The debate over humanity’s extinction has acquired an impressive supporting infrastructure: research institutes, journalism fellowships, policy proposals and diagrams with enough arrows to evacuate an airport. The question of whether artificial intelligence will destroy civilisation now arrives with a second question: who is paying for the discussion?
Kevin Bass has supplied an answer of considerable dramatic ambition. In a viral X post, he describes an “Anthropic Network” linking investors, philanthropic organisations, the AI evaluation group METR and journalism supported by the Tarbell Center. His thesis is that rising Anthropic wealth finances warnings about AI, those warnings justify regulation, and regulation entrenches Anthropic. The machine allegedly cannot stop. Congress should investigate. Presumably before it develops another fellowship programme. The original post
There is a serious institutional question here. An organisation can be legally separate from a company while sharing its funders, professional networks and assumptions. Independence requires more than separate stationery. Nobody needs to telephone a researcher with instructions if the funding system consistently rewards certain questions, career paths and interpretations.
The same applies to journalism. Tarbell acknowledges that, as of 2025, most of its funding originated with Coefficient Giving, formerly Open Philanthropy. It also states that donors have no control over its journalism. Both facts matter. Editorial safeguards deserve consideration; financial concentration deserves scrutiny. A declaration of independence is a starting point for examination, rather than a force field around the newsroom. Tarbell’s ethics policy
Unfortunately, Bass’s argument repeatedly accelerates past what his evidence can carry.
Consider the $7.7 billion. The post presents the donated Anthropic stake as worth more than that amount. His supporting notes describe a ceiling, calculated using a reported ownership stake of less than 0.8 percent and Anthropic’s valuation. An upper bound has been promoted to a confident valuation. Even the attached graphic says “up to.” The footnote appears to have retained legal counsel after the headline left the room.
The location of the shares is similarly unsettled. Bass places them inside Good Ventures Foundation and makes their dominance of its portfolio central to his argument. His notes say the recipient of the donated shares is undisclosed. That missing identification matters enormously when the story depends on following money through a particular institution. You cannot establish financial dependence by supplying the missing account holder through narrative enthusiasm. Bass’s supporting notes
His subsequent research is more restrained still. It reports finding no payment or contract from Anthropic to METR in the filings examined, and no documented transaction into METR from Good Ventures Foundation, Coefficient Giving or Moskovitz personally. Those findings leave room for undisclosed arrangements and indirect relationships. They provide no basis for confidently declaring METR to be on Anthropic’s payroll. Bass’s follow-up investigation
An arrow on a network diagram can represent a grant, employment history, shared premises or a professional association. Each may warrant investigation. They have different implications. Financial control requires evidence about amounts, conditions, discretion and alternatives. Otherwise, the diagram risks becoming a family tree in which everyone is mysteriously paying everyone else’s mortgage.
METR’s own account also deserves careful reading. It says it has not accepted funding from frontier AI companies, while acknowledging substantial free tokens for evaluations, research and engineering. Access and computing resources can create dependencies even without cash payments. This is a concrete avenue for scrutiny: what happens if an evaluator publishes something a cooperating company dislikes? Can it preserve access, reproduce the work and continue operating? METR’s funding update
The comments I read demonstrate how quickly those distinctions disappear. There is applause, comparison with the rating agencies before the financial crisis, and confident repetition of the alleged funding loop. Kevin Simback offers a useful objection: investigating METR’s financial connections does not amount to auditing Anthropic’s finances. The opening claim borrows the authority of an exercise the post has not performed. Simback’s reply
The deeper analytical problem concerns incentives. A donor’s exposure to an AI company could encourage reassuring assessments that protect its value. It could also encourage dramatic assessments that strengthen demand for regulation. Bass emphasises the second possibility. Establishing which effect actually dominates requires evidence about decisions and outcomes. A theory capable of explaining both alarm and reassurance needs some observable condition under which it would be wrong.
Meanwhile, regulatory capture remains a legitimate concern. Compliance costs can favour large incumbents. Preferred evaluators can become institutional gatekeepers. Policymakers should examine who designs proposed standards, who can afford to satisfy them, and who gains authority from their adoption. Those questions become stronger when attached to specific rules and mechanisms.
None of this settles whether a particular AI risk is real. A compromised institution can identify a genuine danger. An impeccably independent institution can make a mistake. Funding analysis helps us decide where to demand stronger verification; it cannot substitute for examining the underlying result.
The useful response is to demand disclosed funding concentrations, enforceable publication rights, transparent methods and genuinely independent replication. Multiple evaluators should be able to disagree without losing their ability to work.
That standard should apply to safety advocates and their critics alike. Bass has assembled material worth examining. His rhetoric asks the reader to treat unresolved connections as a completed prosecution. When the claim is that everyone else is monetising fear, accuracy is an especially useful expense.
No comments yet