Leverhulme Centre for the Future of Intelligence

Stated Theory of Change

LCFI's founding theory of change was that interdisciplinary research -- bringing philosophers, social scientists, cognitive scientists, and computer scientists together -- would help ensure AI develops safely and beneficially. The FLI announcement (2015) stated: "it will have the long-term safe and beneficial development of AI at its core." The founding team included x-risk luminaries (Bostrom, Russell, Price) and the proposal was "developed at CSER" and built on "the pioneering work of FHI, FLI and others."

The mechanism was intentionally broad: not just technical alignment, but the full spectrum of challenges from autonomous weapons to philosophical questions about intelligence. The Leverhulme Trust pitch emphasized "bold, disruptive thinking, capable of creating a step-change in our understanding."

Under Cave's leadership (from 2021), the articulated theory of change shifted to emphasize historical lessons about technological disruption, the importance of regulation, and the need for interdisciplinary ethics to inform governance. Cave frames it as: technology is inherently double-edged, and without proactive interdisciplinary research informing policy, AI could produce outcomes as destabilizing as the industrial revolution (which "fuelled the rise of communism and fascism, colonial expansion and the greenhouse gases that now threaten the biosphere").

Revealed Theory of Change

LCFI's actual operations reveal a multi-pronged theory of change that is diffuse but internally consistent:

1. Education pipeline: Train AI ethics professionals (MPhil, MSt) who disperse into AISI, tech companies, consulting, and policy roles. This is a long-game bet on shaping the next generation of AI governance professionals.

2. Policy influence: Directly engage with UK and EU regulation (CDEI creation, House of Lords testimony, EU Code of Practice feedback, HEAT toolkit). This is the most concrete impact pathway.

3. Intellectual framing: Shift public and academic discourse on AI (AI race critique, AI narratives research, Digital Minds). This operates at the level of ideas -- changing how people conceptualize AI development.

4. Convening power: Use Cambridge's institutional prestige to bring together academics, policymakers, and industry (Next Turing Tests conference, PAI membership, workshops). This creates connective tissue in the AI governance ecosystem.

What LCFI does NOT do, and what separates its revealed theory of change from its founding framing: no technical alignment research, no direct engagement with AI lab safety practices, no frontier model evaluations, no mathematical or empirical safety research. The research output is overwhelmingly humanities and social science.

The most notable divergence between stated and revealed theory: the founding announcement envisioned LCFI as "building on the pioneering work of FHI, FLI and others" -- positioning it in the technical AI safety lineage. In practice, LCFI evolved into an AI ethics and governance centre more comparable to Stanford HAI or the Oxford Institute for Ethics in AI than to MIRI, ARC, or Redwood Research.

Key Assumptions

Assumption 1: The governance/institutional approach matters for reducing AI risk.

• Evidence for: UK CDEI creation, EU AI Act engagement, compute governance report, placement of graduates in AISI. These are concrete governance outcomes.
• Evidence against: The key technical decisions about AI systems are made by labs where LCFI's humanities output has limited direct traction. Governance frameworks (AI Act, US Executive Order) have so far not demonstrably constrained frontier labs' most consequential decisions.
• Testable: Partially. One could evaluate whether LCFI graduates in governance roles make decisions that provably increase safety.
• If wrong: LCFI's entire body of work would be academically interesting but safety-irrelevant. The counterfactual world without LCFI might look very similar in terms of actual AI risk outcomes.

Assumption 2: Interdisciplinary humanities research adds something that technical safety research alone cannot.

• Evidence for: The AI race paper provides a political analysis that pure technical researchers wouldn't produce. Digital Minds addresses moral status questions that require philosophy. The HEAT toolkit bridges regulation and implementation.
• Evidence against: The most concretely safety-relevant LCFI output (compute governance) was co-authored with people who also had technical governance backgrounds. It's unclear that the specifically humanities dimension was necessary.
• If wrong: LCFI's distinctive contribution (the humanities angle) would be decorative rather than load-bearing.

Assumption 3: LCFI can sustain itself after Leverhulme Trust funding ends.

• Evidence for: Diversification to Mercator, Templeton, EU, Accenture. ITH umbrella. Cambridge institutional support. MPhil/MSt tuition.
• Evidence against: No public evidence of renewal. The next Leverhulme Research Centre round doesn't open until November 2026. Supplemental grants (~GBP 4M over 10 years) are insufficient to replace the core GBP 10M grant.
• If wrong: LCFI could contract significantly or be absorbed entirely into ITH, losing its distinct identity and research agenda.

Assumption 4: The education pipeline produces people who make a meaningful difference to AI safety outcomes.

• Evidence for: At least one AISI placement. 50+ alumni across multiple cohorts. Student testimonials describe career-changing experiences.
• Evidence against: No systematic tracking of outcomes. Alumni went into consulting, journalism, policy -- it's unclear how many are in roles with direct safety impact.
• Testable: Yes, through alumni career tracking, though LCFI hasn't published this data.
• If wrong: The education theory of change would be aspirational rather than demonstrated.

Strengths

1. Institutional prestige and convening power. Cambridge's name opens doors in UK government, EU institutions, and international organizations that most AI safety orgs cannot access. LCFI can get a minister to speak at its conference, submit evidence to parliamentary committees, and have its graduates credibly apply to AISI.

2. Genuine policy influence track record. The CDEI creation is a concrete, verifiable outcome. House of Lords testimony is documented. EU AI Act engagement is ongoing. These are not hypothetical -- LCFI has actually shaped UK AI governance institutions.

3. Intellectual quality. O hEigeartaigh's "Machines of Loving Grace" reflection demonstrates sophisticated engagement with AI safety debates. The AI race paper makes a legitimate analytical contribution. The compute governance report is technically substantive. LCFI's humanities-oriented work is high quality within its domain.

4. Education programme. Creating the world's first MSt in AI Ethics for working professionals and an MPhil producing AISI-ready graduates is a genuine structural contribution. If the pipeline works, it creates lasting human capital in AI governance.

5. Digital Minds initiative. This represents a pivot toward a topic (AI consciousness/welfare) that is both genuinely neglected and increasingly policy-relevant. Caviola's EA-adjacent background could reconnect LCFI with the broader AI safety community.

Weaknesses and Risks

1. Funding cliff. The Leverhulme Trust's 10-year grant is approaching its end (2025-2026) with no public evidence of renewal. This is an existential risk for the institution. If the core funding disappears, the supplemental grants (~GBP 4M) and university allocation cannot sustain current operations.

2. Talent drain. The consistent pattern of the most technically capable and policy-oriented people leaving (Ghahramani, Belfield, Whittlestone, Dihal) suggests LCFI functions partly as a stepping stone rather than a permanent home. The people who could produce the most safety-relevant work tend to leave for industry or dedicated policy organizations.

3. Diffuse mission. LCFI simultaneously works on AI narratives, feminist AI, ethical design, compute governance, AI consciousness, law-following AI, Global South AI governance, and immortality philosophy. This breadth is intellectually admirable but makes it difficult to identify a clear, concentrated theory of impact. A focused organization doing one thing well (like GovAI on governance or MIRI on alignment) may achieve more per dollar than LCFI's broad portfolio.

4. Weak connection to the AI safety community. Zero EA/rationalist community engagement, zero Open Phil funding, zero forum posts, no 80K Hours interviews. LCFI is invisible to the community most focused on reducing catastrophic AI risk. This isolation means LCFI's insights don't inform, and aren't informed by, the most active and well-funded safety efforts.

5. Long, indirect causal chain. The mechanism from "humanities research on AI narratives" to "reduced probability of catastrophic AI outcomes" involves many steps and uncertain connections. LCFI's most direct safety impact (CDEI creation, compute governance report) came from specific policy and governance work, not from its core humanities research programme.

6. No impact measurement. After nearly 10 years and GBP 10M+ in funding, there is no independent evaluation of whether LCFI has achieved its founding goals. The REF impact case study is self-reported. Without impact measurement, it is impossible to assess whether LCFI's theory of change is actually working.

Cross-References

• CSER (sister centre under ITH): More focused on existential risk. LCFI was "developed at CSER" but evolved in a different direction. Now co-housed under ITH, creating potential for cross-pollination but also confusion about distinct contributions.
• GovAI (Oxford): More focused, technically grounded AI governance research. Co-authored the compute governance report with LCFI. GovAI has stronger connections to the EA/rationalist community and more targeted policy influence.
• FHI (closed 2024): Was the Oxford spoke of the original LCFI network. Its closure removed one of the founding institutional pillars.
• CHAI (Berkeley, Russell): Russell was a founding LCFI spoke lead. CHAI does technical alignment work that LCFI does not.
• Stanford HAI: Most comparable US organization -- broad interdisciplinary AI research with policy engagement. Similarly criticized for diffuse mission.
• UK AISI: LCFI has placed at least one MPhil graduate there. AISI is the direct policy action arm; LCFI provides the research base.
• CLTR: Whittlestone's destination after leaving LCFI. More focused AI policy organization.

What Would Change This Assessment

• Evidence that the Leverhulme Trust has renewed LCFI's funding would significantly strengthen the sustainability assessment.
• A candid strategic interview with Cave discussing LCFI's impact, priorities, and self-assessment would provide crucial context for understanding the organization's self-model.
• Systematic data on MPhil/MSt graduate outcomes (how many in safety-relevant roles? what decisions have they influenced?) would test the education pipeline theory of change.
• A substantive external critique of LCFI's approach from a credible AI safety researcher would help triangulate the analysis. The current absence of criticism makes assessment harder.
• Evidence that LCFI's humanities research directly influenced a specific AI safety decision (beyond the general governance contributions) would strengthen the case for the humanities approach.
• A major new technical hire or research programme focused on alignment or evaluations would signal a strategic shift back toward the founding vision.

Self-Critique

Weakest claim: My assessment that LCFI's humanities research has "minimal counterfactual impact on AI safety outcomes" is not well-evidenced. I cannot prove the counterfactual -- that UK AI governance would be substantially similar without LCFI. LCFI might have diffuse but real effects that are hard to measure, which is different from having no effects.

Potential bias: I am likely biased toward valuing technical alignment research over governance/institutional work, because the evidence base for technical contributions (papers, tools, benchmarks) is more legible than the evidence base for governance contributions (networks, norms, institutional capacity). LCFI's theory of change may be working in ways that are genuinely important but hard to observe.

What I missed: I could not read Cave's Aeon essay on intelligence as domination, which would provide important intellectual context. I could not access the REF impact case study in full, which might contain concrete evidence of LCFI's policy influence that would update my assessment. I have not spoken to LCFI graduates, who could provide insider perspective on the education programme's effectiveness.

What a thoughtful person who disagrees would say: "You're judging a governance and ethics centre by the standards of a technical alignment lab. That's like criticizing a hospital's administration department for not doing surgery. LCFI's theory of change is that the institutional, political, and conceptual environment around AI matters for safety outcomes, and they're right -- the difference between a world with good AI governance institutions and one without is enormous. CDEI, the AI Act, the AI Safety Institute -- these institutions exist partly because of the intellectual groundwork that centres like LCFI laid. You're undervaluing this because the impact is diffuse and hard to attribute."

Single weakest point: I don't know whether the Leverhulme funding has been renewed. If it has, my concerns about institutional sustainability are overblown. If it hasn't, LCFI may be in more serious trouble than my analysis suggests. This single unknown significantly affects the reliability of the entire assessment.