Research: The EA to AI-Safety Funding Pipeline

Companion pages: The AI-Worldview Poles (the camps), AI Safety Theater (safety as the capture surface), The Funding Ratchet (the donor-advised-fund method, applied to political advocacy), and the Revolving Door map.

The question. How did a philosophical position about effective charity become the mechanism that decided which AI concerns received institutional attention? The unit of analysis here is the money and the plumbing it runs through: grants, grantees, board seats, fellowships, and the disclosure record. Where a person appears, they appear as a node with a documented role. Every dollar figure is the funder’s or the recipient’s own number, an IRS Form 990, or a named outlet’s report. Philanthropic grantmaking, equity investment, congressional fellowships under the Intergovernmental Personnel Act, and the political spending in section 7 are all lawful. The subject is the structure, its concentration, and how much of it can still be counted from outside.

1. The engine: the argument that makes the money rational

The pipeline is not a conspiracy, and treating it as one misses the mechanism. It is a decision procedure applied consistently, and the arithmetic is published.

Nick Bostrom’s 2013 Global Policy paper is the load-bearing document. Because the number of possible future people is astronomically large, any nonzero reduction in extinction risk dominates any definite present good:

“Even if we give this allegedly lower bound on the cumulative output potential of a technologically mature civilisation a mere 1 per cent chance of being correct, we find that the expected value of reducing existential risk by a mere one billionth of one billionth of one percentage point is worth a hundred billion times as much as a billion human lives.”

“One might consequently argue that even the tiniest reduction of existential risk has an expected value greater than that of the definite provision of any ‘ordinary’ good, such as the direct benefit of saving 1 billion lives.”

(Nick Bostrom, “Existential Risk Prevention as Global Priority,” Global Policy 4(1), 2013)

The argument does not need extinction to be likely. It says the expected value of a speculative future outweighs the measured value of a documented present by a margin so large that no present harm can compete inside the same spreadsheet. Everything downstream (which organizations exist, which questions get staffed, which Senate offices have a fellow) is that inequality executed by people who believe it.

The counterweight the grantmaker built into itself. Open Philanthropy never claimed the argument settles everything. It publishes a doctrine it calls worldview diversification: “dividing resources across several different worldviews,” because it finds several plausible and none “clearly all-things-considered much better than the others” (Coefficient Giving, “Worldview Diversification”). The buckets are held apart so that the Bostrom arithmetic cannot eat the malaria budget outright. Section 4 shows what happened to that firewall when money got tight, and what happened when it got loose again.

2. The grantmaker: scale, concentration, and the name change

Renaming the bucket. The portfolio now called Global Catastrophic Risks used to be called “Longtermism.” The funder gave its reason: “The new name better reflects our view that AI risk and biorisk aren’t only ’long-term’ issues” (Our Progress in 2023, footnote 2). The philosophical brand took reputational damage in 2022-23; the operational brand is easier to defend in front of a senator. The money did not move. The label did.

3. The grants: who got paid, how much, and when

3.1 MIRI: the oldest node, and the dependency ratio

Open Philanthropy awarded the Machine Intelligence Research Institute $7,703,750 over two years, announced April 27, 2020, MIRI’s largest grant to that date. With the second half of a 2019 grant, MIRI put the support at $4.38 million a year, “or roughly 60% of our predicted 2020-2021 budgets.” The same post lists $300,000 from the Berkeley Existential Risk Initiative and $100,000 from the EA Funds Long-Term Future Fund, and records that “Open Philanthropy decided not to move forward with its partnership with Ben Delo,” the BitMEX co-founder first credited with part of the grant; “other funders provided the funds” (MIRI, “MIRI’s largest grant to date,” Apr 27 2020).

MIRI’s IRS Form 990 figures (ProPublica Nonprofit Explorer, EIN 58-2565917):

The 2021 spike was crypto, not Open Philanthropy. An anonymous investor gave 3,001 MKR (MakerDAO governance tokens), which MIRI “liquidated … for $15,592,829,” released at $2.5 million a year for 2021-2024 with the rest in 2025; Vitalik Buterin gave 1,050 ETH “worth $4,378,159,” which MIRI ranked third among its contributions “after Open Philanthropy’s ~$7.7M grant in 2020 and the anonymous donation above” (MIRI, May 13 2021).

Read together, the shape is visible without assertion: an organization running at $1.6 million a year in 2015 was running at $7.8 million by 2020, about 60 percent of it from one grantmaker; it banked $25.6 million in 2021, about $20 million of that from two crypto gifts; and in 2022 revenue fell to $1.9 million while expenses stayed at $5.3 million. That is a field’s oldest institution living off a small number of very large decisions. It is what turns a funding pipeline into an apparatus: not that the money is large, but that a few donors’ choices decide whether the institution makes payroll.

MIRI ran its “first fundraiser in six years” in December 2025, targeting $6 million with the first $1.6 million matched “via an SFF grant” (the Survival and Flourishing Fund, Jaan Tallinn’s grantmaker). It raised $1,802,437 in donations plus the $1,600,000 match. MIRI says it aims to hold “at least two years’ worth of reserves,” and credits “generous support we received in 2020 and 2021” for the gap (MIRI’s 2025 Fundraiser).

3.2 Center for AI Safety: the extinction statement’s funder

Open Philanthropy gave the Center for AI Safety $5,160,000 in general support, awarded November 2022, for “technical research and field-building aimed at reducing catastrophic and existential risks from artificial intelligence” (Open Philanthropy grant page, archived March 2023). POLITICO tied the grant to the May 2023 extinction statement, which “was organized by the Center for AI Safety, a group that received a $5.2 million grant from Open Philanthropy late last year” (POLITICO, Brendan Bordelon, Oct 13 2023). The funder claimed the statement as field-building: “The Center for AI Safety, one of our AI grantees, made headlines across the globe with its statement calling for AI extinction risk to be a ‘global priority alongside other societal-scale risks’” (Our Progress in 2023).

The relationship has since broken in public. On March 18, 2026, CAIS posted: “To clarify, the Center for AI Safety has not taken funding from Coefficient Giving / Open Philanthropy for years. We believe the effective altruism movement is, unfortunately, controlled opposition. The less influence it has on AI safety, the better” (@CAIS on X, Mar 18 2026).

The rupture is evidence for the mechanism, not against it. A field a grantmaker incubated is now large enough that its flagship institution can denounce the incubator, which is what successful field-building looks like from the inside. CAIS repudiated the movement and kept the agenda: the extinction framing stayed. See the CAIS profile.

3.3 Georgetown CSET: a policy think tank built from scratch

Open Philanthropy gave Georgetown University $55 million over five years to launch the Center for Security and Emerging Technology in 2019, with former IARPA director Jason Matheny as founding director (Philanthropy News Digest). A further $42 million in 2021 took CSET’s funding “to more than $100 million from 2019 to 2025,” in CSET’s own announcement, which names Open Philanthropy as “CSET’s main funder” (CSET). POLITICO: “CSET is funded almost entirely by Open Philanthropy.”

CSET’s answer: spokesperson Tessa Baker said that while Open Philanthropy remains its largest donor, CSET keeps “complete and independent discretion over the research projects we conduct and the recommendations we make,” and that some of its researchers emphasize AI’s present-day harms alongside long-term risks (POLITICO, Oct 13 2023).

3.4 RAND

POLITICO reported $5.5 million (April 2023) to research “potential risks from advanced AI” and $10 million (May 2023) for biosecurity, both to be spent at the discretion of RAND CEO Jason Matheny, who in September 2023 became one of five members of Anthropic’s Long-Term Benefit Trust. RAND spokesperson Jeffrey Hiday said the money “had no bearing on RAND’s recommendations,” that “RAND has strict guidelines that prevent any funder from influencing our analysis or outcomes,” and that Matheny’s trust seat was reviewed for conflicts, calling him a “financially disinterested trustee” (POLITICO, Oct 13 2023). Anthropic’s current list of trustees (Neil Buddy Shah, Richard Fontaine, Ben Bernanke) no longer includes him (Anthropic, Company). See the RAND profile.

3.5 OpenAI

“In 2017, Moskovitz’s Open Philanthropy provided a $30 million grant to OpenAI.” Its spokesperson Mike Levine said it was “to support work on AI safety, not an equity investment” (POLITICO, Oct 13 2023).

By March 2017 OpenAI’s board was four people: Elon Musk, Sam Altman, Chris Clark and Holden Karnofsky. In 2021 Will Hurd and Helen Toner joined “while Karnofsky resigned,” citing “a potential conflict of interest because his wife, Daniela Amodei, a former OpenAI employee, helped to launch the AI company Anthropic” (TechCrunch, “A brief look at the history of OpenAI’s board,” Nov 21 2023). Toner, previously a senior research analyst at Open Philanthropy and later CSET’s director of foundational research grants, took the seat in September 2021 and was one of the directors who voted to remove Altman on November 17, 2023. TechCrunch reports John Loeber’s reading that Karnofsky chose her as his replacement as speculation, not fact. See Holden Karnofsky and Helen Toner.

The traffic ran both ways earlier: “In 2016, OpenAI CEO Sam Altman led a $50 million venture-capital investment in Asana, a software company founded and led by Moskovitz” (POLITICO).

3.6 Anthropic: where “grant” stops being the right word

Open Philanthropy did not grant to Anthropic. The link ran through people and equity.

• Series A, 2021: $124 million. Moskovitz “was also an early backer of Anthropic … participating in a $124 million investment in the company in 2021” (POLITICO). The actor is Moskovitz personally, as an investor.
• Series B, April 2022: $580 million, led by Sam Bankman-Fried, “joined by Caroline Ellison, Jim McClave, Nishad Singh, Jaan Tallinn and the Center for Emerging Risk Research” (TechCrunch, Apr 29 2022). Bankman-Fried was convicted on seven counts of fraud and conspiracy in November 2023 and sentenced in March 2024; nothing here alleges wrongdoing by any recipient of FTX-derived money.
• Board. Luke Muehlhauser, Open Philanthropy’s senior program officer for AI governance, was “one of Anthropic’s four board members”; Levine said he “holds no financial interest” (POLITICO). He left on May 29, 2024, deciding “to step down from his Board role to focus on his work at Open Philanthropy,” the day the Long-Term Benefit Trust appointed Jay Kreps (Anthropic, May 29 2024).
• Family and staff. Karnofsky, then Open Philanthropy’s director of AI strategy, is married to Anthropic president Daniela Amodei; CEO Dario Amodei is his former roommate (POLITICO). Karnofsky stepped down from Open Philanthropy “to work full-time on AI safety” and joined Anthropic in January 2025 as “a member of the technical staff” reporting to chief science officer Jared Kaplan, working “on the responsible scaling policy and other aspects around safety planning ahead of highly advanced AI systems,” per an Anthropic spokesperson (Fortune, Feb 13 2025).
• Moskovitz’s answer: any returns from his Anthropic investment “will be entirely redirected back into our philanthropic work … My goal in investing stems from the exact same place as our non-profit work: addressing safety and security issues around transformative AI coming to market” (POLITICO).

The careful formulation is tighter than “Open Philanthropy funded Anthropic”: the grantmaker’s principal donor was an early equity investor, the grantmaker’s staff sat on the board, and its former chief executive married into and then joined the company. As of October 2026 Anthropic’s board is Dario Amodei, Daniela Amodei, Yasmin Razavi, Reed Hastings, Chris Liddell and Vas Narasimhan (Anthropic, Company); the remaining tie is Karnofsky’s, as staff. See the Anthropic profile.

3.7 Personnel placement: the Horizon Institute for Public Service

All of this subsection is from POLITICO, Brendan Bordelon, “How a billionaire-backed network of AI advisers took over Washington,” Oct 13 2023.

• Open Philanthropy “effectively created” the Horizon Institute for Public Service in 2022 and “set aside nearly $3 million to pay for what ultimately became the initial cohort of Horizon fellows,” whose salaries it funds in congressional offices, federal agencies and think tanks.
• Placements reported: fellows in the offices of Sens. Martin Heinrich, Mike Rounds, Todd Young and Richard Blumenthal (the last a Horizon AI fellow who came directly from OpenAI); current and former fellows at the Department of Defense, DHS and State, the House Science Committee and Senate Commerce Committee, and at RAND and CSET. Photographs from the closed September 2023 Senate AI Insight Forum show at least two Horizon AI fellows in attendance.
• The mechanism is lawful: the Intergovernmental Personnel Act of 1970 lets nonprofits pay the salaries of fellows working in Congress or the federal government.
• The screening question, both sides: an Open Philanthropy document last updated August 2022 said the fellowship would consider candidates “who share Open Philanthropy’s interests in global catastrophic risks and technology’s long-term impacts,” while also valuing “ideological diversity” and not expecting candidates “to share any particular worldview or affiliation”; POLITICO reported the first clause no longer appears on Horizon’s fellowship page. Horizon co-founder Remco Zwetsloot said the program is “not for the pursuit of particular policy goals,” does not screen for belief in long-term AI risks, and does not direct fellows to particular offices; Levine said Open Philanthropy “did not play an active role in screening, training, or placement of fellows.”

3.8 Capacity building: manufacturing the researchers

Grants to institutions buy output. Grants to fellowships and scholar programs buy the supply of people qualified to hold the opinions, and that compounds.

The funder tells the story of ML Alignment & Theory Scholars (MATS) itself. “We made our first $250,000 grant to the program that would eventually become ML Alignment & Theory Scholars (MATS) in 2019, when it was a side project by some students affiliated with the Stanford Existential Risks Initiative.” The first cohort “had 5 fellows and no permanent full-time staff.” MATS now runs “a year of around 100 scholars with an admission rate of 4-7%,” and reports “over 80% of their alumni are now working full-time in AI safety and security (accounting for a meaningful portion of safety staff at some of the biggest companies and government institutes).” A footnote discloses that MATS’s “few dozen full time staff” include Berger’s wife, “who joined in December as a research manager after a decade as a data scientist at Uber” (Berger, 2025 Letter from the CEO). The letter’s own headline for the portfolio: “Beginning in 2015, seven years before the launch of ChatGPT, we jumpstarted the field of AI safety and security.”

Adjacent grants on the same archived grants page as the CAIS award: $2,047,268 to the Berkeley Existential Risk Initiative for its MATS collaboration with the Stanford Existential Risks Initiative; $1,623,938 for the 2022 cohort of the Open Philanthropy Technology Policy Fellowship; GBP 53,700 (about $63,839) to AI Safety Hub’s Safety Labs program (archived grants page). An 80 percent conversion rate into the field is the number to keep: this is not funding research, it is minting a profession.

4. The asymmetry, in the funder’s own words

The strongest evidence for the asymmetry is not a critic’s characterization. It is the grantmaker’s published reasoning.

(a) The stated priority. “One thing that hasn’t changed since our early days is that we are focused on the highest-stakes risks posed by transformative AI, like catastrophic misuse or loss of control.”

(b) The stated posture toward present-harm regulation. “We want regulation to be proportionate to the risks posed and often feel that proposed regulation targeting smaller-scale harms is overly restrictive. … in practice we are typically skeptical of regulation that we don’t think will move the needle on reducing the worst-case risks.” (Both: Coefficient Giving, “Our Approach to AI Safety and Security”.)

(c) Which bucket was cut first when money was short. After the endowment fell, the funder wrote: “The decline of our available assets should disproportionately affect funding for GHW relative to GCR” (global health and wellbeing relative to global catastrophic risks), because health opportunities cluster near the cost-effectiveness bar while catastrophic-risk opportunities sit far above and below it. It “raised our cost-effectiveness bar for GCR spending” after FTX collapsed, then “decided to keep that new higher bar, while also aiming to roughly double our GCR spending over the next few years” (Our Progress in 2023). When money contracted, the measured bucket (bednets, deworming, vitamin A) was designated to absorb the loss, and the speculative bucket was slated to double. That is the Bostrom inequality operating as written, and it was published, not inferred.

(d) Then the money came back, and the reason given was AI. On July 23, 2026 Coefficient raised its 2026 commitment to GiveWell’s recommendations from $175 million to $1 billion, which “will nearly double the cumulative funding we have previously committed to GiveWell’s recommendations over the last decade,” and lowered its health bar “from ~2,000x to ~1,000x,” which “returns the bar to its 2022-23 level.” It calls this “a one-off surge rather than a new steady state,” driven by “growing expectations of future giving from Good Ventures and other funders.” The second reason it names: “we and Good Ventures generally expect that transformative AI, if navigated safely, could make the world as a whole massively richer and reduce the need for global health and development giving in the future. So both a higher credence in the possibility of transformative AI and shorter estimated timelines push us to spend down a fixed pot of GHW funding faster today” (Berger and Reid, “Increasing Our 2026 Allocation to GiveWell’s Recommendations to $1 Billion,” Jul 23 2026). In September it said it is “significantly accelerating its global health and wellbeing giving in 2026-27” (Coefficient Giving, Sep 8 2026).

This is the strongest fact against reading (c) as a permanent tilt: in 2026 the measured bucket got what the funder’s homepage calls “the largest single shift in this direction so far.” It is also the same premise in different clothes. The 2023 cut and the 2026 surge are both justified by the funder’s view of AI: first as the risk that outranks bednets, then as the event that will make bednets unnecessary, so the health money should be spent before it arrives. Either way, the AI forecast sets the health budget.

(e) The AI-side spending, as reported by the funder. Roughly $25 million in 2024 on LLM-agent capability benchmarks, “some of which are already being used by the U.S. and UK government, OpenAI, and Anthropic”; “roughly $50 million” committed to technical AI safety research in 2024; an open RFP to allocate “at least $40 million” more (Our Progress in 2024 and Plans for 2025). That RFP “helped more than triple our total TAIS giving from $46 million in 2024 to $157 million in 2025, across more than 200 grants.” Berger also writes that “I have been struck by how much more accurate the predictions of people with short AGI timelines have been than mine,” that fast progress “increasingly looks to me like the default path,” and that the funder is “more directly competing with companies for talent” whose safety staff “could earn millions of dollars a year” (2025 Letter from the CEO).

(f) The funder’s case that AI safety is underfunded, at full strength. “Philanthropic funding for climate risk mitigation … was ~$9-15 billion in 2023 (the most recent data available), which we estimate to be ~20x as much as philanthropic funding for AI safety and security in 2024” (Coefficient Giving, “AI Safety and Security Need More Funders,” Oct 2 2025). One twentieth of $9-15 billion implies roughly $450-750 million a year of AI-safety philanthropy worldwide, a rounding error against the $200 billion-plus of annual AI data-center capital expenditure the same post cites. So the asymmetry this page documents is not “x-risk is drowning in money while present harms starve.” It is narrower: inside the small pool of money that existed for thinking about AI at all, one decision-maker’s cost-effectiveness bar set the split, and that bar was built from a premise that does not let present harms compete. Both things are true.

(g) The other side of the ledger is no longer small. On October 14, 2025, ten foundations, co-chaired by MacArthur and Omidyar Network, with Ford, Mellon, Mozilla, Kapor, Doris Duke, Lumina, the Siegel Family Endowment and Packard, launched Humanity AI, a $500 million five-year initiative (MacArthur Foundation). Its first money went out on May 12, 2026: “more than $18 million in new pooled grants,” including $500,000 each to the AI Now Institute and the Center for Democracy and Technology, $3 million to create “AI Civics” led by Data & Society with the Digital Public Library of America, and a $10 million open call (Mellon Foundation, May 12 2026).

(h) The present-harms institutions at founding scale. The Distributed AI Research Institute (DAIR), founded by Timnit Gebru in December 2021 “after being pushed out of Google” (TechCrunch, which reports she was fired), launched with $3.7 million from the Ford Foundation, the MacArthur Foundation, the Kapor Center and the Open Society Foundations (TechCrunch, Dec 2 2021; DAIR). In 2019 launching one AI-policy think tank cost the x-risk funder $55 million, more than fourteen times DAIR’s entire launch funding two years later. By 2025 the present-harms side had assembled a $500 million coalition. The asymmetry was largest exactly when the policy agenda was being set, roughly 2017 to 2023, and it has narrowed since. State it as a period, with the period named.

Why there is no annual budget to compare. Neither AI Now nor DAIR files a return covering those years, for two different reasons. AI Now was founded at NYU in 2017, so its budget sat inside NYU’s accounts; it incorporated only in 2025, and as a pair: AI Now Institute Inc, a 501(c)(3) ruled in June 2025, and AI Now Policy Institute Inc, a 501(c)(4) ruled two months earlier, which separates research from lobbying on paper while sharing a name. DAIR has never been a filing entity: it is a sponsored project of Code for Science and Society, which it joined the day it announced itself, so its money is inside CS&S’s return, an entity whose total revenue across all its projects ran $3.1 million (FY2020), $5.8 million (FY2021), $7.6 million (FY2022) and $17.0 million (FY2023). The $3.7 million is a launch total, not an operating budget, and the two must not be compared as if they were.

5. The disclosure surface: the receipts are being taken down

In “Our Progress in 2024 and Plans for 2025,” the funder wrote that “our grants database offers an increasingly inaccurate picture of our overall work. (Our database generally does not include funding we advise from non-GV donors …) Since our current database is almost totally redundant with Good Ventures’ database, we’re thinking about potential changes and may deprecate it” (Our Progress in 2024).

The individual grant pages are gone from the live web. Old openphilanthropy.org grant URLs redirect to a generic fund landing page; coefficientgiving.org/grants/ returns 404; Good Ventures’ per-grant pages redirect to a generic giving page. Meanwhile the money routed through donors who never appear in the database grew: “over $100 million” in 2024 and “over $200 million” in 2025 (2025 Letter from the CEO).

Every pre-2025 dollar figure on this page now rests on an archived copy, the recipient’s own announcement, or a journalist who read the page while it was up. A public, per-grant, searchable record was more than most foundations ever offered, and it is being retired just as more of the money moves through advised donors who never appeared in it. The stated reason (fungibility, not taking credit for partners’ giving) is coherent. The effect is that the pipeline is harder to audit from outside, whatever the intent.

6. The critique, at full strength

The eugenics-lineage argument. Timnit Gebru and Émile P. Torres argue that the normative framework motivating the drive to build AGI “is rooted in the Anglo-American eugenics tradition of the twentieth century,” that “many of the very same discriminatory attitudes that animated eugenicists in the past (e.g., racism, xenophobia, classism, ableism, and sexism) remain widespread within the movement to build AGI, resulting in systems that harm marginalized groups and centralize power, while using the language of ‘safety’ and ‘benefiting humanity’ to evade accountability,” and that undefined systems like “AGI” cannot be appropriately tested for safety at all (Gebru and Torres, “The TESCREAL bundle,” First Monday 29(4), Apr 14 2024).

The crowding-out argument, from named researchers (all POLITICO, Oct 13 2023):

• Deborah Raji (UC Berkeley): the network’s fixation on speculative harms is “almost like a caricature of the reality that we’re experiencing”; a licensing regime pushed by Open Philanthropy-funded experts “concentrates existing monopolies and entrenches them even further”; and “They’re so organized, they’re so homogenous in their ideology. And so they’re much more coordinated.”
• Suresh Venkatasubramanian (Brown; co-author of the White House Blueprint for an AI Bill of Rights): the focus on cataclysmic AI risk is “speculative science fiction” bordering on “fearmongering”; the strategy is “the same playbook … planting various people in various places with appropriate funding”; the message to lawmakers is “‘You should be scared out of your mind, and only I can help you.’”
• Tim Stretton (Project On Government Oversight): “We have [the] AI [industry] inserting its staffers into Congress to potentially write new laws and regulations around this emerging field. That is a conflict of interest.”
• David Skaggs (former Representative; former chair, Office of Congressional Ethics): “There should be full disclosure of where these folks are coming from.”

The tally. Nirit Weiss-Blatt, an avowed critic, totals the money from public databases: Open Philanthropy “fueled more than $330 million for the AI x-risk cause”; Vipul Naik’s donations list totals $387,097,518; an EA Forum-hosted AI safety funding spreadsheet totals $499,357,062 across 2014-2023, explicitly excluding “near-term AI Safety concerns such as effects on labor market, fairness, privacy, ethics, disinformation”; CSET is the largest single recipient at more than $100 million. She quotes Sayash Kapoor (Princeton): interest in AI x-risk “did not arise organically … It has been very strategically funded by organizations that make x-risk a top area of focus” (Weiss-Blatt, AI Panic, Dec 5 2023). Her numbers trace to named public databases; her characterizations are hers.

Withdrawal from inside the donor class. See section 7: Vitalik Buterin’s March 2026 break with the Future of Life Institute.

7. 2026: the money goes electoral

From 2017 to 2023 the pipeline ran through grants, fellowships and staff. By 2026 the same policy fight was also being fought with political money, from the AI companies themselves and from both sides of the safety question.

Anthropic’s $40 million. On February 12, 2026 Anthropic announced it was “contributing $20 million to Public First Action, a new bipartisan 501(c)(4) that will support public education about AI, promote safeguards, and ensure America leads the AI race” (Anthropic, Feb 12 2026). On July 21 it added another $20 million, “bringing our total support to $40 million,” and said both gifts “were made exclusively to support Public First Action’s public education and policy mission, and cannot be used to influence the election of any candidate for federal, state, or local office” (Anthropic, Jul 21 2026). Reuters reported that two former members of Congress launched Public First Action “to counter the political group Leading the Future,” and that it “supports two political action committees that give money to candidates,” which per FEC records have given to Sen. Pete Ricketts, House candidate Brian Poindexter and Senate candidate Kevin Hern (Reuters via Cybernews, Jul 22 2026).

The other side’s purse. Leading the Future is “a super PAC backed by the venture capital firm Andreessen Horowitz, OpenAI president Greg Brockman, Palantir co-founder Joe Lonsdale, and AI search engine company Perplexity.”

The test race: New York’s 12th, June 23, 2026. Assemblymember Alex Bores, sponsor of New York’s RAISE Act (which “would require leading AI companies to develop, publish, and adhere to formal safety protocols”), ran for Jerry Nadler’s open seat. Outside groups spent “more than $40 million on the race.” Leading the Future “spent $8 million to make sure Bores did not win”; “a combination of super PACs and AI safety groups spent roughly $20 million to support Bores, with the bulk of that money coming from an organization backed by Anthropic,” alongside Dream NYC, “backed by Anthropic AI safety researcher Daniel Ziegler,” and You Can Push Back, “backed by crypto billionaire Chris Larsen.” Bores lost to Assemblymember Micah Lasher, himself a co-sponsor of the RAISE Act, who said: “I have some news for the two big AI companies that have taken such an unusual interest in who won this congressional seat … I won’t be taking my cues from either of you when it comes to protecting our kids, our jobs, and our environment” (Gizmodo, Ece Yildirim, Jun 24 2026). Spending totals differ between outlets; the FEC’s independent-expenditure filings are the record of account.

The Future of Life Institute’s pivot, and its donor’s break. On March 13, 2026 Vitalik Buterin wrote that FLI had pitched him in 2021 “a comprehensive roadmap that focused on improving all major existential risks (bio, nuclear, AI…) as well as general pro-peace and pro-epistemics initiatives”; that it liquidated roughly $500 million from his SHIB gift where he had expected $10 million to $25 million; and that it then made “an internal pivot” toward “cultural and political action as a primary method.” His objection: “My worry is that large-scale coordinated political action with big money pools is a thing that can easily lead to unintended outcomes, cause backlashes, and solve problems in a way that is both authoritarian and fragile, even if it was not originally intended that way.” FLI’s reasoning, as he relayed it: “the situation has changed greatly since 2021, AGI is coming very soon, and their pivot is needed to affect the world fast enough, and to counteract the lobbying war chests of large AI companies.” He praised FLI’s “pro-human AI declaration” for uniting “conservatives, progressives and libertarians, America, Europe, and China … atheists and the Pope” (Buterin on X, Mar 13 2026; The Block). FLI president Anthony Aguirre answered that FLI has “immense respect” for Buterin’s views: “We support a great deal of work focused on coordination and epistemics, as well as on technical solutions around compute governance, secure chips and formal verification that systems work as intended” (Cointelegraph, Mar 13 2026, updated Mar 16).

The 2023 critique was that one funder’s staff arrived in Congress first. The 2026 record is that the frontier labs now fund the fight directly, on both sides: Anthropic’s money for safeguards, OpenAI’s president and a16z’s money against them. The donor whose single 2021 gift became roughly $500 million of FLI’s money calls the turn to political money “authoritarian and fragile.” And the first big test went to a RAISE Act co-sponsor who told both companies he would not take their cues. The money buys a fight, not an outcome.

8. The defense, at full strength

• On motive. Open Philanthropy’s Mike Levine: “The idea that we, our funders, or our grantees are motivated by pecuniary interests is misguided.” It funds “a variety of grantees working on AI policy, including many who don’t particularly share our concerns around catastrophic risks” (POLITICO).
• On regulatory capture, the specific counter-argument. Levine: “Regulations targeting these frontier models would create unique hurdles and costs specifically for companies that already have vast resources, like OpenAI and Anthropic, thus giving an advantage to less-resourced start-ups and independent researchers who need not be subject to such requirements (because they are building less dangerous systems).” (Emphasis original; POLITICO notes many AI experts dispute this.)
• On viewpoint monoculture, with a receipt. “Support a wide range of experts, including those who disagree with us and each other.” The funder paid for both a paper finding language models could help people with little lab training develop new pathogens and a RAND paper finding current models did not substantially raise biological-attack risk: “Both papers were produced by groups we’ve funded.” It ran an essay contest to change its own views and “awarded the top prize to the Forecasting Research Institute for a paper that strongly challenged our thinking at the time” (Our Approach to AI Safety and Security).
• On the counterfactual. CEO Alexander Berger: “Without the efforts we’ve made to develop the field of AI risk, I think that fewer people with AI experience would have been positioned to help, and policymakers would have been slower to act. I’m glad that we were paying attention to this early on, when it was almost entirely neglected by other grantmakers” (Our Progress in 2023).
• On self-criticism. The same post concedes the Wytham Abbey purchase was made “at a time when FTX was making huge commitments to the GCR community, which made resources appear more abundant and lowered our own bar,” and that “the original grant would no longer have risen to the level where we would want to provide funding.”
• On the “strategic funder” trap. Berger, 2026: “I also want us to avoid the common ‘strategic funder’ trap of thinking we have all the answers and just need to slot grantees into our vision, almost like subcontractors” (2025 Letter from the CEO).
• On the “vast pool of money” framing. See 4(f): the funder’s estimate is that AI-safety philanthropy runs at about one twentieth of climate philanthropy.
• On the fellows program. Zwetsloot names the vacuum: “In an ideal world, all the relevant government offices would have permanent in-house staff with critical subject-matter expertise on emerging technologies – which they need in no small part to not be overpowered by corporate lobbying.” He hopes programs like Horizon’s will one day be “no longer necessary.” Other organizations also fund congressional tech fellows, including TechCongress and the AAAS (POLITICO).
• From inside the movement. Jaan Tallinn after the November 2023 OpenAI board crisis: “The OpenAI governance crisis highlights the fragility of voluntary EA-motivated governance schemes. So, the world should not rely on such governance working as intended” (Semafor, quoted in Weiss-Blatt). See the Jaan Tallinn profile.

The strongest form of the defense concedes the mechanism and disputes the valence: a small number of people decided this mattered before anyone else did, funded it into existence and staffed it, and that is what philanthropy is for in a neglected field. The critique does not defeat that. What survives is narrower: a field brought into existence by one bar-setter inherits that bar-setter’s premises, including the premise that documented present harm cannot outrank speculative future harm, and no amount of good faith in the funder changes what that premise does to the agenda downstream.

9. The mechanism, stated plainly

Six moves, each lawful, ordinary and defensible on its own. The apparatus is the sequence.

1. A premise that makes the arithmetic one-sided. Bostrom 2013: any nonzero extinction-risk reduction outweighs any definite present good (section 1).
2. A concentrated purse that accepts the premise. Good Ventures through Open Philanthropy: over $7 billion since 2014, over $1 billion in 2025, one couple’s fortune behind most of it (section 2).
3. Institutions built from zero to the funder’s specification. CSET at $100 million-plus, CAIS at $5.16 million, MIRI at roughly 60 percent dependency in 2020-21 (section 3).
4. A profession minted. MATS from a $250,000 student side project to 100 scholars a year, 80 percent into the field (3.8).
5. Placement into the offices that write the rules. Horizon fellows in four named Senate offices, three federal departments, two committees, RAND and CSET, lawful under the IPA of 1970 (3.7).
6. The bar, applied under stress. When assets fell, the measured bucket was designated to absorb the cut and the speculative bucket was slated to double; when they rose, the health money was pulled forward on the strength of the same AI forecast (4c, 4d).

The result is not that anyone lied. By the time Congress went looking for someone who understood AI, the people available to ask had been selected, trained, paid and placed by an organization whose founding premise was that the question worth asking was extinction. The pipeline did not suppress the present-harms question. It arrived first, with staff.

Verdict

The EA to AI-safety funding pipeline is a real apparatus and an unusually legible one, because for most of its life it published its own receipts. What they show is not a plot. It is a premise, executed with discipline, by people who wrote the premise down.

One. The asymmetry was real and was largest when the agenda was set, roughly 2017 to 2023. It has narrowed since: Humanity AI’s $500 million is not a rounding error, the pipeline’s own flagship grantee now calls the movement “controlled opposition,” and in 2026 the funder moved $1 billion toward GiveWell’s charities, with its AI forecast among the stated reasons. Write the asymmetry as a period with a name on it, not as a permanent condition.

Two. The funder’s claim that AI safety is underfunded is also true on its own published arithmetic, and does not contradict the first point. The scarce resource was never dollars measured against Nvidia’s revenue. It was the question: who got paid to decide which risk the word “AI” would mean in a Senate hearing room. For years that was allocated by one bar, held by one organization, funded by one couple.

Three. The audit trail is closing. The grant pages are delisted, the database “may” be deprecated, and over $200 million a year now moves through advised donors who never appear in it.

Four. The fight has gone electoral. The labs now pay for it directly, on both sides, and the first big test went to a candidate who told “the two big AI companies” he would take cues from neither.

Related

• The AI-Worldview Poles: the accelerationist, doomer, present-harms and national-security camps
• Watching the Watchers: the people the money placed, mapped machine by machine
• AI Safety Theater: safety as the capture surface
• MIRI and the Off-Switch Problem: the oldest grantee, in its own words
• The Funding Ratchet: donor-advised funds and fiscal sponsorship in political advocacy
• Revolving Door · AI Governance Tracker · Ratchet Atlas
• Profiles: Holden Karnofsky · Helen Toner · Jaan Tallinn · Center for AI Safety · Anthropic · RAND · Joe Carlsmith