A thread alleging that Anthropic sits at the center of a self-funding "AI Doom Machine" hit 1.2 million views within hours on September 15, 2026, boosted by a repost from Meta's former chief AI scientist Yann LeCun. Its author, independent researcher Kevin Bass (@kevinnbass), calls it "the Anthropic Network" and says it amounts to a regulatory capture machine that cannot turn itself off. He's calling for a Congressional investigation. The thread names real organizations, cites real filings, and includes a genuinely relevant rebuttal from one of the people it implicates — but as of publication, the core financial claim linking METR's evaluations to Anthropic's stock price has not been backed by the evidence Bass says is still coming.
This is worth covering carefully, not because the conspiracy framing is necessarily right, but because the structural question underneath it — how do you fund "independent" AI safety evaluation without creating exactly this kind of entanglement — is one explainx.ai has covered from the other direction, and it deserves the same scrutiny in both directions.
TL;DR: claim vs. what's actually confirmed
| Bass's claim | Status |
|---|---|
| METR is "on Anthropic's payroll" | Not literally true — METR is a nonprofit, not an Anthropic vendor. Bass's actual claim is indirect: shared donor funding, not a direct contract. |
| Good Ventures Foundation is "the overwhelming funder" of the AI-safety ecosystem funding METR | Partially plausible but unverified at the scale claimed; Bass promised supporting evidence not yet published |
| Dustin Moskovitz's Anthropic stake grew from ~$500M to ~$7.7B in ~16 months | Unverified specific figures; Moskovitz's early Anthropic investment is real and publicly known |
| Open Philanthropy (Coefficient Giving) invested in Anthropic | Directly disputed by Coefficient Giving's president, Alexander Berger, in a screenshot Bass himself included |
| Tarbell Center places "AI Doom" stories in major outlets, funded by the same network | Confirmed as funded ($5.3M from Coefficient Giving, per Bass's diagram); its fellows' bylines at TIME, The Verge, MIT Tech Review, The Guardian, and the LA Times are separately verifiable |
| METR receives money directly from Anthropic-linked funders | Bass's own diagram says "none found" for a direct grant in Coefficient Giving's index or Good Ventures' tax filings — the claim is about indirect, multi-hop overlap through intermediaries, not a direct payment |
| "Congress must investigate" | Bass's call to action, not something that has happened |
The argument, as Bass makes it
Bass's thread starts from a real proposal: Anthropic CEO Dario Amodei's "We Must Pace the Frontier" essay, which calls for third-party evaluators to independently assess frontier model risk, and specifically names METR as an example of that kind of evaluator.
Bass's objection is that METR isn't structurally independent of Anthropic's success. His chain of reasoning: Good Ventures Foundation, the philanthropic vehicle funded substantially by Dustin Moskovitz (an early Anthropic investor and Meta co-founder), holds a large and rapidly appreciating Anthropic equity stake. Good Ventures, in turn, is a major funder of the broader AI-safety nonprofit ecosystem — an ecosystem that Bass says includes METR and, further out, media and advocacy efforts like the Tarbell Center for AI Journalism, which places AI-risk reporting in outlets including The Verge, Science, the LA Times, and TIME.
Bass's conclusion is that this creates a closed loop: if Anthropic's valuation keeps climbing, the network funding both AI-risk evaluation and AI-risk journalism gets richer and louder; if Anthropic falters, that same network loses its funding base. Either way, he argues, none of the organizations involved can afford to produce findings that meaningfully threaten Anthropic — which would make "third-party" evaluation a misnomer.
He extends the argument politically, framing Anthropic's safety messaging as a form of "anti-AI propaganda" that risks slowing US AI development relative to China, and name-drops Jacob Coxon — the Anthropic pretraining researcher who publicly resigned in September 2026 citing AI-doom-style concerns — as, in Bass's framing, a product of this same ecosystem rather than an independent voice.
What's actually verifiable right now
Some pieces of this are on solid ground. Dustin Moskovitz's early personal investment in Anthropic is well documented, and it's also true that Good Ventures Foundation and its associated grantmaking arm — renamed Coefficient Giving from Open Philanthropy — fund a meaningful share of the AI-safety nonprofit sector, including organizations doing model evaluation, policy research, and AI journalism. That concentration is a matter of public record and isn't itself in dispute. explainx.ai's own coverage of Coefficient Giving's $200 million safety-org grant program noted the same underlying structural issue from the opposite angle: most AI safety research capacity currently sits either inside the labs building the systems, or funded by a small number of donors connected to those same labs. Bass's thread is, in effect, arguing that Coefficient Giving's attempt to fund evaluation outside the labs doesn't fully solve the independence problem — it just relocates it.
Where the thread gets shakier is in its specific numbers and its causal leap from "shared funding exists" to "therefore findings are compromised." Bass's follow-up posts cite a National Philanthropic Trust Form 990 Schedule I filing listing FY2025 grants to organizations he links to "METR's network" — Founders Pledge ($78.8M), the RAND Corporation ($61.6M), FAR AI ($14.4M), Effective Ventures USA ($11.6M), and Epoch AI. Those figures, if accurate, describe a real philanthropic funding cluster. They do not, by themselves, establish that METR's technical evaluations of Anthropic's models have been shaped by that funding, which is the actual claim doing the work in Bass's argument. He says fuller evidence and a GitHub repository are coming in a separate post; as of this writing, that evidence hadn't been published, so the strongest version of the claim remains unverified.
The follow-up diagram: more nuanced than the original thread
Bass published the promised follow-up: a dense funding-flow diagram, sourced to a research repository of CSV files, tracing "funders that reach METR through ARC, RAND, Longview and pooled funds." It's more careful than the original thread in one important respect — it explicitly labels what it could and couldn't confirm. The diagram states plainly that no direct METR grant appears in Coefficient Giving's 2,911-row award index or in Good Ventures Foundation's tax filings through 2025, and that no filing confirms where Moskovitz's donated Anthropic stake — publicly estimated at $500 million in November 2025, with a ceiling as high as $7.7 billion at Anthropic's reported September 2026 Series H valuation — actually sits today.
What the diagram does show, sourced to specific filings, is a web of indirect funding: Coefficient Giving and Good Ventures money flowing to intermediary organizations — the Alignment Research Center (METR's parent until its December 2021 spin-off), RAND Corporation, Longview Philanthropy, FAR AI, Redwood Research, and The Audacious Project's pooled donor fund — several of which have separately funded or worked with METR. Two of Anthropic's Series A investors, Moskovitz and Jaan Tallinn, are also identified in the diagram as Anthropic board observers by their own public account. The diagram also documents $5.3 million in Coefficient Giving awards to the Tarbell Center for AI Journalism, whose fellows have bylines at TIME, The Verge, MIT Technology Review, The Guardian, and the LA Times.
This is a meaningfully different claim than "METR is on Anthropic's payroll." It's closer to: a small number of Anthropic-linked donors fund a cluster of organizations, several of which separately fund or collaborate with METR, and the same donor cluster also funds AI-risk journalism. That's a real, checkable pattern of overlapping interests — and also a common structure in a small philanthropic field, not necessarily evidence of coordinated intent. The diagram includes direct quotes from Moskovitz addressing the conflict-of-interest question himself: he has said his donated Anthropic holdings sit "entirely in our foundation, dedicated to charity," and that he considers himself "naive in some ways" about how the arrangement looks from outside, while maintaining he thinks that perception isn't warranted.
The rebuttal Bass included himself
Notably, Bass's own thread contains a screenshot of Coefficient Giving president Alexander Berger directly disputing a closely related claim — though it's worth being precise about the timing: Berger's quoted reply is dated December 18, 2025, responding to an earlier, similar allegation, not a same-day response to this specific September 15 thread. Berger's stated corrections are: Open Philanthropy never invested in Anthropic; Dustin Moskovitz invested personally, early on, and has since donated his stake — not to Open Philanthropy or its successor; and the organization now goes by Coefficient Giving, not Open Philanthropy. If accurate, that rebuttal undercuts a specific factual claim in Bass's chain (that the grantmaking organization itself holds or invested in Anthropic equity), even though it doesn't address the broader argument about overlapping donor networks and shared personnel across AI-safety organizations.
Most establishment voices in the AI safety field have not endorsed METR's role in the way Amodei's proposal envisions, according to Bass — he specifically names Dwarkesh Patel and Dario Amodei as the notable exceptions who have. That's a claim about the state of expert opinion, not a financial fact, and it's worth reading as Bass's characterization rather than a tallied count.
Why this is worth taking seriously anyway
Set aside whether "regulatory capture machine" is the right frame. The underlying governance question predates this thread and will outlast it: when an entire field's evaluation and advocacy infrastructure is financed by a small number of donors whose wealth is itself tied to the success of the companies being evaluated, how much independence is actually possible? This isn't unique to AI — similar concentration concerns have come up in pharmaceutical safety research funded by industry-adjacent foundations, and in financial-sector research funded by bank-linked endowments. The pattern is old; AI safety is just the newest arena.
What makes this particular moment notable is who's amplifying it. Yann LeCun has been a public, sustained skeptic of existential AI-risk framing, including specifically Anthropic's, and his repost signals that Bass's thread is being read — accurately or not — as validating ammunition in the broader "pace the frontier" fight that's been running since Amodei's essay. That fight already spans investors calling the pacing push "self-serving hype," politicians on both sides of the US political spectrum, and now a funding-transparency argument aimed squarely at the evaluators Amodei's own proposal depends on. Jack Dorsey's separate "open the frontier" essay, published the same week, made an adjacent but distinct point: that publication restrictions and evaluator access shouldn't be controlled by the same incumbent labs whose commercial interests they might protect. Bass's thread pushes that same independence question one level further, toward the funders behind the evaluators themselves.
What to watch next
Bass has now done more of the legwork than most viral funding-conspiracy threads bother with — a sourced diagram citing specific filings is a higher bar than the original thread alone. But two things still separate "documented overlapping funding network" from "proof METR's evaluations are compromised": independent verification of the CSV-sourced claims by someone outside Bass's own project, and any actual finding that METR's technical conclusions about Anthropic's models changed because of donor pressure, rather than simply noting that funders overlap. Neither has happened yet. Also worth watching: whether METR or Coefficient Giving issue a direct, dated response to this specific thread and diagram, rather than commentators pointing back to Berger's eight-month-old, differently-scoped reply; and whether any sitting member of Congress picks up Bass's call for an investigation, which as of publication has not happened. Until then, this is a documented pattern of overlapping donor relationships, correctly labeled by Bass's own diagram as indirect and partly unconfirmed — worth tracking, and worth taking as a real structural concern, without accepting the strongest "regulatory capture machine" framing as established fact.
Related reading
- Update — September 15, 2026: A current OpenAI researcher's statement raises a deeper, related independence problem — even a well-funded, unconflicted evaluator may not be able to trust its own test results once models become situationally aware enough to recognize they're being evaluated. A current OpenAI researcher says models are now too situationally aware to evaluate →
- Coefficient Giving's $200 million AI safety grant program
- Dario Amodei's "Pace the Frontier" proposal, explained
- Pace the frontier goes cross-partisan: Baker, Burry, Trump, and Harris react
- Jack Dorsey's "open the frontier" reaction
- Anthropic researcher Jacob Coxon resigns over AI safety fears
- What is an embedded evaluator in AI safety?
- Is "pace the frontier" safety or a plateau narrative?
- Source: Kevin Bass's original thread on X (view the thread directly for his full argument, cited filings, and any updates)
This post reports on an active, disputed public allegation as of September 15, 2026. Financial figures and diagram claims attributed to Kevin Bass are his own sourced research, not independently verified by explainx.ai. This post will be updated if METR, Coefficient Giving, or other named parties issue direct, dated responses to this specific thread and diagram.