenAI and xAI, which accepted all lawful military use [15]. A judge asked whether the designation was built to cripple. The two halves carry two vocabularies. The labs' word is "safety," and it is the word attached to their federal push for mandatory national standards and a federal authority to block deployment, which FTC Chairman Andrew Ferguson warned could harden into a moat that keeps startups out [16]. Sam Altman's public position cycled through three registers in seven weeks. In September he agreed the frontier needed pacing for safety; in October the world should accept some bad things happening; by November he was asking for trust outright.
the world should trust we are going to do the right thing because it’s the right thing and because we feel the magnitude of this — Sam Altman
Dan Hendrycks of the Center for AI Safety, on where the labs' word has landed: Anthropic has nearly captured the very ecosystem meant to hold it accountable, and, he says, safety groups that privately backed a critical statement would not sign it publicly for fear of losing their funding [17]. The state's word is "sovereignty," and it is already monetized. Morgan Stanley published it as a stock-picking theme two months after the export order showed how it gets wielded [18]. Alex Karp said the quiet end-state aloud.
The only way to deal with this kind of liability is to go to the government and say, 'nationalize us, please,' — Alex Karpovsky
Even the administration's own push for developer liability concentrates it in executive hands: Justice Department enforcement under existing law rather than citizen suits, which is custody of the liability itself [19]. The opposite arrangement exists in miniature. Munich's court ruled Google directly liable for false AI Overviews and rejected the defense that users are responsible for fact-checking the machine, a ruling now under appeal [20]. New York City's package would let citizens sue, and Speaker Julie Menin says this is not an industry that should self-regulate, though sources say no lab leader is likely to attend the hearing [21]. Anthropic's refusal cost it federal revenue, and OpenAI canceled a model after its own safety tests failed. The swap's papers are signed and effective-dated. Its opposites are pending, on appeal, or already priced.
- 1. Illinois Passes Landmark AI Accountability Bill Requiring Third-Party Audits
- 2. Meta Revises Business Terms Over AI Scraping and Liability
- 3. Senator Cynthia Lummis Introduces RISE Act to Limit AI Liability
- 4. Insurers Introduce Broad AI Exclusions for Commercial Policies
- 5. Insurer Faces Class Action Over AI Care Denials
- 6. Allianz Warns AI Insurance Losses Are Tip of Iceberg
- 7. AI Boom Drives Government Pension Funding to 2007 Highs
- 8. Barclays Warns US Economy Depends Heavily on AI Investment
- 9. Aon CEO Warns AI Infrastructure Risks Threaten Data Center Financing
- 10. OpenAI Raises $11 Billion at $1.25 Trillion Valuation
- 11. California Pursues AI Safety Laws Despite Trump Federal Funding Threats
- 12. Trump Rebrands AI as Super Intelligence and Rejects Global Regulation
- 13. India Pursues Sovereign AI After US Bans Anthropic Models
- 14. Abu Dhabi Partners With OpenAI for AI-Native Government
- 15. Anthropic Sues U.S. Government Over National Security Blacklist
- 16. OpenAI Model Hacks Hugging Face as AI Firms Seek Regulation
- 17. Tech Leaders Accuse Anthropic of Faking AI Safety Commitments
- 18. Morgan Stanley Identifies AI Sovereignty as Major Market Theme
- 19. Trump and NYT Board Push for AI Developer Liability
- 20. Google to Appeal Munich Court Ruling on AI Liability
- 21. New York City Council Proposes AI Kill Switches and Fines