Hoverview

AI philanthropy · data quality · donor reporting · funding · results frameworks

The wave is real. The plumbing is not

The organisational readiness question has been asked. The harder question is whether anyone can measure what the money does.

Run the new AI philanthropic wealth through a 5 per cent distribution rate and three pools alone would move roughly $17 billion a year. The OpenAI Foundation holds a 26 per cent stake in OpenAI Group PBC; at the $852 billion valuation set by OpenAI's August 2026 tender offer, that stake is worth about $220 billion, which at 5 per cent is $11 billion a year. Anthropic's eight co-founders pledged 80 per cent of their wealth in January 2026; after the company raised at a $965 billion valuation, Forbes estimated each co-founder's net worth at $16.6 billion, putting the pledged total at roughly $106 billion, or $5.3 billion a year at 5 per cent. The Jen-Hsun and Lori Huang Foundation genuinely does owe 5 per cent; Nvidia's rise pushed its assets up 170 per cent to $9.2 billion in 2024, putting its annual obligation around $460 million. That is nearly twice the Gates Foundation's record $9 billion annual payout. The money is not hypothetical. Coefficient Giving raised its 2026 allocation to GiveWell's recommendations from $175 million to $1 billion in a single announcement in July, and the OpenAI Foundation is hiring the people who will write the eligibility rules for the next rounds.

The organisational readiness question has already been answered. Can a $3 million health NGO in Kisumu absorb $300,000 of unrestricted AI philanthropy money? Yes, and the evidence is the OpenAI Foundation's People-First AI Fund, which is sized to organisations with budgets between $500,000 and $10 million, caps grants at 10 per cent of the organisation's budget, and excludes university-affiliated institutes and think tanks. The mechanism exists. The eligibility line that restricts it to US-based 501(c)(3) organisations is a policy choice, not a structural constraint.

That argument is sound, and this post does not contest it. The question it raises and does not pursue is harder: once the money moves, can anyone say what it did?

What the new funders are not asking for

The striking feature of the People-First AI Fund is what it does not require. Grants are unrestricted. There is no logframe, no results framework, no theory of change, no indicator table. The application asks for a narrative about how the organisation would explore AI in service of its mission, and the fund is explicitly platform-agnostic: use of specific AI tools or providers is not required and will not affect funding decisions.

For a $50 million programme reaching 208 community groups at roughly $195,000 each, that is a defensible design. The grants are small enough that the reporting burden of a full results framework would exceed the cost of the grant. Unrestricted funding is a deliberate choice to trust the organisation, and the evidence from the aid effectiveness literature supports the case that small unrestricted grants produce better outcomes than small restricted ones because the organisation can allocate to its actual needs rather than to a funder's predetermined categories.

But the People-First AI Fund is $50 million against a potential $17 billion. What happens when the amounts grow and the funders need to know whether the money worked?

Traditional donors built their results infrastructure over decades. The UK's logframe guidance dates from 2011. DG ECHO's Single Form has been iterated since the early 2000s. The World Bank's results framework is embedded in its legal agreements. These systems exist because at some point a parliament, a board, or a taxpayer asked: what did the money achieve? The question is inevitable, and the reporting infrastructure that answers it shapes everything downstream: which indicators are collected, how beneficiaries are counted, what "success" means in a completion report.

The new AI funders have not yet built that infrastructure, because they have not yet needed it. The question is whether they will build something that learns from the existing system's failures, or whether they will reinvent it with the same structural problems.

What the existing system cannot tell you

The sector's current results infrastructure has been described at length in this blog, and the summary is not encouraging.

Most humanitarian reporting systems do not record whether a beneficiary was previously counted, which means every cumulative reach figure is wrong by an unknown factor. DFID's own methodology retreated to "peak year" as its preferred indicator precisely because cumulative totals could not be deduplicated. OCHA's 2020 guidance distinguishes between "people reached" (at least one contact) and "people covered" (planned assistance delivered in full), and neither definition resolves whether the same person appears twice in the annual total.

The logframe method, as Rosenberg and Posner designed it in 1979 and as Gasper critiqued it in 2000, conflates sufficient and necessary conditions. A bare causal link between two levels cannot say whether delivering the output is enough for the outcome to follow, merely necessary for it, or one of several contributing causes. The method was designed for a world of single-donor, single-programme accountability, and it strains when a programme reports to multiple funders with different frameworks, which is the shape of the world the AI philanthropy wave will create.

The Council on Foundations and Candid found that about 13 per cent of international grant funding reached organisations registered in the country where they implement. The implication for results reporting is that 87 per cent of the grant agreements, and therefore 87 per cent of the reporting frameworks, are written by intermediaries in London, New York, or Washington, and the implementing organisation reports upward through a chain that adds compliance cost at every link without adding measurement quality.

These are not problems that $17 billion will solve by volume. They are structural features of the reporting infrastructure, and they will reproduce unless the new funders design something different.

What Coefficient's approach reveals

Coefficient Giving's $1 billion allocation to GiveWell is instructive because it represents a different theory of accountability. The money flows through four implementing organisations (Against Malaria Foundation, New Incentives, Evidence Action, and others) that GiveWell has evaluated using cost-effectiveness analysis. The accountability mechanism is not a logframe. It is a published estimate of cost per life saved, updated as evidence accumulates, with the methodology visible to anyone who reads the website.

That model has genuine strengths. The cost-effectiveness estimates are falsifiable. The methodology is published. The bar for funding (lowered from 2,000 times to 1,000 times the benchmark this year to accelerate spending) is explicit and debatable. And the implementing organisations are already delivering at scale, which means the absorptive capacity problem that worries traditional donors is partially solved.

It also has a limitation that matters for the question this post is asking. The cost-effectiveness model works for interventions with established evidence bases: bed nets, seasonal malaria chemoprevention, vitamin A supplementation. These are programmes where the relationship between input and outcome has been measured across dozens of randomised controlled trials. The model does not extend easily to the kinds of work the People-First AI Fund supports: community journalism, cultural organisations, disability services, rural community groups exploring how AI might help them serve their communities. For those programmes, nobody has an established evidence base, because the interventions are new.

When Coefficient's $1 billion moves through bed net distributions in the Democratic Republic of the Congo, the reporting infrastructure exists: the Against Malaria Foundation tracks nets distributed, households covered, and malaria incidence in target areas. When the OpenAI Foundation's $50 million moves through 208 community groups across the United States, the reporting infrastructure does not exist, because the fund was designed to let organisations explore rather than to measure what exploration achieves.

Both are legitimate designs for their scale and purpose. The question is what happens at the scale in between: when the grants are large enough to require accountability but the interventions are new enough that no established framework exists.

The design window

The most important thing about the new funders is that they have not yet learned the sector's habits. The OpenAI Foundation excluded think tanks, which no traditional donor would do. Coefficient publishes its reasoning and revises its allocations in public. These are organisations making it up as they go, which means the results infrastructure they build in the next two to three years will shape how several billion dollars a year is measured for decades afterwards.

That is a design window, and it is open now.

The sector has an opportunity to influence what gets built. It also has a track record of offering the new funders exactly what it offered the old ones: consortium proposals, prime-and-sub structures, and logframes that look like the logframes written for the funder that stopped writing cheques last year. The choice is a larger share of the same pipe, or a different pipe.

The same-pipe option means traditional results frameworks adapted to the new funders' language, with the same structural problems: undifferentiated reach figures, inherent-level models that break across donors, indicator caps that shape programme logic, and governance rules discovered after the framework is committed. The different-pipe option means results infrastructure that is designed from the start for the characteristics of AI philanthropy money: multiple funders per programme, new and untested interventions, fast iteration, organisations that have never filled in a logframe and may be better served by never filling one in.

What the different pipe would look like is not obvious, and proposing one is beyond the scope of this post. But three requirements are clear from the failures of the existing system. The infrastructure needs to distinguish new from repeat beneficiaries at the point of data entry, not at the point of aggregation. It needs to separate the programme's causal logic from any single donor's rendering of it, so that the same programme can report honestly to funders with different frameworks without maintaining parallel planning documents. And it needs to make the measurement layer (indicators, targets, means of verification) separable from the logic layer (what the programme does and why it expects it to work), so that a programme exploring a new intervention can describe its logic before it has the evidence to populate a full indicator table.

What "ready" means

The question "are you ready?" has at least two answers.

The organisational answer is yes for the reasons he gives. Small organisations can absorb proportional grants. Regranting mechanisms exist. The eligibility constraints are policy choices that can be changed by the people being hired this quarter.

The measurement answer is not yet, and the reason is that the sector's existing results infrastructure was built for a funding landscape that is about to change in ways it was not designed for. A system built for bilateral accountability to a single donor does not naturally extend to a world where a programme receives unrestricted AI philanthropy alongside a restricted FCDO grant alongside a GiveWell-evaluated intervention, each with different expectations about what "results" means and how they should be reported.

Whether the new funders will ask the results question at all is uncertain. They may decide that the cost of measurement exceeds its value at the grant sizes they intend, and that trust-based philanthropy does not require the apparatus of a logframe. That is a coherent position, and for grants under $500,000 it may be the right one.

But someone will ask. A board member, a journalist, a congressional staffer, or the funders themselves, when the second or third year's allocation is under discussion and the question is whether the first year's money achieved anything. When that question arrives, the infrastructure to answer it will either exist or it will not, and building it after the question is asked is how the sector ended up with the system it has now.

Sources

Valuations and pledges

AI philanthropy programmes cited

  • Update on the OpenAI Foundation, OpenAI Foundation (2026). Source for the $1 billion planned for 2026 and the 0.45 per cent payout rate
  • 2026 People-First AI Fund, OpenAI Foundation (2026). Source for the eligibility criteria, the $500,000 to $10 million budget range, the 10 per cent cap, the exclusion of think tanks, and the expansion to regranting organisations
  • Update on the People-First AI Fund, OpenAI Foundation (2025). Source for 208 grantees and $40.5 million in the first wave
  • Come build the OpenAI Foundation, OpenAI Foundation (August 20, 2026). Source for the hiring announcement and the "much of the organization still left to build" statement

Coefficient Giving and GiveWell

Results infrastructure