Giving What We Can

Stated Theory of Change

GWWC's path to AI risk reduction runs through a multi-step causal chain:

1. Convince people to publicly pledge 10% of income to effective charities
2. Pledgers use GWWC's research to identify high-impact giving opportunities
3. Some fraction of pledgers direct donations to AI safety organizations
4. Those organizations use the funding to do work that reduces AI risk

GWWC itself does not claim to be an AI safety organization. Their direct theory of change is about creating a cultural norm of effective giving. AI safety is one of several cause areas that beneficiaries of this norm may choose to fund. GWWC's AI cause area page recommends CHAI, GovAI, and the Long-Term Future Fund.

Revealed Theory of Change

GWWC's actions are well-aligned with their stated mission of norm change through pledging. The revealed priorities -- based on where they invest time and resources -- are:

1. Pledge growth as the primary metric. The BHAG (1M pledgers), the 2026 targets (1,500 new pledges), and the hiring profile (Head of Marketing, Partnerships Associate) all optimize for pledge count. This is the KPI the organization lives and dies by.

2. Platform as infrastructure. The 2026 strategy explicitly makes the donation platform "more central." This positions GWWC as the effective giving rails -- a neutral conduit that benefits from volume regardless of cause area allocation.

3. Evaluating evaluators as the research edge. This is where GWWC adds unique intellectual value. The work with ACE shows concrete impact on ecosystem quality. But research is being deliberately constrained (they explicitly won't expand scope) in favor of growth.

4. EA-adjacent but broadening. Pledge booths at EAG events, but expanding to Sentient Futures Summit and influencer partnerships. The tension between EA-native growth (easier, more concentrated) and mainstream expansion (harder, more scalable) is the central strategic choice.

Where revealed theory diverges from AI safety relevance: GWWC devotes minimal visible attention to AI safety as a cause area. The AI page recommends CHAI (no longer a top community recommendation) and GovAI. No staff member appears to specialize in AI safety research. The $6M to animal welfare in 2024 is highlighted; no equivalent AI safety figure is disclosed. This suggests AI safety is a low priority within GWWC's multi-cause-area approach.

Key Assumptions

Assumption 1: Public pledging changes behavior at scale.

• Evidence for: 10,000+ pledgers, $355M in recorded donations, behavioral science literature on pre-commitment.
• Evidence against: 70% of pledgers stop recording after 5 years. Half of impact from 2.75% of pledgers. The mechanism may work for the committed minority and fail for the majority.
• Testable: Yes, through cohort retention analysis and randomized pledge vs. no-pledge studies.
• If wrong: The pledge is a selection effect (committed people were going to give anyway) rather than a treatment effect (the pledge caused them to give).

Assumption 2: Cultural norm change is achievable at the target scale.

• Evidence for: Religious tithing norms (Zakat, Ma'aser) show that 2.5-10% giving norms can persist across billions of people for centuries. Sunstein's social tipping point theory provides a mechanism.
• Evidence against: Current growth is concentrated in EA-adjacent populations. Expanding beyond this base is unproven. The 10% target may be too high for mainstream adoption (One for the World's 1% threshold suggests lower entry points may reach more people).
• Testable: Partially -- expansion to non-EA events will test mainstream receptivity.
• If wrong: GWWC becomes a niche EA institution rather than a mainstream cultural force.

Assumption 3: Effective giving leads to effective AI safety work.

• Evidence for: Pledger donations flow to Long-Term Future Fund, Emerging Challenges Fund, and directly to AI safety organizations through the platform.
• Evidence against: No public data on AI safety donation fraction. GWWC's AI recommendations (CHAI, GovAI) are outdated. Pledgers are worldview-diverse and may systematically prefer global health or animal welfare over AI safety.
• Testable: Yes, if GWWC published cause area donation breakdowns.
• If wrong: GWWC's impact on AI safety funding is negligible even as total donations grow.

Assumption 4: The marginal dollar to GWWC produces more than $1 in effective donations.

• Evidence for: GWWC's self-reported 6x average multiplier.
• Evidence against: CEARCH's 1.06x marginal multiplier estimate. The gap between average and marginal suggests diminishing returns.
• Testable: Yes, through independent evaluation of marginal cost-effectiveness.
• If wrong: Donating to GWWC is equivalent to donating directly to recommended charities, eliminating the case for operational funding growth.

Strengths

1. Unusual institutional honesty. The acknowledged mistakes page, the transparent (if declining) multiplier reporting, and the "what we won't do" list are rare among nonprofits. This builds long-term credibility.

2. Evaluating evaluators is a genuine innovation. No other organization fills this niche. The concrete impact on ACE's methodology demonstrates systemic value that transcends money moved.

3. No-fee donation platform eliminates perverse incentives. By decoupling revenue from donation volume, GWWC avoids the structural conflict that plagues many donation platforms.

4. High personal commitment. CEO donating 50% of income, nearly all staff pledged, several staff doing salary sacrifice. This level of skin-in-the-game is exceptional.

5. Clean organizational trajectory. Smooth leadership transition, successful spinout from EV, disciplined strategic focus. No governance scandals despite being deeply embedded in EA during the FTX era.

6. Multiplier potential at scale. If the 6x multiplier holds as GWWC grows, the BHAG would redirect $3B annually to effective charities for ~$50M in operational cost. Even a 2x multiplier at that scale would be significant.

Weaknesses and Risks

1. Pledge attrition is the existential threat. 70% silent after 5 years. If most non-recording pledgers have stopped giving effectively, the pledge is less a lifelong commitment and more a 2-3 year giving phase for most people. This would fundamentally undermine the BHAG's math.

2. Extreme donation concentration. Half of impact from 2.75% of pledgers. The organization's measured effectiveness depends on a handful of wealthy individuals. If they leave or redirect, impact metrics collapse.

3. Coefficient Giving dependency. 50% of operational funding from one source. If Open Phil deprioritizes effective giving support (as it has deprioritized other areas), GWWC faces a funding crisis. The tipping model is nascent and unproven at the needed scale.

4. Marginal multiplier near 1x. If CEARCH's estimate is correct, additional operational funding produces roughly $1 in effective donations per $1 spent. This undermines the growth case -- why fund GWWC's expansion rather than donate directly?

5. AI safety contribution is unmeasured and likely small. Without cause area donation data, the AI safety impact is speculative. The outdated AI recommendations suggest low organizational attention to this cause area.

6. EA brand risk. GWWC's founding story, community, and growth channels are deeply intertwined with EA. FTX demonstrated this vulnerability. Mainstream expansion requires distancing from EA without losing the EA base.

Cross-References

• GiveWell: GWWC evaluates and recommends GiveWell. Most GWWC global health donations likely flow through GiveWell recommendations. Complementary relationship.
• Founders Pledge: GWWC evaluates Founders Pledge. CEO sits on a Founders Pledge committee (disclosed conflict). Different target audience (tech founders vs. general public).
• One for the World: Direct comparison -- 1% pledge targeting university students. Lower threshold, different demographic. Tests whether lower entry points drive more total impact.
• The Life You Can Save: Similar mission, broader approach. GWWC's 10% is more demanding but potentially more transformative per person.
• Coefficient Giving / Open Philanthropy: Primary funder. CG funds both GWWC and many organizations GWWC evaluates, creating structural conflicts.
• EA Funds, Longview: GWWC evaluates and recommends their funds. Channel for donations including some AI safety funding.

What Would Change This Assessment

• Published AI safety donation data. If GWWC showed that, say, 15%+ of donations flow to AI safety causes, the indirect theory of change would gain substantial credibility.
• Cohort retention data showing recent improvement. If 2022+ cohorts retain at 50%+ instead of 30%, the pledge model's scalability looks much better.
• Updated AI cause area recommendations. Replacing CHAI/GovAI with current top AI safety organizations would signal genuine attention to this cause area.
• Marginal multiplier re-evaluation. An independent evaluation showing marginal multiplier above 3x would restore the case for operational funding growth.
• Mainstream growth evidence. If non-EA channels (influencer partnerships, broader events) drive pledges at comparable rates to EA events, the BHAG becomes more credible.
• Financial independence. If the tipping model or other earned revenue reduces CG dependency below 30%, the biggest structural vulnerability is mitigated.

Self-Critique

Limitations of this analysis:

• UK financial data is unavailable. The operational budget estimate (~$2M/year) is derived from impact evaluation self-reports, not audited financials.
• The AI safety assessment relies heavily on absence of evidence (no reported AI safety donation fraction) rather than evidence of absence. GWWC could be channeling meaningful AI safety donations without publicizing it.
• I have not accessed the full Google Docs impact evaluation, which may contain more granular retention and cause area data than the summary pages.
• The CEARCH marginal multiplier estimate could not be fully verified -- I found the number in search results but did not read the full methodology paper.

Where this analysis could be wrong:

• I may be underweighting GWWC's long-term cultural norm change value. If GWWC succeeds in normalizing 10% giving even for 100,000 people, the downstream effects on all cause areas (including AI safety) could be enormous over decades -- even if the current AI safety donation fraction is small.
• The 30% retention rate may look worse than reality if many pledgers are donating effectively but not recording with GWWC. The 2026 pledge fulfillment initiative may surface higher actual compliance.

Weakest claim: My assessment that GWWC's AI safety impact is "likely small" is based on absence of data rather than direct evidence. If GWWC published cause area data showing significant AI safety flows, this claim would need major revision.

What would a thoughtful disagreer say? "GWWC is building the infrastructure of effective giving. In a world where AI safety becomes the consensus top cause area, having millions of pledgers who are committed to giving effectively -- wherever the evidence points -- is exactly what the field needs. You're evaluating GWWC on its current AI safety contribution rather than its option value for the future."