BRENNEN

Industries / Government, education and community / Nonprofits

Foundations

Foundations are organizations that hold a pool of money, usually given by one person, family or company, invest it, and give part of it away each year as grants to charities, researchers and other causes.

1867

George Peabody gave $1 million to set up the Peabody Education Fund for "the promotion and encouragement of intellectual, moral, or industrial education" among the young of the poorer Southern and Southwestern states; the Rockefeller Archive Center calls it what many consider to be the first modern American philanthropic foundation.

resource.rockarch.orgmsa.maryland.gov

30

Milestones achieved across 6 domains.

8

To watch for: announced, not yet confirmed outside.

12

Open deliverables. Nobody has shipped them.

In one minute

Where it started

Endowments that pay for good works are old: in the 1500s Ottoman jurists argued over cash waqfs, endowments of money lent out at interest whose income paid teachers and preachers. The modern foundation, a fund run by trustees who choose whom to give to, took shape in the United States with George Peabody's education fund of 1867, which at first left out Black children because it supported only existing schools, and with the Cleveland Foundation, which in 1914 pooled many donors' gifts under one board. Since 1976 US law has required private foundations to pay out about 5% of their assets each year.

Where it stands

A 2018 study led by Harvard's Hauser Institute counted more than 260,000 foundations holding close to US$1.5 trillion; 90% had less than US$10 million, and 58% did not work with any other foundation. US foundations gave an estimated $117.15 billion in 2025, 19% of all US giving, by Giving USA's count, and Europe's 175,203 public-benefit foundations spend about €76 billion a year, by Philea's. The largest are changing pace: the Gates Foundation plans to spend $200 billion and close in 2045, and the new OpenAI Foundation has committed $25 billion.

What is open

Much of what foundations could learn about themselves sits unread in tax returns and PDF reports: what they own, whether they kept public pledges to give more and with fewer conditions, and what their evaluations found. AI models that read long documents, web agents that check websites, and US tax returns now filed as machine-readable data put these jobs within reach of one person or a small team. Other open jobs are fair tests of AI as funders begin to use it, such as whether an AI screener scores the same application differently when only the applicant's name changes; only 1% of foundations Candid surveyed in 2025 used AI to screen applicants.

Domains

AchievedWatch forOpen

Advances to build with

1 research note

Open in foundations

3 of 12 open deliverables

Open · The endowment

What foundations own

A public dataset of the stocks, bonds and other investments that the 1,000 largest US private foundations list with their tax returns, for each year, with every holding matched to a standard security identifier and an industry. A second table flags holdings that clash with a foundation's own stated mission, such as tobacco shares held by a health foundation, for a person to review.

Built with what exists

The IRS's bulk e-file data for Form 990-PF and the NCCS tables; OpenFIGI, a free service that matches security names to standard identifiers; a small AI model for reading attached schedules, such as Meta's Muse Ember, built for extraction and long-document parsing (see the Note "The jobs agents repeat"); Python.

Credit · unclaimed

Open · Grantmaking

Looser terms, checked

A measure, foundation by foundation, of the share of grant dollars given as general support in each year from 2018 to 2024, taken from the purpose line of each grant on US tax returns, comparing foundations that signed the 2020 Council on Foundations pledge with those that did not.

Built with what exists

The IRS e-file data for Form 990-PF, which lists each grant's recipient, amount and purpose, through the NCCS tables or ProPublica; the published list of pledge signers; a classifier such as Gemini Decision 1, which returns probabilities for fixed-answer questions (see the Note "A thousand questions, one pass"); Python.

Credit · unclaimed

Open · Reviewing applications

Name swaps for AI screeners

A public test of whether AI models used to screen grant applications score the same application differently when only the applicant's name, organization or location changes. Write 200 fictional applications to a fictional funder, make matched copies that differ in one detail, score every copy with several models under one published rubric, and report the gaps.

Built with what exists

AI models such as Claude, Gemini and GPT through their programming interfaces; Gemini Decision 1, which returns probabilities for fixed-answer questions and so can show small shifts that a final score hides (see the Note "A thousand questions, one pass"); Python for paired statistics.

Credit · unclaimed

Where it sits