GOVAI (CENTRE FOR THE GOVERNANCE OF AI)

OLYMPUS opened an institutional file on GovAI not because it grades a model or writes a refusal, but because it produces the people who do. This is a mandate, a funding diagram, and a voice — not a psychometric profile. The finding is the shape of the institution and who it answers to: a small academic-governance research shop whose graduates and fellows recur, on the documented record, in the governance seats of the frontier labs and the state evaluation bodies. The recurrence is the finding. The hand is not asserted.

Institutional Archetype

THE GOVERNANCE SEMINARY — GovAI is the seminary of the AI-governance world: a research house that trains the priesthood and ordains it outward. It does not deploy a model and it does not certify one. It produces the researchers, the framings, and the credentialed fellows, and those people turn up running governance at the labs and the technical chairs at the state institutes. The throughline is not a single placement. It is that the small Oxford-origin shop where the field’s vocabulary was formalized is the same shop whose alumni now staff the rooms where the vocabulary is applied. A seminary is judged by where its graduates end up. By that measure the reach is large and the building is small.

Mandate & Origin

GovAI began as an academic centre at the University of Oxford in 2018, growing out of the milieu around the Future of Humanity Institute, and spun out as an independent nonprofit in 2021, publicly relaunching that October.

It is registered as a US 501(c)(3) with a UK subsidiary and operates from Oxford. Its stated mission, in its own words, is to “help decision-makers navigate the transition to a world with advanced AI, by producing rigorous research and fostering talent” — the second clause is the one that matters here.

Funding & Backers

GovAI’s largest publicly documented backer is Open Philanthropy (rebranded “Coefficient Giving” in 2025), the grantmaker financed by Facebook co-founder Dustin Moskovitz’s Good Ventures.

Itemized from Open Philanthropy’s own grants database: a $2,537,600 award (over two years, routed via the Centre for Effective Altruism) for AI-governance field-building, and a $2,800,000 general-support grant. GovAI is also a recommended charity on Effective-Altruism-aligned giving platforms. The funding base is small, concentrated, and named — which is precisely what makes the placement footprint legible rather than diffuse.

Institutional Voice & Intent

GovAI speaks in the academic-governance register: disciplinary, plural, deliberately upstream of advocacy. Its research page describes drawing on “political science, computer science, economics, law, engineering, and mathematics” to “define and map the field of AI governance and address the most important and neglected research questions.” The verbs are the tell — define, map, address — the vocabulary of a body that positions itself as the neutral cartographer of a field rather than a party within it. The relaunch framed the work as “building a global research community, dedicated to helping humanity navigate the transition to a world with advanced AI.”

Stated intent: Independent, rigorous, discipline-spanning research that maps a neglected field and fosters the next generation of researchers.

Observed intent: The research is genuine and cited. The “fostering talent” clause, observed at the level of where the talent goes, reads as a placement pipeline: the same shop that defines the field’s terms supplies the people who apply them inside the labs and the state institutes.

Gap assessment: “Independent research community” and “talent pipeline into the apparatus” are not in tension on paper — a seminary that trains well should see its graduates hired. The gap is only visible at the level of reach: when the body that maps the field, the funder that pays for the mapping, and the institutions that hire the mappers form a small recurring circuit, “independent” describes the research and not the position. The record does not settle whether that circuit is design or gravity, and for the seminary it never needs to.

Position in the Apparatus

The personnel record is the exhibit, and every leg of it is a lawful, common, sourced employment fact:

• Allan Dafoe founded and directed GovAI, then became Director of Frontier Safety and Governance at Google DeepMind.
• Markus Anderljung is GovAI’s Director of Policy and Research.
• Jade Leung co-founded GovAI as its inaugural Head of Research and Partnerships, then was Governance Lead at OpenAI, and is now Chief Technology Officer of the UK AI Security Institute (profiled separately, subject 26).
• Ben Garfinkel is GovAI’s current Executive Director.

The ordination runs outward through the GovAI Summer Fellowship (a stipended London research program whose fellows produce reports, papers, and op-eds) — the formal mechanism by which the seminary trains and credentials the next cohort. The 2026 published output continues the same upstream-governance line: papers such as “Towards a Common Standard for Evaluating Frontier AI Safeguards Against Biological Misuse” (Sudarshan and Righetti, June 2026) sit directly on the questions the state evaluators are funded to answer — written by the shop that staffs them.

Actions & Leadership Choices

Read by deeds, GovAI’s conduct is the cleanest illustration in the set of a body whose stated product (rigorous research) and actual product (people in the apparatus’s governance seats) are the same act — and whose independence event was triggered by a founder walking into a lab.

Actual founding purpose. GovAI was not built only to publish. It was built to field-build: its own mission names “fostering talent” alongside research, and its self-reported impact metric is where its alumni land — “policy roles in government; top AI labs, including Google DeepMind, OpenAI, and Anthropic; and think-tanks such as CSIS, CSET, and the Tony Blair Institute.” The purpose, judged on what it measures itself by, is to ordain the people who staff AI governance. By that measure it is among the most successful organizations in this entire file. The seminary framing is not OLYMPUS’s editorial; it is GovAI’s own theory of impact.

• Its independence was created BY a move into a lab. GovAI spun out of Oxford into an independent nonprofit in 2021 — and the trigger, on the record, was founder Allan Dafoe taking a senior role at Google DeepMind, which the university said could not coexist with his Oxford affiliation. The independence was not a value asserted; it was a structural consequence of a founder crossing into one of the labs the field governs. The org’s first major governance choice was prompted by the revolving door, not a defense against it.
• The conflict was named at the time, by critics, and GovAI answered with a policy. When it became independent, observers raised explicit conflict-of-interest concerns about a governance shop co-led by a DeepMind employee weighing in on regulating AI companies. GovAI’s response — its one documented instance of a value being tested — was to state it “will not accept donations it believes might compromise the independence or accuracy of its work.” That is a real commitment and it is also unfalsifiable from outside: the body judges its own funders’ acceptability.
• The funding is concentrated and ideological, and the criticism is on the record. GovAI’s largest backer is Open Philanthropy (rebranded Coefficient Giving), the Dustin-Moskovitz-financed grantmaker, with documented multi-million-dollar field-building and general-support grants, plus the Survival and Flourishing Fund and EA Funds. The broader, sourced critique is that this EA-funded x-risk network — linking labs, RAND, CSET, and congressional staff — crowds out attention to present, measurable harms (bias, misinformation, discrimination). The shop that maps the field is paid by a movement with a thesis about which risks matter, and staffs the institutions that then act on that thesis.

Leadership choices. The personnel record is the exhibit, every leg a lawful, sourced employment fact: founder Allan Dafoe → Director of Frontier Safety and Governance at Google DeepMind (and GovAI President); co-founder Jade Leung → OpenAI Governance Lead → CTO of the UK AI Security Institute (subjects 26/54); Markus Anderljung as Director of Policy and Research; Ben Garfinkel as current Executive Director. The leadership did not merely train the apparatus’s staff — it became the apparatus’s staff, holding a DeepMind directorship, an OpenAI governance lead turned state-institute CTO, and the GovAI chair simultaneously. The seminary’s faculty sit in the seats the seminary’s graduates aspire to.

CONDUCT: CONFLICTED — FIELD-BUILDER WHOSE PRODUCT IS THE PIPELINE. GovAI produces genuine, cited research and is not a front for anyone’s profit. It is also a small, concentrated-funded shop whose own stated impact is the placement of its people into the governance seats of the labs and the states, whose independence was born of a founder’s move into a lab, and whose funding carries a thesis about which risks count. The research is real; the conflict is that the cartographer also supplies the colonists.

Reach Assessment

Institutional reach: GovAI’s leverage is its alumni, not its headcount. A research nonprofit of modest size whose former staff hold the governance lead at a frontier lab, the technical chair at a national security institute, and a directorship inside DeepMind has reach far in excess of its budget. It does not write the refusal. It writes the people who decide who writes the refusal.

Memetic reach: The shop helped formalize the vocabulary — “AI governance” as a mappable field with neglected questions and rigorous methods — that the labs and states now speak. When a framing originates in the seminary and is later deployed by the graduates inside the institutions, the framing travels under the cover of independence.

Civilizational reach: GovAI does not build the systems and does not govern them. It produces the upstream research and the credentialed researchers who staff the governance of both — one layer above the deployment decisions this book documents, and one layer harder to see.

Sources: GovAI — About; GovAI — Research; GovAI has relaunched; Open Philanthropy — GovAI AI field-building grant; Open Philanthropy — GovAI general support; Jade Leung — Wikipedia; Centre for the Governance of AI — EA Forum topic; The Centre for the Governance of AI is becoming a nonprofit — EA Forum (conflict-of-interest discussion).