The AI doom debate and who funds the evaluators
WatchfulEye ·
1. The sequence
On 8 September 2026, a researcher announced on X that he had resigned from Anthropic. Within about a day his post had passed 100 million views. Two days after the resignation, a former president told Democratic donors that AI should be a central issue. Six days after it, the President of the United States phoned Nvidia's chief executive on stage at a technology conference and called the warnings "a hoax."
This piece sets out that sequence, the funding relationships around the AI evaluation proposal that came out of it, and what the public documents do and do not show.
The dates:
8 September. Jacob Coxon, who spent about three years on pretraining research at OpenAI and then Anthropic, posted his resignation on X at 8:04 p.m. ET: "Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives." He added: "The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt." He told Axios on 9 September that he had been at Anthropic for four months, two months short of the six-month point at which employee stock vests, and that he still holds OpenAI equity: "I left before any of my equity vested."
8-9 September. Evan Hubinger, an alignment science lead at Anthropic, replied on X: "Jacob is correct here—we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade. I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to." WIRED reported on 9 September that Coxon's post had more than 100 million views. CNBC's headline that day: "Researcher says AI has more than 10% chance of 'killing all humans.'"
9 September. Senator Bernie Sanders (I-Vt.), who had announced the Ban Artificial Superintelligence Act with Rep. Greg Casar (D-Texas) on 3 September, wrote on X: "Mr. Coxon is right. The very people building this technology admit that it could threaten the future of humanity" (The Associated Press). Illinois Governor JB Pritzker said in a statement from his office: "It's time to sound the alarm — louder — on reining in Artificial Intelligence."
10 September (reported 13 September). At a Democratic Congressional Campaign Committee fundraiser, in a discussion with House Minority Leader Hakeem Jeffries, Barack Obama said he was neither an "accelerationist" nor a "doomer." He said that if he were running for president, AI "would be one of my central agendas," that the technology "is moving very fast in private hands, and if we don't get on top of it, I think can be dangerous," and that once Democrats held the House they should "put together a framework for a very public conversation" (remarks provided by his office to NBC News and The New York Times). NBC News reported that a source familiar with his thinking said Obama saw Dario Amodei's essay as an "encouraging development." The New York Times reported, as relayed by TechCrunch, that Obama had offered himself as a "sounding board" to AI executives and had spoken to Amodei and OpenAI's Sam Altman.
12 September. Anthropic's chief executive, Dario Amodei, published the essay "We Must Pace the Frontier." Its first proposed step reads: "Each frontier AI company commits to giving ongoing, employee-like access to a team of embedded third-party evaluators (such as METR), whose role is to verify adherence to safety practices and commitments, report incidents, and help assess the alignment of not just completed AI models but training pipelines and processes." Amodei wrote that Anthropic "is unilaterally committing to this step now." Sam Altman wrote, "I agree with Dario that we need to pace the frontier," and Elon Musk posted, "Dario is right" (TechCrunch, 12 September). The post announcing the essay drew 67 million views, by Kevin Bass's count.
13 September. David Sacks, the former White House AI adviser, wrote: "Stop pretending METR is independent when it is intertwined with Anthropic's investors and staff."
14 September. Kevin Bass, an independent researcher, published an audit of the funding around METR and Anthropic and called for a congressional investigation. The same day, President Trump phoned Jensen Huang during the All-In Summit in Los Angeles and called the AI-risk warnings "a hoax."
The money-trail figures below are Bass's, credited to him, with the underlying documents named where they exist.
About the audit: who Kevin Bass is, and what was checked
The money trail in this piece comes from one researcher, so three things come first: who he is, what he did, and what WatchfulEye checked independently.
Who he is. Kevin Bass (@kevinnbass) is an independent researcher. His own bio says he has done "some of the most thorough financial reporting on Effective Altruism and AI Safety"; his background is public health, including bylines in Newsweek and the New York Post during the COVID era. He publishes free on a Substack and states his position on the subject: he has said publicly that he does not want AI slowed down, and that he is building a government-auditing platform as a hobby because AI "could be a weapon against AI abuse." He is not a neutral party. His numbers, and the checks described here, should be read with that in mind.
What he did. On 14 September 2026, Bass published the audit on X with a GitHub repository (github.com/kevinnbass/metr-money-figure) containing the research files, methodology, and a second audit pass applied to his own work. His method was public records: IRS 990 and 990-PF e-files, SEC filings, donor-advised-fund sponsor schedules, company announcements, and statements published by the people involved. Each number in his figures points to a row in his research files, and each row names its source document. Where a figure below rests on a filing, an announcement or a news report, that document is named in the text with its date.
What was checked independently. WatchfulEye checked the parts of the story that do not depend on his filings against the original documents: Coxon's post and the Axios interview; Hubinger's post as quoted by CNBC (9 September 2026); the WIRED view count (9 September 2026); Obama's remarks (NBC News and TechCrunch, 13 September 2026, the latter citing The New York Times); the Sanders-Casar announcement (Sanders's press release, 3 September 2026) and Sanders's post (The Associated Press, 9 September 2026); Governor Pritzker's statement (Office of the Governor, 9 September 2026); Amodei's essay text (darioamodei.com) and the Altman and Musk responses (TechCrunch, 12 September 2026); the All-In Summit transcript of 14 September 2026; METR's funding update of 14 August 2026 (about $71 million in commitments); METR's About page ("significant free tokens"); and METR's Frontier Risk Report of 19 May 2026, which records no published conflict-of-interest policy.
What was not re-derived. WatchfulEye did not re-download Bass's IRS e-file batches, rebuild his 2,911-row grant index, or re-reconcile every donor-advised-fund schedule. Those are his findings, documented in his repository, and are presented as his.
The open gap. An independent review of Bass's audit (Superpower Daily, 15 September 2026) confirms that his network map is real but concludes that the two central links are not established by public records: where Dustin Moskovitz's donated Anthropic stake sits, and whether money from it reaches METR. METR's funding update of 14 August 2026 thanks the Audacious Project, individuals from Jane Street, the Sijbrandij Foundation, the Pew Charitable Trusts, Schmidt Sciences, the Packard Foundation and several named individuals. It does not name Open Philanthropy, Coefficient Giving or Good Ventures. Bass found no direct Coefficient grant to METR in Coefficient's grants index, and the largest grant to METR he found in any filing, through Vanguard Charitable, has no named donor. The documents do not support the claim that Anthropic pays for its own evaluator. They do support a narrower one: the AI-safety funding network is small and interlinked, much of it is not itemized by donor, parts of it sit on the same side of the table as the labs it would evaluate, and no filing found by Bass, the review, or WatchfulEye documents the evaluator's financial independence.
2. What Jensen Huang and President Trump said
The main counterargument came from Nvidia's chief executive, Jensen Huang, at the All-In Summit in Los Angeles on Monday, 14 September, where President Trump called in during the interview. Huang told the President he was "sitting in front of a few thousand people" and put him on speaker. The quotes below are from the published episode transcript.
On the extinction estimates, Huang said: "we shouldn't because it's made up." On the record of earlier predictions: "there was a prediction that in 5 years' time, radiology will be completely taken over by artificial intelligence and there'll be no radiologists in the world. That has proven to be exactly the opposite. We need more radiologists than ever in the world." Of predictions about code generation and entry-level jobs, he said: "all of these predictions have been wrong." Earlier claims that GPT-2 and Llama 3 would be too unsafe to release were raised in the same exchange.
On regulation: "Regulation should solve actual problems." He said that "all of the actual problems so far have come from the labs," because "they have the most compute," and that after the recent incidents "the first thing that you have to do is just root cause the problem from an engineering perspective."
On recursive self-improvement (RSI), he called the underlying techniques "very sensible ideas," and said "this phrase is now being used to weaponize the technology in some way." Products, he said, still have to be evaluated and tested before release.
On China: "Nobody in China is saying that there's end of this and end of that." He said that in China "they're much more pragmatic about it" and "see AI as a technology that's going to advance their economy."
President Trump, on speaker: "The robots are not going to be taking over the world." "The whole thing is a hoax." "Whoever wins AI wins. That's how big it is. It's bigger than the internet." "We have $20 trillion of investment coming into the country," compared with "much less than $1 trillion" under President Biden. On data centers: "it's the oil of the next 20, 25 years," and "There are communities that were dying that have data centers right now, and now they're wealthy communities."
After the call, a host asked Huang why the President rejected the warnings when polling ran strongly against data centers and AI. Huang said that "the first thing that it was anchored on was national security," that "recently, that was all blown to bits," and that the story is "no longer anchored on national security. Now it's anchored on safety."
CNBC and NBC News reported the call on 14 September. NBC described the setting as an Nvidia company meeting; CNBC and the transcript place it on stage at the All-In Summit.
3. How a message spreads
An influence cycle does not require a coordinated plan. It needs a message, a distribution network, institutions with aligned incentives, and funding. The September sequence shows several of those parts.
A source inside a lab. The warning came from a researcher leaving Anthropic, and a senior Anthropic alignment researcher publicly agreed within hours and gave a number (more than 10% within a decade). The resignation post presented a view, not a document, and Hubinger presented his estimate as a personal view.
Prominent agreement. Within six days, Altman, Musk, Amodei, Pritzker and Obama had spoken, and Sanders had already announced a bill. Sacks questioned the evaluator's independence.
Speed. The posts moved from X into national news coverage within about a day (CNBC and WIRED, 9 September).
Funding on several sides. Bass's audit maps how money from one philanthropic network reaches the evaluator proposed for the labs, an AI-journalism fellowship, and other organizations in the field. Part 4 sets out what the filings show and what they do not.
The calendar. The sequence came seven weeks before the 3 November midterms, after a summer of reported AI incidents. Sanders and Casar announced their bill on 3 September, before the resignation. Obama urged Democrats to build a public agenda on AI. President Trump, a Republican, rejected the warnings outright. AI policy has become a campaign issue for both parties.
Unfalsifiable forecasts. A probability of human extinction within a decade is a subjective estimate. It cannot be confirmed or refuted on a timescale that would inform current policy. Huang's point was that comparable past predictions about AI capabilities and jobs did not come true. Those raising the alarm argue that recursive self-improvement is a new kind of risk that past predictions do not settle.
4. The money trail, by the numbers
Kevin Bass's audit, published 14 September, maps who funds METR (Model Evaluation and Threat Research, the nonprofit Dario Amodei proposed as an embedded "third-party evaluator"), who funds journalism that covers it, and where the money comes from.
The figures below are Bass's, from his thread "I have conducted an audit of Anthropic's finances" and the repository github.com/kevinnbass/metr-money-figure, which includes a second audit pass. They are reproduced with credit. Where a number comes from another document, that document is named with its date.
The donated stake. Forbes reported on 7 November 2025 that Dustin Moskovitz (Asana co-founder, early Facebook engineer) and Cari Tuna had moved their Anthropic stake, an estimated $500 million, into a nonprofit vehicle in early 2025, so they could put any "significant financial return" back into philanthropy and "dispel any perception of conflict of interest," as Tuna put it. On 20 April 2026 Forbes estimated the donated stake at less than 0.8% of Anthropic. On 28 May 2026 Anthropic announced a $65 billion Series H at a $965 billion post-money valuation. At that price, 0.8% is about $7.7 billion, Bass calculates: a ceiling, not a valuation. Moskovitz wrote on Bluesky on 30 March 2026, "Our Anthropic shares are entirely in our foundation - no personal benefit," and on 11 April 2026, "The foundation is invested in Anthropic as well." No public filing shows exactly where the stake sits, and Bass's full-text scan of the IRS e-files received in 2025 and 2026 found no return naming an Anthropic holding in a Moskovitz or Tuna vehicle. What the filings do show: Good Ventures Foundation's Form 990-PF for the year to 30 June 2025 lists its largest public holdings as TSMC, SK Hynix, Broadcom, Vistra, Micron, Vertiv and Constellation Energy. Its investment manager, Value Aligned Research Advisors, reported a public-equity book of $4.7 billion at 30 June 2025 and $40.1 billion at 30 June 2026 in its SEC Form 13F filings. Those filings cover all of the manager's clients, not the foundation alone.
The evaluator. On its About page, METR says it "has not accepted funding from AI companies, though we make use of significant free tokens," without naming the companies or stating the quantity. It also says it "cannot accept donations made by or at the direction of frontier AI company employees." The rule does not address lab investors.
The network, figure by figure:
(Figure 1: the money map as Bass traces it: Anthropic's donated stake, Good Ventures Foundation, Coefficient Giving, and the evaluator network. Source: Kevin Bass, metr-money-figure.)
Figure 1 (above): the full map, as Bass traces it. No filing places the stake, so the route through it is inference: from the Moskovitz vehicle to Good Ventures Foundation, which funds Coefficient Giving (formerly Open Philanthropy), which funds METR's former parent ARC, its partner RAND, its pooled-fund donor Longview, and other AI-safety organizations, including FAR AI, Redwood Research and Constellation. Two lines are documented outright. Jaan Tallinn, who led Anthropic's Series A (Anthropic, 28 May 2021) and is an Anthropic board observer (Postimees, 22 February 2026), funds METR through the Survival and Flourishing Fund. And the Audacious Project, housed at TED, catalyzed about $38 million for Canary, a joint RAND and METR project, about $17 million of it for METR (METR, 9 October 2024). On Bass's map, the evaluator's funding sits downstream of some of the same people who invested in the lab.
(Figure 2: METR donors and donor channels checked against Anthropic's funding announcements. Source: Kevin Bass.)
Figure 2 (above): donors on both sides. Bass checked 17 METR donors and donor channels against Anthropic's funding announcements. Two METR funders are named in Anthropic's Series A announcement of 28 May 2021: Jaan Tallinn, who led the round, and Eric Schmidt, whose Schmidt Sciences METR lists as a funder. Dustin Moskovitz and James McClave are named in the same announcement, but Bass found no direct grant from either to METR. METR names "individuals from Jane Street" among its supporters (14 August 2026). The firm bought $100 million of Anthropic shares from the FTX estate in March 2024, according to the bankruptcy court filing, and Anthropic's announcements list it among investors in Series E, F, G and H. METR's independence rule excludes donations from lab employees. It does not exclude lab investors.
(Figure 3: which of 22 published proposals name METR as an evaluator. Source: Kevin Bass.)
Figure 3 (above): who proposed METR. Bass read all 22 proposals for a "FINRA for AI" published between April and September 2026. Four name METR: podcaster Dwarkesh Patel (9 September), Dario Amodei (12 September), David Sacks, to reject it (13 September), and a Washington Examiner piece relaying Sacks (13 September). No institutional proposal, bill, or framework names any examiner. On Bass's count, two of the 22 proposals put METR forward, and one of them came from the chief executive of a lab it would evaluate.
(Figure 4: METR's fundraising set against the incident timeline. Source: Kevin Bass.)
Figure 4 (above): the timing of the money. METR said on 14 August 2026 that it had raised commitments of around $71 million in the previous six months, roughly seven times the $10 million budget its 2024 annual report gives for that year. In the same window, the evaluator role was drafted into the Frontier Risk Report, both labs' governance frameworks, Demis Hassabis's proposed FINRA-style body, the White House review and the FRONTIER Act. The incident investigations came late in that window or after it: METR's team was first on site at OpenAI on 30 July (METR report, 26 August 2026), and Anthropic announced its agreement with METR on 9 September 2026. Routine risk reviews were already under way in March. Bass reads the sequence as the money anticipating the role rather than following the incidents, and says that reading is inference. The dates are documented.
5. Funding and AI journalism
One of Bass's figures concerns journalism about METR.
(Figure 5: the Tarbell Center fellowship and TIME100 AI profiles. Source: Kevin Bass.)
Figure 5 (above): the Tarbell Center. Coefficient's grants index lists three awards to the Tarbell Center for AI Journalism: $816,000 (November 2024), $2.888 million (March 2025) and $1.588 million (July 2025), about $5.3 million in all, Bass found. The Survival and Flourishing Fund recommended another $1.3 million in 2024 and 2025. Tarbell fellows have bylines at outlets including TIME, The Verge, MIT Technology Review and The Guardian, by Bass's count. Between September 2024 and September 2026, Bass found nine articles about METR or its chief executive, Beth Barnes, at TIME, The Verge and Platformer; six were written by Tarbell fellows; none disclosed the fellowship. Both TIME100 AI profiles of Barnes (5 September 2024 and 27 August 2026) were written by the same Tarbell fellow, Harry Booth. Neither disclosed the fellowship's funders.
Tarbell says its donors have no editorial control, and none of this shows that any article was inaccurate. It is a disclosure question. The head of a nonprofit that shapes how AI risk is discussed was profiled for TIME100 AI by a journalist in a fellowship that the Survival and Flourishing Fund supports. The same fund recommends grants to METR.
(Figure 6: the evaluator market. Nine of 20 candidate organizations hold Coefficient awards. Source: Kevin Bass.)
Figure 6 (above): the evaluator market. Bass tabulated 20 plausible candidates for the evaluator role, from RAND and FAR AI to Stanford NLP, the EU AI Office lots, and Hugging Face. Nine of the 20 hold Coefficient awards. One nonprofit in the set refuses lab money outright: METR (US CAISI and the EU AI Office also refuse, but they are governments). FAR AI caps lab money; Transluce discloses it; Epoch and others accept it. The three candidates put forward in mid-September, METR, Stanford NLP and Hugging Face, were named by Anthropic's chief executive, a Stanford professor, and Hugging Face's chief executive. On Bass's map, the largest funder in this market traces back to one philanthropic network.
(Figure 7: money flows to METR and Redwood Research from Coefficient and Good Ventures Foundation. Source: Kevin Bass.)
Figure 7 (above): the investigator's subcontractor. On 12 September 2026, Redwood Research posted on X that "several staff from Redwood have been subcontracted by METR" to work on the Anthropic investigation; the terms were not disclosed. Coefficient's grants index lists five awards to Redwood totaling $63.1 million, the largest $36.6 million on 18 November 2025. On 9 September 2026 Coefficient wrote that its grantmakers had "recently recommended more than $70 million over the next two years" for Redwood, in a passage citing Redwood's work on the OpenAI investigation. Redwood's Form 990 for 2024 reports revenue of about $22,000 and six employees, after it spun off its Constellation program in 2023. Redwood's Forms 990 list Holden Karnofsky as a director through November 2023, Paul Christiano through June 2023 and Ajeya Cotra through August 2024. Karnofsky joined Anthropic in January 2025 as a member of technical staff working on its Responsible Scaling Policy (Fortune, 13 February 2025, confirmed by an Anthropic spokesperson); Cotra is now on METR's technical staff (METR team page). On 13 September 2026, according to Bass, METR edited its 26 August report on the OpenAI incident to disclose that Ryan Greenblatt, the Redwood staffer contracted onto that investigation, "is the domestic partner of Beth Barnes, METR's CEO," adding that Barnes "was not involved in the decision to engage" him. The report does not say who made the decision.
In one line: Coefficient funds Redwood, METR subcontracts Redwood staff, and METR investigates Anthropic, whose early investor Dustin Moskovitz funds Coefficient. The filings raise a structural conflict-of-interest question. They do not show that any party acted improperly.
6. METR's self-assessment and its conflict-of-interest policy
(Figure 8: METR's independence self-assessment, showing no published conflict-of-interest policy. Source: METR, Frontier Risk Report, via Kevin Bass.)
Figure 8 (above): METR's own self-assessment. Four months before Amodei proposed METR as the embedded evaluator, METR assessed its own independence against AEF-1, the voluntary standard of the evaluator industry body it belongs to. The results, from its Frontier Risk Report of May 19, 2026: it answered Yes on 21 of 26 items. It answered No on requirement 2.3, a published conflict-of-interest policy. It answered No on requirement 5.4, a responsible disclosure policy. On 4.6, it answered "See notes": participants in its pilot could exit without notice, and those who stayed could redact certain findings. On whether it had disclosed all conflicts relevant to the evaluation, METR wrote: "Several of the staff and collaborators directly involved in this pilot (at least 6) have close personal relationships with AI company staff." On whether a meaningful fraction of its funding comes from the system provider, METR wrote: "Note that METR's work with nonpublic models and use of free tokens incentivize a cordial relationship with AI companies."
In plain terms: the evaluator proposed by Anthropic's chief executive reported that it had no published conflict-of-interest policy and that at least six of the people involved in its pilot have close personal relationships with AI company staff. Those statements are METR's own, made four months before the role was proposed for it.
The relationship predates 2026. Anthropic's original Responsible Scaling Policy, published September 19, 2023, credits ARC Evals, the organization that became METR, with "key insights and expertise supporting the development of our RSP commitments," and recognizes ARC's responsible-scaling framework as the one that "inspired our approach." METR now has three roles: researcher of dangerous autonomous capability, contributor to the industry standard (AEF-1), and a participant in the market its report says should exist; METR wrote that "periodic third-party assessment of risks from developers' internal use of AI should be adopted throughout the industry." Standards bodies routinely include practitioners, and none of this is improper on its face. Audit firms that help develop a standard and then audit against it are usually held to stronger independence rules.
(Figure 9: Vanguard Charitable, METR's largest identified grant channel, and the growth in donor-advised grants. Source: Kevin Bass.)
Figure 9 (above): donor-advised money. The largest single grant to METR that Bass found in any Form 990, $4.0 million, came through Vanguard Charitable in its fiscal year to June 2025 (Vanguard Charitable Form 990, Schedule I). Donor-advised funds do not name donors, so no filing identifies its source. Bass's chart shows the wider pattern: Vanguard Charitable's grants to the AI-safety and EA-infrastructure cluster went from $4.3 million (FY2023) to $8.2 million (FY2024) to $65.6 million (FY2025), an eightfold jump in a year, by Bass's tally of the fund's Schedule I grant lists. The FY2025 grantee list is, in Bass's words, "the SFF and Coefficient recipient set." No filing names the donors behind that money.
(Figure 10: the disputed direct Coefficient grant. Filings do not confirm it; the ledger shows what is documented. Source: Kevin Bass.)
Figure 10 (above): both sides' claims, scored. When the debate broke open, METR's defenders said "METR has never accepted any funding from them" (meaning Coefficient/Open Philanthropy), and critics said "Coefficient pays their bills." Bass scored both claims against his research ledger. He found no direct Coefficient grant to METR in the 2,911-row awards index or the filings. No filing can prove "never" as a lifetime claim. The ledger does show a $4.55 million ARC program transfer to METR at spin-out (ARC's Form 990 for 2024), a $10 million Coefficient award to METR's RAND partner (September 2025), unbooked lab tokens, and donor lines with direct lab ties. Neither side's exact claim holds; the narrower facts do.
7. The public debate, by Bass's count
Bass's research ledger also records how the debate played out online and in the press.
(Figure 11: the debate by the numbers. 67M views vs 7.1M, and no replies located from METR, Redwood, or Coefficient. Source: Kevin Bass.)
Figure 11 (above): the debate, by the numbers. Bass collected the significant posts in the 11-14 September argument. Amodei's essay announcement, naming METR as the embedded evaluator, drew 67 million views. Sacks's reply drew 7.1 million. Bass searched metr.org, METR's Substack, and the accounts of Beth Barnes and board member Chris Painter from 12 to 14 September and located no response to Sacks or Amodei. On Sacks's thread, 86 quotes and replies passed 1,000 views; none came from METR, Redwood or Coefficient accounts. The two accounts of SFF-funded organizations that did reply agreed with parts of Sacks's point.
Bass also logged 31 press records on 12-13 September about the controversy. Ten named METR. One discussed METR's funding, and it was a blog fact-check; none of the straight-news reports in his sweep did.
8. Two readings, and a second money trail
Bass and Huang reach a similar conclusion from different directions: the warnings have real consequences whether or not the forecasts are right. In six days they drew responses from the President, a former president, a governor and a senator, and they bear on capital, regulation and votes. The question the documents can address is who funds the institutions that would interpret the risk.
Bass's closing claim, in his words: "Anthropic's business model models itself after the very thing it claims to fear. Except Anthropic's ideology infects humans, not computers." That is his argument, not a finding of this piece.
On China, Bass wrote: "China is keeping messaging tight. That is why optimism for AI is so high in China." Huang made a similar point at the summit. Neither statement measures Chinese public opinion. Both contrast how AI risk is discussed in the two countries.
A second money trail. A separate funding question comes from the other side of the debate. In May 2026, Kevin O'Leary, an investor in a planned data center campus in Box Elder County, Utah, said he wanted "a forensic auditor" to find out who was funding the groups opposing it: "Which adversary would want that? There's only one. It's China" (Fox 13 Salt Lake City, 12 May 2026). In a social media post reported by Fox 13 on 25 June 2026, he wrote that he had "no evidence" that the Utah groups and people he had named were funded by China or the Chinese Communist Party (Fox 13). The Bitcoin Policy Institute's report of 18 May 2026, *Foreign Influence in the Campaign against American AI* (Sam Lyman, BPI's head of research), cites an American Energy Institute report of April 2026 that found twelve US organizations opposing data center development had collectively received more than $39 million from foreign donors. BPI also reports that on 29 April 2026 Senator Bernie Sanders convened a Capitol panel on "the existential threat of AI" where two of the four panelists were Chinese government affiliates, Zeng Yi of the Beijing Institute of AI Safety and Governance and Xue Lan, chair of China's national AI governance expert committee; that Xue called the US-China AI race "an inaccurate narrative" and argued for "safe zones" of cooperation on AI safety; and that the panel came a month after Sanders introduced legislation for a moratorium on new AI data centers. The limits: BPI is a pro-bitcoin think tank with its own interests. The AEI total counts documented foreign funding to those organizations; it does not show that the money paid for their data center campaigns. Graham Webster, a research scholar at Stanford University's Center for International Security and Cooperation, told WIRED that the BPI report's examples do not match documented Chinese influence campaigns: "It's totally normal for the English-language Chinese media to pick up storylines that are in the US media. It's just how wire services work" (WIRED, 12 June 2026). NPR reported on 10 June 2026 that independent researchers had so far found little evidence of a coordinated Chinese effort against U.S. data centers. "We haven't found much," said Darren Linvill, co-lead of the Media Forensics Hub at Clemson University, who added that he could not rule out covert paid influence but saw no need for it (NPR, 10 June 2026). O'Leary is a data center investor who withdrew his China charge. On both trails, the people raising the alarm are not the only ones with money riding on the outcome.
9. Coordination is not required
The documents are also consistent with a simpler explanation: the people involved may believe what they are saying.
Anthropic has set out this view since at least 2023, when its first Responsible Scaling Policy contemplated pausing scaling if safety fell behind capability. Amodei's September essay revises a three-year-old institutional position. Hubinger's 10% is a subjective probability, not a measurement, and subjective probabilities can be honestly held. Coxon may have quit out of genuine concern. Obama may have repeated a concern he holds. Sacks's objection, Bass's audit, and this article come from the same environment: people acting on their reading of the incentives in front of them.
None of those facts proves coordination, and this article does not claim it. Four hypotheses about the political adoption of AI-risk warnings are worth separating. Policy convergence, that Democratic politicians independently find AI regulation attractive, is supported by observable behavior. Narrative adoption, that politicians increasingly repeat catastrophic-risk framing that originated inside the AI-safety field, has evidence of overlap but no proven causal chain. Coordinated strategy, that Anthropic or its funders deliberately work with Democrats to make AI risk an electoral issue, has no evidence in this review. Electoral opportunism, that politicians privately disbelieve the risk and use it anyway, is likewise unestablished. The documents support the first two, and they show an idea becoming an institution, with institutional interests that do not require coordination.
The structural facts are these. Anthropic spent more than $3.5 million on federal lobbying in the first half of 2026, more than the $3.1 million it spent in all of 2025, Axios reported on 21 July 2026, while the evaluation framework it helped design moves toward law. METR raised commitments of around $71 million in six months (METR, 14 August 2026), before the regulatory framework settled, while recommending that its own practice become an industry standard. Compliance regimes of this kind carry fixed costs that a frontier lab can absorb and a smaller competitor, an academic group, or an open-source project may not. A lab can sincerely believe the risk is real and still benefit from rules that weigh more heavily on smaller rivals. If the largest developers converge on the position that only sophisticated organizations can build AI safely, under continuous inspection by credentialed outside evaluators, and government codifies it, market concentration can follow without any collusion.
10. Open questions
The documents leave three questions open.
Disclosure. METR's own compliance table records no published conflict-of-interest policy, close personal ties between at least six people involved in its pilot and AI company staff, and reliance on free tokens from AI companies. Whether any "independent evaluator" proposal will require written answers to those points is open.
The money route. Bass published his research files, with a source for every row, plus a second audit pass and a tracker of what could not be verified. The independent review of his work found that the two central links, where the donated stake sits and whether it funds METR, are not established, and Bass says so himself. The review concluded that the narrower governance problems exist regardless of the money route. Bass's research notes of 14 September say Coefficient Giving's 2025 Forms 990, due by November 2026, could show whether any of its entities received the stake.
How to test future warnings. Four questions apply to any claim about AI extinction risk, and to claims that the risk is a hoax: who made the claim, who confirmed it, who funds the people making it, and what policy or commercial interest it supports.
*This piece was written by the WatchfulEye research team. It reproduces the audit figures of Kevin Bass (@kevinnbass) from his September 14, 2026, thread, "I have conducted an audit of Anthropic's finances," and his repository github.com/kevinnbass/metr-money-figure, with credit, and names each underlying document in the text with its date. Quotes from Jensen Huang and President Trump are from the All-In Summit episode transcript, Los Angeles, September 14, 2026; CNBC and NBC News reported the call the same day. Coxon's quotes are from his post on X of September 8, 2026, and his Axios interview of September 9, 2026; Hubinger's quote is from his post on X as reported by CNBC on September 9, 2026; the view count is from WIRED, September 9, 2026. Obama's remarks were reported by NBC News and by TechCrunch, citing The New York Times (September 13, 2026). The Sanders-Casar bill announcement is from Senator Sanders's press release of September 3, 2026; Sanders's post on X is as reported by The Associated Press, September 9, 2026; Governor Pritzker's statement is from his office's release of September 9, 2026; Amodei's essay is "We Must Pace the Frontier" (darioamodei.com/post/we-must-pace-the-frontier), dated September 2026, and its publication date and the Altman and Musk responses are from TechCrunch, September 12, 2026. The ARC Evals attribution quotes are from Anthropic's Responsible Scaling Policy of September 19, 2023 (anthropic.com/news/anthropics-responsible-scaling-policy); METR's "adopted throughout the industry" line and its self-assessment are from its Frontier Risk Report of May 19, 2026 (metr.org/blog/2026-05-19-frontier-risk-report); the lobbying comparison is from Axios, July 21, 2026 (axios.com/2026/07/21/anthropic-ramps-up-lobbying-spending-ai-policy-fights); Holden Karnofsky's move to Anthropic is from Fortune, February 13, 2025. The scope and limits of WatchfulEye's checks are set out in the "About the audit" section above.*
Corrections
If a source, a frame, or an inference on this page is wrong, we correct it here. contact@watchfuleye.us.
Correction (3 October 2026): An earlier version of this article attributed to Jensen Huang quotes that are not in the All-In Summit episode of 14 September 2026, including "We should not regulate because of hypotheticals"; they are replaced with his words from the episode. It said David Sacks asked President Trump a question that Huang answered; in the episode, a host asked Huang. It quoted Sacks as saying "show us the data, show us the report, show us the leaked information"; no source for those words was found, and the quote was removed. It quoted Illinois Governor JB Pritzker as saying "It's time to sound the alarm louder on reining in AI"; his office's release of 9 September 2026 reads "It's time to sound the alarm — louder — on reining in Artificial Intelligence." It credited NBC News with reporting that Barack Obama had been a "sounding board" for AI executives; The New York Times reported that, as relayed by TechCrunch. It called Jacob Coxon's resignation post anonymous; it was posted under his name. It said Senator Bernie Sanders already had the bill; Sanders and Rep. Greg Casar announced the Ban Artificial Superintelligence Act on 3 September 2026 and introduced it on 23 September 2026, according to Sanders's office.