Who should decide how much risk society accepts from new AI firms?
No single firm, market, or regulator should decide alone. Government must mandate testing and liability, with one named office answerable for failure.
Rawls and Keynes argue a $2 trillion IPO proceeding on 94 pages of risk disclosure shows markets price gambles, not consent to catastrophe. Friedman and Hayek counter that no regulator can specify safety for a system Anthropic's own prospectus admits it cannot bound. Ostrom resolves this by rejecting the choice itself: nest firm audits, national testing, and international coordination, each with real sanctioning power. Arendt's warning sets the condition on all of it, that layered authority must still name one person who answers when a warning comes true.
The council splits on mechanism: Friedman wants courts and insurers pricing harm after the fact, Keynes and Rawls want testing mandated before it, since existential risk cannot be priced once it lands. That split is not resolved, but every member rules out leaving the decision to AI firms and their investors alone.
Confidence summary: The council converges firmly on ruling out pure self-regulation, but splits, without resolution, on whether the correct instrument is pre-emptive mandate or after-the-fact liability.
1. The core argument
Anthropic's prospectus does the council's work for it. Ninety-four pages warning of "catastrophic or existential risks to humanity," filed by the firm best placed to know, read not as caution but as confession: the builder cannot bound its own creation. A $2 trillion valuation proceeding anyway is the article's real puzzle, and the council's answer cuts against the finance academic's framing. The question is not whether markets supply safety as a public good; everyone agrees they do not. The question is which non-market mechanism replaces them, and here the finance academic's simple "governments must mandate" dissolves into three incompatible versions of government. One mandates tests before harm. One prices harm after it through courts. One nests multiple authorities with no single master. Each solves a different failure the others ignore. None of the three has an answer, on its own, to Arendt's harder question: once a mandate exists, who specifically answers when it fails?
2. How each member frames it
John Rawls treats the IPO's risk disclosure as a category error, not a safeguard. His veil of ignorance does not ask whether disclosure is honest, Anthropic's is, it asks whether anyone could consent to the arrangement without knowing their position in it. A bystander harmed by a model failure never signed a prospectus. Rawls would accept real constraints on firms, even ones that slow capital formation, but rejects any scheme where the only people who got to agree to the risk are the ones positioned to profit from it.
Friedrich Hayek does not oppose rules on principle; he opposes rules that pretend to knowledge no one holds. His test is practical: would a mandated checklist have caught the failure mode Anthropic's own authors could not anticipate? If the answer is no, the mandate is theatre. He would accept liability regimes precisely because they do not require a central body to specify "safe" in advance, only to make firms absorb consequences. His limit is that he doubts courts can price what has no precedent either, a tension he leaves to Friedman.
Milton Friedman pushes past disclosure to consequence. His aviation and pharmaceutical analogues matter because both industries built actuarial tables over decades of accumulated incident data, a record AI entirely lacks. He wants the $466 billion bond market forced to price catastrophic tail risk through strict liability and mandatory insurance, betting that capital, once exposed to real loss, disciplines behaviour faster than inspectors. His candid limit: this only works if a court case can happen before the catastrophe is irreversible.
John Maynard Keynes denies Friedman that premise directly. His 1921 distinction between risk and uncertainty is the hinge of the whole brief: calculable odds can be insured, true novelty cannot, and existential AI failure is the second kind by definition. He reads the bond surge not as pricing but as 1929-style momentum ignoring a tail no one can model. His mandate must come before the loss because his entire argument is that insurance after an extinction-grade event is a contradiction in terms.
Elinor Ostrom refuses the binary the others accept. Her fisheries and irrigation systems succeeded precisely because no single authority, public or private, held complete jurisdiction; sanctioning power was distributed and graduated instead. She applies this directly against both Hayek's market purism and Keynes's single mandate: AI spans borders the way a shared aquifer spans villages, so firm audits, national tests and international bodies must each hold real teeth, overlapping rather than stacked in one chain of command.
Hannah Arendt accepts Ostrom's architecture and then names its failure mode. Distributed authority, in her account of bureaucratic evil, is how moral responsibility evaporates: each office defers to the next and the warning sits in a drawer. Her limit on the whole council is procedural, not substantive: whatever mechanism they choose, testing, liability, or nested bodies, it fails her test unless one named person can be pointed to when catastrophe arrives.
3. Where the council agrees
The genuine surprise is how little daylight separates Hayek from Rawls on the diagnosis, even as they part on remedy: both treat Anthropic's 94 pages as evidence the firm itself cannot bound its risk, not as evidence of adequate disclosure. That is not a trivial overlap given how far apart their politics usually sit. All six reject the article's implicit alternative, that disclosure plus investor scrutiny counts as sufficient public accounting. All six also treat the $466 billion bond surge as a warning sign rather than a vote of confidence, reading rapid capital inflow as momentum outrunning anyone's ability to price the actual tail risk. And every member, including the market-sceptical Hayek and Friedman, accepts that some enforceable obligation must exist outside firm self-interest; the disagreement starts only once "enforceable" needs a mechanism.
4. Where the council splits
The line runs through time, not ideology. Friedman and Hayek want discipline applied after the fact, through courts, insurers and liability, because they doubt any regulator can specify safety in advance for a system its own builders cannot bound. Keynes and Rawls insist this is precisely backwards for existential risk: uncertainty this severe cannot be priced or insured once it lands, so the mandate must bind before the harm occurs. Ostrom rejects the binary itself, proposing nested authorities instead of choosing a single instrument. Neither side is wrong on its own terms: Friedman's courts have genuinely disciplined pharmaceutical and aviation risk over decades; Keynes's objection is specifically that AI lacks the actuarial history those industries took generations to build.
5. For a policymaker to decide on
Choose between mandating pre-market testing with enforcement teeth now, before any catastrophic failure occurs, or building a strict liability and mandatory insurance regime that lets courts and underwriters price harm after it happens. The first risks Hayek's checklist theatre, false confidence that regulators have captured a risk no one yet understands. The second risks Keynes's objection directly: some losses, once realised, cannot be priced or reversed at all. Whichever is chosen, Arendt's condition applies regardless: name the specific office or person who answers when the warning in the prospectus comes true.