The Billionaire Building AI’s Rulebook

AI regulation is usually sold as protection against dramatic technical dangers: cyberattacks, biological weapons, or models that escape human control. But regulation also creates standards, evaluators and institutions with authority to decide whether those standards have been met. When the product is a system that answers questions, the stakes extend to what it is allowed to say.

Governments borrow expertise. Researchers identify risks, policy organizations turn them into proposals, fellows enter public institutions, and journalists establish which concerns deserve attention. Much of the architecture is built before legislators vote. Dustin Moskovitz has helped finance that architecture at extraordinary scale.

I. The investor and the regulator

The Facebook co-founder participated in Anthropic’s $124 million Series A in 2021. That was the total round, not his individual investment. Separately, Good Ventures, founded by Moskovitz and Cari Tuna, funds philanthropy through recommendations from Coefficient Giving, formerly Open Philanthropy.

Coefficient says its AI safety, security and field-building commitments grew from $168 million in 2024 to $351 million in 2025, with more than $1 billion projected for 2026. Those are funding commitments and recommendations, not a tally of cash spent lobbying; Coefficient also works with other donors.

The grant portfolio includes Redwood Research and other AI research organizations. Horizon lists Coefficient among its donors and places fellows in government. Tarbell disclosed Coefficient as its majority funder as of 2025; its policy says donors have no editorial control. The influence question is larger than direct instructions: who can afford to build the field from which public officials recruit their experts?

Other AI-safety advocates pursue legislation directly. Rep. Ted Lieu named ControlAI among supporters of his AI Kill Switch Act proposal. That is a distinct advocacy relationship; the records cited here do not establish a Coefficient-to-ControlAI funding link.

Anthropic also sought to shape California’s rules. SB 1047 proposed safeguards for very large models, but Governor Gavin Newsom vetoed it in 2024. SB 53 became law in 2025. It requires large frontier developers to publish risk frameworks and includes incident reporting and whistleblower protections. Its large-developer duties use a revenue threshold above $500 million; the law does not impose identical obligations on every startup.

Anthropic endorsed SB 53, saying it and other leading companies already followed many of the relevant practices. That is the economic question worth pressing. A company that has built its compliance machinery may face less additional expense than a competitor that must build it from scratch. Regulation can restrain incumbents while also making entry more expensive.

This is an argument about competitive incentives, not proof of a measured financial windfall. The law also burdens Anthropic, and Moskovitz’s 2021 investment alone does not establish his current personal exposure. But the alignment deserves scrutiny: an investor in a frontier laboratory helped fund the policy field, while that laboratory advocated rules resembling practices it already used.

II. When safety becomes political

A rule against helping someone engineer a pathogen can point to a concrete harmful act. Labels such as misinformation, extremism or political bias first require someone to define them. Turning that definition into a benchmark makes it scalable. It does not remove the political judgment.

The Anti-Defamation League’s AI Index evaluates models on categories including antisemitic, anti-Zionist and extremist outputs. Its published recommendations call for audits and outside engagement, including safeguards concerning Israel and Palestine. Jonathan Greenblatt has claimed ADL’s engagement led to changes in Sora’s refusal behavior. That is his account of the effect, not an independent audit of the model.

ADL is a separate example of pressure on model behavior, not a documented recipient of Moskovitz funding here. Nor does SB 53 enact ADL’s political categories. The concern is how a private benchmark can gain reputational authority, then become a standard companies feel compelled to satisfy—and potentially influence future procurement or regulation.

What counts as anti-Zionism, how criticism of Israel is classified, and which arguments a model should refuse are political questions. Users should be able to examine those choices. An assistant can accept one premise and challenge another, choose which history matters, or attach a warning to one position while treating its rival as ordinary knowledge. Most users see the answer, not the evaluation that shaped it.

III. The politics of philanthropic field-building

The pattern extends beyond AI. Good Ventures committed $60 million toward a $120 million Abundance & Growth Fund. The fund covers several areas, including housing and economic growth. Housing deregulation can promise more homes while shifting decisions away from residents who prefer existing local restrictions. Calling a policy “abundance” does not settle whose preferences should prevail.

California’s official Proposition 12 donor disclosures list $4 million from Open Philanthropy Action Fund supporting the animal-confinement measure. Philanthropy that finances changes to how a market operates is exercising political influence, even when the goal is presented as reducing suffering.

Coefficient’s expansion in biosecurity raises another version of the same question: who funds the research and professional networks from which governments draw advice? Emergency powers, public-health mandates and acceptable risk are matters for public argument, not questions settled by the size of a research budget.

AI makes this influence different in kind. The same assistant may explain whether a housing reform worked, summarize a pandemic dispute, or decide which sources deserve credibility. It is becoming an interface through which people understand the other systems philanthropy seeks to change.

IV. The system being built

Government has enormous formal authority. Private institutions can nevertheless shape the agenda long before that authority is exercised: by financing expertise, defining risks and supplying the people available to implement policy.

The question is not whether AI should have rules. It is who builds them, whose assumptions become compliance requirements, and whether the public can challenge those choices before they disappear into software.

When a machine answers a political question, the struggle over that answer may already be over. The political achievement will have occurred long before the machine speaks.