Oxford Martin AI Governance Initiative (AIGI)

Theory of Change

AIGI's theory of change rests on a single, distinctive idea: international AI governance should certify state jurisdictions, not individual firms. Co-Director Robert Trager's flagship proposal (arXiv, August 2023) calls for an International AI Organization (IAIO) modeled on ICAO (aviation), IMO (maritime), and FATF (money laundering). The IAIO would certify that a state's domestic AI regulatory regime meets international standards. Non-certified jurisdictions face import bans on AI goods and export restrictions on AI inputs like chips.

Trager, IGCC podcast (June 2024): "I think a body like that, in AI, would be a great idea... you can have ties, for instance, to the trade regime, and you can say that if a jurisdiction doesn't have certification, if it's violating international standards, then other countries can say, well, we're not going to import any AI whose supply chains have something coming from that jurisdiction."

The broader framework has three pillars: (1) develop international safety standards (through AI Safety Institutes), (2) create an international reporting regime for large training runs (compute providers report to governments, who report to an international body, analogous to money laundering reporting), and (3) the jurisdictional certification body.

Why certify jurisdictions and not firms? Trager argues (Just Security podcast, February 2024) that no frontier AI state will allow an international organization to inspect its AI companies: "that would be a highly proliferating moment to have an international organization that was governed internationally going and doing something like that. And everybody has always thought that the IAEA has had quite a few spies in there."

AIGI combines this international governance framework with a newer Technical AI Governance programme (TAIG) that funds DPhil students in engineering/computer science to build the technical expertise governance currently lacks.

What They Do

AIGI is a research programme within Oxford Martin School, University of Oxford, established circa 2023 with ~8 core staff, 7 DPhil students, ~40 research affiliates and visiting fellows (~55 in network). Five research areas: International AI Governance, Technical AI Governance, Frontier AI Governance, China AI Governance, Social Impact.

Policy engagement. Co-Director Trager delivered ~100 talks in 2024 at venues including the AI Seoul Summit, Geneva AI for Good Summit, and Paris AI Action Summit. AIGI submitted a formal response to the US NTIA AI Action Plan RFI (April 2025). Affiliates Marta Ziosi and Nitarshan Rajkumar served as Vice-Chairs drafting the EU General-Purpose AI Code of Practice under the EU AI Act, which has been signed by OpenAI, Anthropic, Google, and other major providers. Staff have been seconded to the EU AI Office and UK AISI.

Convenings. In 2024-2025: 8+ high-level workshops on risk tiers, open-source AI governance, safety frameworks alignment, and AI Summit reform. ICML 2025 TAIG Workshop (80 submissions, 64 accepted). WAIC Shanghai cross-border crisis coordination workshop. New Delhi pre-summit dialogue (officially designated Pre-Summit Event by India's MeitY). Track II US-UK-China AI safety dialogue with Carnegie Endowment and Tsinghua University.

Research publications. 9 reports in 2024, expanding in 2025. Notable: Verification for International AI Governance (Harack, endorsed by Yoshua Bengio), Chain-of-Thought Hijacking (Barez, 99% attack success rate on Gemini 2.5 Pro, covered by Fortune), "Open Problems in Technical AI Governance" (agenda-setting), Annual AI Governance Report with ITU, Agentic Inequality, comparative analysis of safety frameworks and standards.

Technical AI Safety. Fazl Barez (joined February 2025) leads technical safety research including interpretability and alignment. He taught Oxford's inaugural "AI Safety and Alignment" course in Engineering Science. Toby Ord published widely discussed analysis arguing RL scaling will plateau due to information-theoretic inefficiency ("a millionth as much information to be gained per FLOP in RL training compared with pre-training").

Key People

Robert Trager (Co-Director): Political scientist, UCLA PhD, also International Governance Lead at GovAI and Senior Research Fellow at Blavatnik School. The operational engine of AIGI -- personally responsible for most policy engagement. ~100 talks in 2024. The jurisdictional certification approach is his core intellectual contribution.

Michael Osborne (Co-Director): ML Professor, Oxford Engineering Science. Co-founder and CSO of Mind Foundry, a defence/national security AI company ($44M total raised, counter-drone detection, acoustic tripwire). Known for "The Future of Employment" paper (13K+ citations). His day-to-day governance involvement at AIGI appears limited; he primarily provides ML expertise and Oxford institutional weight.

Fazl Barez (TAIG Lead, Senior Researcher): Key technical hire. Collaborations with Anthropic, UK AISI, CSER Cambridge. N2G algorithm adopted by OpenAI. Brought genuine AI safety research capability to what was previously a pure governance shop.

Toby Ord (Senior Researcher): Author of The Precipice, co-founder of EA movement. Former FHI. Moved to AIGI after FHI closed April 2024. Provides brand recognition and intellectual weight on existential risk.

Team of ~55 total. Notable visiting fellows: Allan Dafoe (GovAI founder, Google DeepMind), Henry de Zoete (former PM AI adviser, led Bletchley Summit), Sam Daws (former UN SG staff, 35 years multilateral diplomacy).

Money and Incentives

Budget: completely opaque. As a university programme (not a separate legal entity), AIGI has no public financial reporting. No 990 (not a US nonprofit), no UK charity accounts. Budget unknown.

Estimated scale. Based on comparables: 8+ staff salaries (GBP 39K-58K range for postdocs), 7 DPhil stipends (~GBP 25-35K each), programme management, and extensive international travel for ~100 talks/year suggests an annual budget of GBP 500K-1.5M. Oxford Martin programmes typically receive GBP 250K-1M over 2-5 year terms, so AIGI likely draws on additional sources or receives above-average allocation.

Primary funding: Oxford Martin School endowment. James Martin donated GBP 70M+ in 2005; additional $100M in matched funding by 2010. The endowment supports ~25 interdisciplinary research programmes. AIGI's funding is fundamentally different from most AI safety orgs -- it draws on an established university endowment, not EA/longtermist philanthropy.

Coefficient Giving/Open Phil: $35K (October 2024, via BERI). Trivially small. AIGI is not dependent on the CG/Open Phil ecosystem that funds most AI governance organizations.

OpenAI co-funding question. Oxford Martin School and OpenAI jointly fund a GBP 1M generative AI programme (2025-2026), with OpenAI API credits and compute access. It is unclear whether AIGI benefits from this. If it does, this creates a conflict of interest: AIGI studies governance that directly affects OpenAI.

Incentive analysis. AIGI's university endowment funding gives it unusual independence from the funder dynamics that affect most AI safety orgs. No single external funder has leverage. The main incentive risk is institutional: Oxford Martin School's priorities shape AIGI's existence. If Oxford Martin decides AI governance is no longer a priority, AIGI dissolves -- exactly as happened to FHI under the Philosophy Faculty.

Conflict of interest: Mind Foundry. Co-Director Osborne co-founded and serves as CSO of Mind Foundry, a defence AI company ($44M raised, counter-drone and military sensor applications). This is a direct conflict of interest with his AIGI governance role. No disclosure found in any AIGI publication.

What Others Say

No direct criticism of AIGI exists. Zero EA Forum, LessWrong, or Alignment Forum posts. Zero news articles critiquing AIGI. The org is too new (~2 years) and too academic for targeted criticism, and operates outside the EA community where most AI safety critique occurs.

The strongest indirect criticism comes from structural analysis of international AI governance feasibility:

Maas et al. (2025), "International Agreements on AI Safety": Reviews all major proposals including Trager's jurisdictional certification. Identifies barriers: verification difficulties (even with hardware/cryptographic tools), enforcement gaps, geopolitical tensions (US skepticism of multilateralism), sovereignty concerns, and the fundamental problem that no precedent exists for "major powers limiting themselves on a technology for which they don't have military substitutes."

Trager himself (Just Security podcast, 2024): "arms control has, everybody knows arms control is challenging, but it's actually more challenging for dual-use technologies like advanced AI, that there aren't military substitutes for." He acknowledges there is no convincing historical example of what he is proposing.

FHI precedent. FHI's closure after 19 years at Oxford shows the institutional risk. "Death by a thousand paper cuts" -- hiring freezes, fundraising restrictions, bureaucratic obstruction. AIGI operates in the same university, though under a different and more receptive school (Oxford Martin vs Philosophy Faculty).

What's Absent

Budget and all financial details -- completely opaque
Conflict of interest disclosure for Osborne's Mind Foundry role
Articulation of AIGI vs GovAI division of labor (multiple shared personnel, overlapping work)
Any presence in EA/rationalist community forums (zero posts)
Published impact evaluation or self-assessment
Clarity on whether OpenAI co-funding reaches AIGI
80,000 Hours or EA community media appearances
Osborne's actual day-to-day governance involvement
Diversity data despite international governance focus