Founders Pledge

Stated Theory of Change

Founders Pledge states two complementary theories of change for AI safety:

1. Diversify AI safety funding. Approximately 80% of philanthropic AI safety funding comes from Open Philanthropy/Coefficient Giving and the Survival and Flourishing Fund. FP argues this concentration creates single points of failure, worldview lock-in, and geographic bias. FP's mechanism: channel tech entrepreneurs' pledged equity into AI safety through expert-recommended organizations, creating a second pipeline of donors independent of the CG/SFF duopoly.

2. Fund governance, resilience, and Track 2 diplomacy. Within AI safety, FP bets on governance and policy interventions rather than technical alignment research. The specific theory: great power competition between the US and China creates a "race to the bottom" on AI safety. Backchannel diplomacy, government capacity-building, and resilience infrastructure can slow this race and create cooperation pathways. FP uniquely emphasizes "resilience" (preparing for scenarios after deployment fails) as a neglected complement to prevention-focused approaches.

The causal chain: Tech entrepreneur pledges equity -> FP research identifies high-impact AI safety orgs -> Member donations flow to recommended orgs -> Those orgs influence government policy, build diplomatic channels, and advance technical safety -> AI risk is reduced.

Revealed Theory of Change

FP's actions reveal several important divergences from the stated AI safety theory of change:

AI safety is a small part of a much larger operation. FP moved $238M to charities in 2024. The GCR Fund -- which covers AI, nuclear, biosecurity, and general GCR -- had $5.1M contributed and $2.1M granted. Even assuming all GCR grants went to AI (they don't), AI safety is roughly 1% of FP's total money moved. The real center of gravity is climate ($90M+ in anonymous donations) and global health (via GiveWell partnership). AI safety is a genuine intellectual commitment but not the financial core of the organization.

The real product is donor cultivation, not research. Of ~75 staff, roughly 13 are researchers. The rest are community managers, growth leads, advisors, product managers, grants officers, and operations. The Glassdoor criticism -- "the research team ultimately serves high-value donors" -- resonates with the staffing structure. FP's primary function is getting tech founders to give more, and to give to effective causes. The research exists to make the giving advice credible.

The governance/policy tilt is revealing. 6 of 9 recommended AI organizations are governance/policy-focused. Only FAR AI and CHAI do technical safety. This reflects either (a) a genuine conviction that governance is the highest-leverage intervention, (b) the GCR team's expertise profile (Ruhl was an international security/policy scholar, not a technical AI researcher), or (c) both. The absence of technical alignment, interpretability, or evaluation-focused grantees (Redwood, METR, ARC, Anthropic's alignment team) is a clear choice.

Key personnel departures may have hollowed out AI capacity. Christian Ruhl (GCR Lead, prolific researcher with publications in Lawfare, Bulletin of Atomic Scientists, and The Atlantic) departed to Coefficient Giving in June 2025. Tom Barnes (author of the AI report, AI policy secondee to UK government) appears to have moved to AISI. The new GCR Lead, Libbie Prescott, has a molecular biology PhD and government policy background -- competent but different expertise. If Barnes has indeed left, FP currently has no dedicated AI safety researcher.

Climate is where FP's institutional character shines. The $90M in anonymous climate donations, the Deploy/US right-leaning group strategy, and the willingness to fund across political lines reveal an organization that can take genuinely contrarian positions. This institutional character -- making unconventional bets in neglected areas -- is exactly what FP claims to do in AI safety but where the evidence is thinner.

Key Assumptions

Assumption 1: Tech entrepreneurs will direct meaningful amounts to AI safety.

• Evidence for: 60% of members with liquidity invest in "long-term" causes. FP's community programming actively shifts members toward EA-aligned giving.
• Evidence against: Of $1.7B deployed, the AI safety fraction is unknown but likely small. Members have choice, and many give to causes (local arts, universities) that FP wouldn't recommend. The pledge minimum is only 2%.
• Testable? Yes -- FP could disclose the AI safety fraction. They haven't.
• If wrong: FP's AI safety impact is limited to the small GCR Fund, not the broader pledge pool.

Assumption 2: FP's research capacity is sufficient to identify high-impact AI safety organizations.

• Evidence for: The ITN framework is sound. The AI report articulates a coherent theory of change. FP recommends several orgs that CG also funds (CNAS, FAR AI, CSET).
• Evidence against: The AI report was authored by a PPE graduate with limited technical background. Glassdoor reviews describe "pseudo-research" and "uninformed literature review." SoGive found systematic overoptimism in the climate evaluation. The two most qualified GCR/AI researchers have departed. No external review of AI safety recommendations exists.
• Testable? Yes -- an independent review of FP's AI portfolio against CG's, Longview's, and SFF's would reveal overlap and differentiation.
• If wrong: FP may be directing money to suboptimal AI safety organizations, reducing the marginal impact of each donated dollar.

Assumption 3: Governance and Track 2 diplomacy are the highest-leverage AI safety interventions.

• Evidence for: The US-China competition dynamic is real. Governments are under-resourced. The CNAS military AI workshop was the first of its kind ("more person-hours were spent thinking about autonomous systems CBMs than cumulatively had been spent before"). CLTR influenced UK MoD AI Strategy.
• Evidence against: Technical alignment progress may be more decision-relevant for long-term AI safety. Governance interventions can be slow, reversible, and vulnerable to political shifts. Track 2 diplomacy's causal chain to actual risk reduction is long and uncertain.
• Testable? Partially -- outcomes of specific diplomatic initiatives can be tracked, but the counterfactual (would risk have been higher without them?) is inherently uncertain.
• If wrong: FP's AI safety portfolio is optimized for the wrong lever, missing the technical safety and alignment work that could more directly reduce catastrophic risk.

Assumption 4: Diversifying AI safety funding away from CG/SFF concentration is valuable.

• Evidence for: CG itself agrees, publishing "AI Safety and Security Need More Funders." Monoculture funding creates blind spots. FP's donor base (tech entrepreneurs) is different from CG's (Moskovitz wealth) or SFF's (Tallinn wealth).
• Evidence against: If FP's AI portfolio heavily overlaps with CG's recommendations (both fund CNAS, CSET, FAR AI), the "diversification" is nominal -- different donor pipeline, same grantees. True diversification would require funding approaches CG won't fund.
• If wrong: FP adds intermediary overhead without genuine strategic diversification.

Strengths

1. Genuinely innovative philanthropic infrastructure. The pledge model -- getting founders to commit equity before liquidity -- is a genuine innovation that taps a different donor population than traditional philanthropy. $12.9B pledged is a massive pool, even if conversion to AI safety funding is low. Pledge Ventures (85% carry to charity) is creative institutional design.

2. Institutional willingness to take contrarian positions. The climate strategy (funding right-leaning groups, backing Deploy/US) demonstrates FP can make bets that most EA-adjacent organizations would consider too politically uncomfortable. If this institutional character extended to AI safety, it could enable funding approaches that CG and Longview won't touch.

3. Track 2 diplomacy is distinctive and neglected. FP's emphasis on US-China backchannel diplomacy for AI risk is genuinely differentiated. Few other AI safety funders work in this space. The Conclave on Great Powers and Extreme Risks (co-designed with NAS) and the Carnegie Endowment AI cooperation project are specific contributions that other funders aren't making.

4. Significant leverage through advisory function. FP's influence extends beyond its direct grants. The $2.5M Carnegie "Averting Armageddon" project was advised by FP. Advisory recommendations to members with liquidity could direct far more capital than the GCR Fund alone.

5. Financial sustainability planning. Pledge Ventures, member operational funding (45.5%), and diversified foundation support reduce the existential risk to FP's operations. The "loss-making by ~$10M/year" model is sustainable as long as fundraising continues, and the move from founder-CEO to professional CEO reduces key-person risk.

Weaknesses and Risks

1. AI safety research capacity may be effectively zero. If both Christian Ruhl and Tom Barnes have left, FP has no dedicated AI safety researcher. Libbie Prescott (GCR Lead) has biosecurity expertise, not AI. The GCR Fund includes AI as one of five risk domains, but without a specialist, the ability to identify high-impact AI safety funding opportunities is severely limited.

2. No independent evaluation of AI safety recommendations. SoGive critiqued FP's climate work and found systematic overoptimism. No one has done the equivalent for AI safety. FP recommends 9 AI orgs and directs donor capital to them without any published external validation. For a meta-charity whose value proposition is "help you give better," this is a critical gap.

3. Donor-serving incentive structure. FP's primary relationship is with tech founder donors, not with AI safety grantees or the AI safety research community. If donor preferences diverge from the most impactful AI safety interventions (e.g., if founders prefer funding governance organizations that engage constructively with AI labs rather than organizations that challenge them), FP's recommendations may follow donor comfort rather than impact analysis.

4. Scale of AI safety funding is not yet consequential. $2.1M in GCR grants across 12 organizations in 2024, with AI as one of five domains, means FP's direct AI safety funding is likely in the hundreds-of-thousands-of-dollars range. This is valuable as seed funding for small organizations but not consequential for the field's most significant funding needs.

5. The Jaan Tallinn paradox. FP frames its AI safety value as diversifying funding away from the CG/SFF duopoly while having Jaan Tallinn (the SFF funder) as an advisor. This doesn't invalidate FP's work but undermines the independence claim. Compare to Longview Philanthropy, which faces the same structural irony (funded by CG to be an "independent alternative" to CG).

6. Governance gaps. No published conflict of interest policy. Trustee Andreas Haug chairs both the UK board and the Pledge Ventures Investment Committee. US board members serve 0.45 hours/week. No board members with AI safety expertise. Multi-entity structure (UK, US, Germany) without consolidated financial reporting.

Cross-References

Longview Philanthropy: Direct peer as a funder intermediary trying to diversify AI safety funding away from CG concentration. Both are funded by CG for operational support. Key differences: Longview targets UHNW donors ($100K+), has 6 dedicated AI grantmakers, and raised $20M for its Frontier AI Fund. FP channels tech founder pledges and has $5.1M in its GCR Fund (covering all risk domains). Longview's AI safety funding is likely 10-20x FP's.

GiveWell: FP partners with GiveWell for global health research and recommends GiveWell top charities. GiveWell does not fund AI safety. The relationship gives FP credibility on its global health recommendations but doesn't extend to GCR/AI.

Coefficient Giving: FP's largest named operational funder ($15.3M). CG also employs FP's former GCR Lead (Christian Ruhl). CG funds the organizations FP recommends (CNAS, FAR AI, CSET). The overlap raises the question: is FP providing genuinely independent evaluation, or routing donors to CG-validated grantees?

CLTR, CNAS, FAR AI: Three of FP's recommended AI orgs. CLTR has influenced UK AI policy. CNAS has produced the military AI CBM workshop. FAR AI organized Western-Chinese AI safety dialogues. These are credible organizations. The question is whether FP's recommendations add information beyond what CG has already determined.

What Would Change This Assessment

• If FP disclosed the AI safety fraction of member donations: Knowing whether $5M or $50M of the $238M moved in 2024 went to AI safety would fundamentally change the assessment of FP's AI safety significance.
• If FP hired a senior AI safety researcher to replace Ruhl/Barnes: This would signal continued commitment and rebuild the analytical capacity needed for credible AI safety grantmaking.
• If an independent reviewer evaluated FP's 9 AI safety recommendations: A SoGive-style analysis of FP's AI portfolio would provide the external validation that currently does not exist.
• If FP funded confrontational AI safety approaches: Funding organizations like PauseAI, compute governance advocates, or AI watchdog groups would demonstrate genuine independence from its tech-founder donor base.
• If the GCR Fund grew significantly: A $20M+ GCR Fund (comparable to Longview's FAIF) would make FP a consequential AI safety funder rather than a small supplement.

Self-Critique

Sources I should have checked but didn't:

• The full 140-page "Navigating Risks from Advanced AI" PDF (not extractable). This is the most important missing source -- it would reveal whether FP's AI safety analysis has genuine depth or is the "pseudo-research" Glassdoor alleges.
• Christian Ruhl's EA Forum AMA (blocked domain). Would reveal candid discussion of GCR Fund strategy.
• Confirmation of Tom Barnes' current employment at AISI. If he's still at FP part-time, the research capacity assessment changes.
• UK Charity Commission annual report narratives (financial tables captured but qualitative governance details may not have been).

Where this analysis is potentially biased:

• I may be too harsh on FP's AI safety research capacity. The team may have depth that doesn't appear on Glassdoor or in public profiles. The full AI report might be rigorous.
• I may overweight the personnel departures. Academic-to-practice organizations regularly cycle researchers. Ruhl moving to CG could reflect FP's strength (it trained someone CG wanted) rather than weakness.
• The SoGive critique of climate work may not generalize to AI safety work, which uses different (qualitative/rhetorical) methodology.

What a thoughtful person who disagrees would say: "You're evaluating FP's AI safety work as if it were a standalone AI safety funder. It's not. FP's value-add is the pipeline -- connecting tech entrepreneur wealth to effective giving, including AI safety. The GCR Fund is small, but the advisory influence on member donations could be orders of magnitude larger. The Track 2 diplomacy work is genuinely novel. And the climate program demonstrates that FP can execute brilliantly when it has the right people and capital. If AI safety funding to FP grows like climate funding did, the assessment would be completely different."

My single weakest claim: That FP's AI safety research quality is questionable. I'm relying on Glassdoor reviews (inherently biased), a climate-domain critique (may not generalize), and staff departures (may be normal turnover). I have not read the full AI report. The claim could be entirely wrong if the full report demonstrates genuine analytical depth.

What information would most change my view: The full AI report PDF, read in detail. If it demonstrates the kind of analytical rigor that Christian Ruhl's published work shows -- specific causal models, quantified uncertainty, engagement with counterarguments, novel analysis -- then my assessment of FP's AI safety research capacity would shift significantly upward. If it reads like a literature review with recommendations bolted on, the Glassdoor critique would be substantiated.