Coefficient Giving (formerly Open Philanthropy)

Stated Theory of Change

CG's stated AI safety theory of change is: philanthropic capital, directed through a research-intensive process of identifying high-impact opportunities, can fill the structural gap between the massive commercial incentives driving AI capabilities and the weak incentives for AI safety. This works through three mechanisms:

1. Visibility: If society can see what AI systems are capable of and what risks they pose, policymakers and the public can make informed decisions. CG funds evaluations, benchmarks, and public data infrastructure to close this information gap.

2. Safeguards: Technical and policy safeguards need to be developed before they are needed, because AI capabilities advance faster than safety infrastructure. CG funds safety research and policy frameworks that establish norms before crises occur.

3. Capacity: The AI safety field lacks the talent base to match the pace of AI development. CG funds career pipelines, training programs, and institutional capacity to build the field.

The broader organizational theory of change is "hits-based giving" -- take many high-risk bets, expect most to fail, and let a few successes carry the portfolio. This is embedded within a "worldview diversification" framework that allocates capital across different ethical frameworks.

Revealed Theory of Change

CG's actions largely match its stated theory, with some important divergences:

What matches: CG actually spends on all three pillars. It funds evaluations (Epoch AI, METR), technical safety research ($40M+ RFP), governance policy (CSET, RAND, bipartisan legislation), and capacity building (MATS, BlueDot, career transition funding). The scope and specificity of the technical AI safety RFP demonstrates genuine domain expertise.

What diverges:

1. CG's revealed priority is pragmatic safety within continued development, not all possible approaches to AI risk reduction. CG funds groups that "disagree with each other" on technical questions but appears to draw a line at organizations whose theory of change involves stopping or significantly slowing frontier AI development. PauseAI, ControlAI, and MIRI's recent pause advocacy do not appear to be funded. This is consistent with Anthropic's interests (continued development with safeguards) but may not represent the full space of legitimate safety approaches.

2. CG is building itself as the central hub for AI safety philanthropy, not just as one funder among many. The rebrand to CG, the 13-fund structure, the partnership team, the custom donor portfolios -- all of these position CG as the default platform for large-scale AI safety giving. The "Need More Funders" post calls for diversity but channels new funders through CG's advisory framework.

3. CG's governance investments have been more impactful than its technical safety investments, relative to spending. The RSP framework, funded partly through CG grants to METR and Carnegie Endowment, is now used by all major AI labs. This is a governance hit that has shaped industry norms far beyond its grant dollar amount. Meanwhile, the technical safety portfolio's concrete achievements (benchmarks, individual research papers) are important but have not yet produced a decisive safety technique.

4. CG's self-identified mistake -- being too slow on technical AI safety -- reveals something about its internal dynamics. CG admitted in its 2024 annual letter that it should have scaled technical safety spending earlier and is "now playing catch-up." In a monopsony, the funder's timing errors are the entire field's timing errors.

Key Assumptions

Assumption 1: Philanthropic funding can meaningfully influence AI safety outcomes at scale.

• Evidence for: CG-funded benchmarks are used by governments and labs. RSP framework now industry standard. MATS alumni at major labs and government.
• Evidence against: AI safety philanthropy (~$500M/yr) is 1/400th of AI development spending ($200B+/yr capex). Labs' commercial incentives dwarf philanthropic influence. Safety progress depends on lab cooperation that philanthropy cannot compel.
• Testable: Compare safety outcomes in areas with CG funding vs. without. Track whether CG-funded policy frameworks actually constrain lab behavior during competitive pressure.
• If wrong: The entire AI safety philanthropic enterprise is a feel-good exercise. The real levers are lab governance, regulation, and international agreements -- none of which philanthropy can directly deliver.

Assumption 2: CG's worldview (AI development should continue with safeguards) is the correct framework for reducing AI risk.

• Evidence for: Stopping AI development is practically impossible given geopolitical competition. Safeguards-within-development is the only approach with a chance of being implemented by actual AI labs.
• Evidence against: If development speed is the primary risk factor, then funding that makes development seem safer (by providing safety frameworks) could net-increase risk by reducing pressure to slow down. The "race to the bottom" dynamic may overwhelm any safeguards.
• Testable: Track whether RSP commitments are honored or abandoned under competitive pressure.
• If wrong: CG has spent $580M+ making the development of transformative AI appear more responsible while not materially reducing the probability of catastrophe.

Assumption 3: CG can remain an independent evaluator of AI safety despite deep structural ties to Anthropic.

• Evidence for: CG's COI policy exists. CG funds diverse organizations, including some critical of Anthropic. CG publicly acknowledged the concentration problem.
• Evidence against: Moskovitz is an Anthropic investor. Karnofsky (501(c)(3) board) works at Anthropic. Multiple staff moved to Anthropic. Anthropic wealth could reshape the field. PauseAI accusation. CG does not fund organizations opposed to continued AI development.
• Testable: Would CG fund an organization that directly challenged Anthropic's RSP as insufficient or that called for pausing Anthropic's frontier development?
• If wrong: CG is effectively Anthropic's philanthropy arm, not an independent funder of AI safety.

Assumption 4: Concentrated funding through a single sophisticated funder is better than distributed funding through many less sophisticated funders.

• Evidence for: CG's hits-based approach has produced verifiable successes (Baker Nobel Prize, YIMBY movement, cage-free campaigns). CG's technical sophistication in the AI safety RFP exceeds what most funders could produce. Speed and decisiveness matter when timelines are short.
• Evidence against: Monopsony dynamics distort the field's strategy. Correlated blind spots are undetectable by the blind spot holder. FTX collapse showed the danger of funder concentration. The "discovery problem" means approaches that don't fit CG's worldview are structurally unfindable.
• Testable: Compare the AI safety ecosystem's strategic diversity now vs. in a counterfactual with 5-10 funders of similar size.
• If wrong: CG's $580M+ created a large but strategically narrow field, missing approaches that could have been more effective.

Strengths

1. Unmatched scale and institutional knowledge in AI safety philanthropy. No other funder combines CG's budget, technical sophistication, and decade-long track record. The $40M+ technical AI safety RFP demonstrates world-class understanding of the research landscape.

2. Genuine intellectual honesty, unusual among major funders. CG publishes annual reviews, self-critiques, notable lessons pages. It funded a contest for essays challenging its AI risk views and gave the top prize to research contradicting its thinking. It publicly admitted being too slow on technical safety.

3. Concrete, verifiable achievements across multiple cause areas. The YIMBY movement, cage-free campaigns, malaria vaccine, and David Baker's Nobel Prize are real hits. The RSP framework is now industry standard. Epoch AI's data is widely cited.

4. Strategic positioning for the coming Anthropic wealth wave. If $37.8B in Anthropic co-founder pledges materializes and flows through EA-aligned channels, CG is the organization best positioned to direct that capital. This could increase CG's influence by an order of magnitude.

5. Bipartisan approach to AI safety. In a highly polarized political environment, CG's explicit commitment to bipartisan AI safety engagement is strategically valuable and rare.

Weaknesses and Risks

1. Structural Anthropic conflict is the elephant in the room. CG's primary funder (Moskovitz) is an Anthropic investor. CG's co-founder (Karnofsky) works at Anthropic. Multiple CG staff moved to Anthropic. The organizations CG funds evaluate Anthropic's safety. This is not one conflict -- it is a web of entanglement that no COI policy can fully address.

2. Monopsony power distorts the field. With 50-80% of AI safety philanthropy, CG's preferences become the field's strategy. Organizations optimize for CG's worldview. Approaches CG does not fund (pause advocacy, radical regulatory approaches) are structurally marginalized. The "discovery problem" means promising ideas that don't match CG's framework never get tested.

3. Only ~3 grant investigators manage $130M+/year in AI safety grants. This is an extraordinary concentration of decision-making power. Each investigator controls ~$40M+ in annual funding decisions. If they have correlated blind spots, the field has correlated blind spots.

4. No independent evaluation of CG's own effectiveness. CG evaluates its grantees but no one evaluates CG. There is no external audit of whether $580M+ in AI safety grants has been well spent. The hits-based approach makes evaluation harder, but the absence of even an attempt is a governance failure.

5. Board lacks independent AI safety expertise. All boards are dominated by insiders (Tuna, Moskovitz, Berger) with only 2 clearly external members. There is no independent AI safety expert providing board-level oversight of $130M+/year in AI safety spending.

6. The approaching Anthropic wealth could make everything worse. If $37.8B in Anthropic employee donations flows through CG-adjacent channels, the conflict of interest becomes structural. Organizations funded by Anthropic employees evaluate Anthropic's safety. The independence of the entire AI safety ecosystem could be compromised.

Cross-References

• Anthropic: The deepest entanglement in CG's network. Personnel pipeline, Moskovitz investment, Karnofsky marriage/employment. CG and Anthropic share a social and intellectual ecosystem. CG's theory of change (safeguards within continued development) aligns with Anthropic's core business model.

• METR: CG-funded evaluations org. METR has stronger COI policies than CG itself -- it does not accept money from AI companies, executives, or "people we work directly with on major risk management decisions." METR's independence standards are higher than CG's.

• SFF (Survival and Flourishing Fund): The second-largest AI safety funder (Jaan Tallinn). Together with CG, accounts for ~80% of philanthropic AI safety funding. SFF provides some diversification but shares many of CG's worldview assumptions.

• Longview Philanthropy, AISTOF: Independent AI safety donor advisories that CG explicitly supports as healthy diversification. CG was "glad that AISTOF existed before and was doing some independent funding."

• PauseAI, ControlAI: Organizations whose theories of change involve slowing or stopping frontier AI development. CG does not appear to fund these, which marks a boundary of its pluralism. This is the most direct test of whether CG's funding is independent of Anthropic's interests.

• GiveWell: CG's origin -- spun off from GiveWell in 2017. They share board members (Hassenfeld) and office space. GiveWell handles the "straightforward charity" allocation (~10% of CG's total portfolio).

What Would Change This Assessment

• CG funding PauseAI, ControlAI, or similar organizations would significantly weaken the claim that CG's AI safety grantmaking is structurally aligned with Anthropic's interests.
• An independent evaluation of CG's AI safety portfolio finding either strong positive impact or significant gaps would update the effectiveness assessment dramatically.
• Moskovitz divesting from Anthropic and Karnofsky leaving CG's 501(c)(3) board would reduce the structural conflict.
• A major AI safety failure at an Anthropic-evaluated org that CG funded would test whether the evaluator independence concern is real.
• CG publishing grant rejection data would directly test the monopsony/discovery problem critique.
• The Anthropic IPO and subsequent employee giving patterns will reveal whether CG captures this wealth and whether it creates untenable conflicts.

Self-Critique

What sources should I have checked but didn't: Holly Elmore's Substack (primary source for the PauseAI accusation), Good Ventures Form 990-PF (actual grant dollars), EA Forum comment threads (where the most pointed community criticisms live), the full Vipul Naik donation database (27K words, only partially analyzed), and CG's internal team composition in more detail.

Where is this analysis potentially biased: I may have given too much weight to the Anthropic conflict narrative. CG funds hundreds of organizations, and the Anthropic connection -- while real -- may be less causally important than the structural monopsony dynamic. I may also underweight CG's genuine intellectual honesty and pluralism.

What would a thoughtful person who disagrees say: "CG is more transparent and self-critical than any comparable organization. The Anthropic conflict is managed through recusal and disclosure. CG funds groups that disagree with each other. The alternative to CG dominance is not a vibrant marketplace of funders -- it is less total funding for AI safety. The perfect is the enemy of the good."

Single weakest claim: That CG's grantmaking is structurally aligned with Anthropic's interests. While the structural conditions for this alignment exist (Moskovitz investment, Karnofsky employment, revolving door), I have not established that specific grant decisions have been influenced. The PauseAI accusation is secondhand and unverified. CG may simply not fund pause advocacy because it genuinely believes pause is bad strategy, not because it serves Anthropic.

What information would most change my view: CG publishing a detailed breakdown of grant rejections with reasons, or an independent evaluation of CG's AI safety portfolio, or evidence that CG has funded organizations fundamentally opposed to continued frontier AI development.