Oxford Martin AI Governance Initiative (AIGI)

Stated Theory of Change

AIGI's stated theory has two components that operate at different timescales:

Near-term mechanism: Produce governance research, train technical governance researchers (TAIG DPhils), convene multi-stakeholder workshops, and place affiliates at regulatory bodies (EU AI Office, UK AISI). This creates expertise and influence at the institutions that write AI rules today.

Long-term mechanism: Build the intellectual and institutional foundations for international AI governance through the jurisdictional certification framework (IAIO). Certify state regulatory jurisdictions -- not firms -- for AI compliance, with trade enforcement. This requires decades of diplomatic work and is explicitly modeled on how ICAO/FATF/IMO developed.

The causal chain: Research + convenings + diplomacy -> international standards development -> jurisdictional certification body -> trade-enforced compliance -> reduced probability of catastrophic AI outcomes.

Revealed Theory of Change

AIGI's actions are broadly consistent with its stated theory, with some notable emphases:

Diplomatic engagement is the revealed priority. Trager's ~100 talks/year, Track II US-China dialogues, presence at every major AI summit, and direct engagement with the EU, UK, India, and China governments show that policy influence -- not research publication -- is the primary output. The affiliate model (50+ people in the network) leverages a small core team for maximum diplomatic reach.

The NTIA RFI reveals strategic code-switching. AIGI's website presents balanced international governance rhetoric. The NTIA response, targeting the Trump administration, uses hawkish US-leadership framing, advocates renaming the AI Safety Institute to "AI Security Institute," and recommends export controls and data center expansion. Either AIGI genuinely believes in US hegemonic AI governance or it strategically adapts its message to its audience. Both possibilities are important: the first suggests AIGI's international framework is US-centric at its core; the second suggests sophisticated policy communication skills.

Technical AI governance is a genuinely novel contribution. The TAIG programme and Fazl Barez's hire represent a real attempt to bridge the gap between ML and governance. Chain-of-Thought Hijacking is a genuine safety contribution. Toby Ord's RL scaling analysis has governance implications. This is not just governance with extra steps -- it is building a new subfield.

GovAI overlap blurs institutional identity. With Trager, Garfinkel, Dafoe, and Pluckebaum all holding dual positions, it is difficult to determine where AIGI ends and GovAI begins. The jurisdictional certification paper appears on both organizations' websites. In practice, AIGI appears to be the "Oxford academic arm" of a broader governance network centered on GovAI, with AIGI providing the university prestige and DPhil pipeline while GovAI provides the independent nonprofit structure and policy placement.

Key Assumptions

1. International governance is feasible for a technology with no military substitute.

• Evidence for: ICAO/FATF/IMO exist and function. Trager's model borrows proven enforcement mechanisms (trade-based certification). Voluntary responsible scaling protocols have achieved some industry uptake.
• Evidence against: Trager himself admits "there isn't really a convincing single example of a case where major powers are limiting themselves on a technology for which they don't have military substitutes." AI is fundamentally dual-use. US-China strategic competition makes cooperation harder than in aviation or maritime. Verification is technically difficult (Maas et al. 2025).
• Testable: Yes, over decades. Intermediate test: does a jurisdictional certification pilot emerge within 5 years?
• If wrong: AIGI's core theoretical contribution is an elegant but impractical thought experiment, and the real governance action happens through domestic regulation and bilateral agreements rather than multilateral frameworks.

2. Oxford's institutional environment will sustain AIGI.

• Evidence for: Oxford Martin School is more receptive to interdisciplinary research than the Philosophy Faculty that killed FHI. The endowment provides stable funding.
• Evidence against: FHI died of "a thousand paper cuts." Oxford's bureaucratic culture killed a 19-year-old institute with global renown. AIGI is only ~2 years old. The hiring and fundraising constraints that damaged FHI are institutional to Oxford, not specific to one faculty.
• If wrong: AIGI suffers the same fate as FHI within 5-10 years.

3. The TAIG pipeline will produce people who are genuinely useful for governance.

• Evidence for: Garfinkel notes AI governance has "very few people" and most topics have "one person and two months of work" as the frontier. There is real demand for people with both ML and governance expertise.
• Evidence against: DPhil training takes 3-4+ years. The governance landscape may change dramatically before the first cohort completes. Technical research skills may not translate to policy effectiveness.
• If wrong: AIGI trains technically skilled researchers who contribute to academic literature but don't influence actual governance decisions.

4. Convening power translates to governance influence.

• Evidence for: ICML workshop drew EU AI Office head and UK AISI participation. Paris, Seoul, Geneva summit engagement is at the highest diplomatic level. Ziosi/Rajkumar Vice-Chairs shaped the EU GPAI Code of Practice.
• Evidence against: Workshops and summits frequently produce communiques without binding commitments. Trager acknowledges (IGCC podcast) that summits allow "political leaders to kick the can down the road." No evidence any AIGI workshop changed a specific policy outcome.
• If wrong: AIGI organizes well-attended events that produce interesting discussion but no measurable governance improvement.

Strengths

Institutional independence. AIGI is funded primarily by Oxford Martin School's endowment, not by CG/Open Phil or AI labs. This gives it independence from the funder dynamics that constrain most AI governance organizations. No single external entity can pressure AIGI's research agenda.

Global diplomatic reach. For a 2-year-old programme with 8 core staff, AIGI's geographic and institutional reach is remarkable: EU AI Office, UK AISI, OECD, UNESCO, ITU, Carnegie Endowment, Tsinghua University, India's MeitY, WAIC Shanghai. Trager is personally connected at the highest levels of international AI governance.

Novel theoretical contribution. The jurisdictional certification approach is the most detailed proposal for international AI governance that avoids the IAEA problems (proliferation risk of inspecting firms, sovereignty concerns). Whether or not it succeeds, it provides a concrete alternative to vague calls for "global AI governance."

Technical-governance bridge. TAIG, Barez's research, and Ord's scaling analysis are genuinely novel attempts to connect technical AI research with governance. Most governance orgs lack technical depth; most safety orgs lack governance expertise. AIGI is one of the few attempting to bridge both.

Visiting fellow network as force multiplier. Dafoe (DeepMind), de Zoete (UK government), Rajkumar (EU Code of Practice co-founder of UK AISI), Daws (UN diplomacy) -- these connections provide AIGI with access and influence far beyond its core staff size.

Weaknesses and Risks

Osborne conflict of interest. Co-Director Osborne co-founded Mind Foundry, a defence AI company ($44M raised, counter-drone and military sensor applications). He simultaneously co-directs an initiative studying AI governance. This undisclosed conflict undermines AIGI's credibility on military AI governance and commercial AI regulation.

Financial opacity. No public financial data of any kind. For an organization advocating governance transparency, the inability to see its own budget, funding sources, and allocation is a credibility gap.

GovAI entanglement. The heavy personnel overlap (Trager, Garfinkel, Dafoe, Pluckebaum) makes it difficult to evaluate AIGI as an independent entity. Is AIGI generating distinctive value or providing an Oxford University letterhead for work that GovAI was already doing?

Institutional fragility. Oxford's bureaucratic environment killed FHI. AIGI has no independent legal status, no separate governance board, and no financial autonomy. If Oxford Martin School deprioritizes AI governance, AIGI has no fallback.

Impact measurement vacuum. No published evaluation of whether any workshop, report, or secondment changed a specific policy decision. The volume of output (100 talks, 9 reports, 8 workshops) is impressive, but volume is not impact.

Zero community visibility. AIGI has zero presence in the EA/AI safety community. No forum posts, no 80K Hours appearances, no engagement with the people most focused on catastrophic AI risk. This either reflects a deliberate strategic choice or means AIGI is missing the constituency most likely to push for the strongest governance measures.

NTIA RFI tension. The hawkish, US-leadership framing in the NTIA response ("American leadership in artificial intelligence," renaming AISI to "AI Security Institute") sits uneasily with AIGI's international governance mission. If AIGI is strategically adapting its message, that is sophisticated but raises questions about which position reflects genuine beliefs.

Theoretical feasibility. Trager's own acknowledgment that no historical precedent exists for major powers limiting themselves on a dual-use technology without military substitutes is a fundamental challenge to AIGI's core theory.

Cross-References

vs. GovAI: The most important cross-reference. AIGI shares key personnel with GovAI (Trager, Garfinkel, Dafoe, Pluckebaum). GovAI is an independent nonprofit focused on field-building and talent placement; AIGI is a university programme focused on international governance frameworks and technical governance training. The division of labor appears to be: GovAI trains and places governance professionals; AIGI provides the Oxford academic platform for international diplomacy and DPhil training. They are complementary rather than competing, but the organizational boundaries are blurry.

vs. FHI: AIGI is the direct institutional successor to Oxford's AI governance tradition. FHI incubated the ideas, GovAI spun out, FHI died, and Toby Ord moved to AIGI. AIGI faces the same institutional risks that destroyed FHI, but operates under a more supportive school and with a smaller, less iconoclastic team.

vs. CSET (Georgetown): Both are university-affiliated AI governance research organizations. CSET is US-focused with national security orientation; AIGI is international with diplomatic orientation. CSET has deeper quantitative analysis; AIGI has broader geographic engagement.

vs. Concordia AI: Beijing-based partner in US-China dialogue. Concordia AI provides the Chinese side of the Track II dialogue that AIGI convenes. They are collaborators, not competitors.

vs. IAPS (Institute for AI Policy and Strategy): IAPS affiliate Renan Araujo is also an AIGI research affiliate. Both organizations work on AI governance policy, but IAPS is US-focused and more operational; AIGI is international and more academic.

What Would Change This Assessment

• Publication of AIGI's budget and funding sources would clarify the scale of operations and whether OpenAI money reaches AIGI.
• Evidence that the jurisdictional certification approach has been adopted by any state or international body as a framework for negotiation would significantly increase confidence in the theory of change.
• Disclosure and management of the Osborne-Mind Foundry conflict would reduce the governance risk.
• A clear articulation of AIGI vs GovAI division of labor from Trager or institutional leadership would clarify AIGI's distinctive value.
• Evidence of specific policy impact traceable to an AIGI workshop, report, or secondment would demonstrate that volume of engagement translates to governance outcomes.
• If Oxford Martin School begins to constrain AIGI (hiring freezes, bureaucratic friction) the way the Philosophy Faculty constrained FHI, this would signal existential institutional risk.
• If the NTIA framing reflects genuine US-hegemonist views rather than strategic communication, the assessment of AIGI's international governance credibility would need significant downward revision.

Self-Critique

Weakest claim: The assessment that AIGI's institutional independence (endowment funding) makes it more credible than CG-funded organizations. In practice, Oxford's institutional constraints may be as distorting as funder pressure -- just in different ways (academic incentive to publish, institutional conservatism, bureaucratic friction).

Potential bias: This analysis may overweight the Osborne-Mind Foundry conflict because it is the most concrete governance finding. In practice, if Osborne's involvement at AIGI is minimal (which the evidence suggests), the conflict may have limited practical impact.

What a thoughtful critic would say: "AIGI is Robert Trager plus an Oxford letterhead. The jurisdictional certification idea is elegant but will never be adopted by any major power. The TAIG programme trains people for an academic subfield, not for actual governance. The convenings are expensive conferences where elites congratulate each other. And the zero EA community presence means AIGI is invisible to the people who care most about AI risk."

What a thoughtful defender would say: "AIGI is the rare AI governance organization that operates at the actual international diplomatic level, with real access to EU AI Office heads, UK government advisers, Chinese scholars, and Indian policymakers. The jurisdictional certification approach is the only detailed proposal that avoids the IAEA problems. The TAIG programme is building a field that needs to exist. And the independence from EA funder dynamics is a genuine advantage."

Single weakest claim: That AIGI's convening and engagement activities translate to measurable governance impact. The evidence shows volume and reach but not influence on specific decisions.

Information that would most change my view: Internal data on whether any AIGI workshop, report, or secondment demonstrably changed a specific policy decision or regulatory outcome.