Centre for the Study of Existential Risk (CSER)

Stated Theory of Change

CSER's stated theory of change has three components:

1. Study: Existential risks are understudied in academia. By housing x-risk research at Cambridge, CSER lends institutional credibility and attracts researchers who wouldn't work at less prestigious institutions.

2. Policy influence: Research findings are translated into policy recommendations through direct government engagement (APPG, UN, EU, DSIT secondments, Bletchley Declaration), public advocacy (Martin Rees' media presence), and cross-sector workshops.

3. Field-building: By hosting conferences, the MPhil program, visiting researchers, and collaborative projects, CSER grows the global community of people working on x-risk.

The causal chain is: rigorous interdisciplinary research at a prestigious institution produces actionable insights, which inform governance decisions through established policy channels, which reduces catastrophic risk. Simultaneously, CSER's field-building activities expand the pool of researchers and practitioners working on these problems.

Revealed Theory of Change

CSER's actions suggest a theory of change that differs from the stated one in important ways:

What they actually do: CSER functions primarily as a prestige-based convening platform and talent incubator rather than a research powerhouse. Its most impactful outputs (the Malicious Use report, the APPG, Intelligence Rising, the compute governance paper) are convening/synthesis products that bring together researchers from multiple institutions, not novel research findings. Its researchers leave for higher-impact positions at DeepMind, UK AISI, Columbia, and policy organizations.

Where stated and revealed diverge:

• CSER's website frames AI as central to its mission, but at most 20-25% of its research output is AI-related. The revealed priority is broad x-risk research weighted toward biosecurity and environment.
• CSER claims to "develop collaborative strategies to reduce these risks," but the specific strategies or their demonstrated risk reduction are never articulated. The outputs are papers, workshops, and policy engagement -- inputs to risk reduction, not demonstrations of it.
• The rapid director churn suggests CSER may not be able to retain the caliber of leadership it attracts. This implies the director role is used as a prestigious stepping stone rather than a career destination.

Talent pipeline function: The pattern of staff departures (Belfield to DeepMind, Avin to UK AISI, Whittlestone to CLTR, Connelly to Columbia) suggests CSER's most significant revealed function may be training and credentialing researchers who then have larger impact elsewhere. This is valuable but distinct from the stated theory of change.

Key Assumptions

1. Breadth enables cross-cutting insights (evidence: mixed) CSER assumes studying AI, bio, climate, and nuclear risks together produces insights unavailable to specialized organizations. Some evidence supports this: the Intelligence Rising game, which applies governance simulation across risk domains, and the cross-cutting risk methodology work. But no published research demonstrates specific cross-domain insights that couldn't have been produced by specialized organizations.

• Testable: Do CSER's cross-domain publications get cited by researchers in multiple risk domains?
• If wrong: CSER's breadth simply dilutes its AI safety impact without compensating benefits.

2. Cambridge prestige converts to policy influence (evidence: moderate) CSER's Cambridge affiliation gives it access to UK government (DSIT secondments, APPG), international institutions (UN, EU, OECD), and high-profile donors (Feinberg). The APPG and Bletchley involvement demonstrate real access.

• Testable: Can specific policy changes be traced to CSER input?
• If wrong: CSER has prestigious access but no policy leverage -- a common failure mode for academic policy engagement.

3. Governance-focused AI work reduces AI risk (evidence: weak) CSER's entire AI theory of change runs through governance rather than technical safety. This assumes that the binding constraint on AI safety is governance, not technical capability. The compute governance paper is the strongest case for this -- compute is a concrete governance lever.

• Testable: Do AI governance frameworks that CSER contributes to actually constrain frontier AI development?
• If wrong: Governance interventions without technical safety solutions are insufficient.

4. The field-building mission can survive chronic instability (evidence: uncertain) The MPhil program and Feinberg endowment bet on CSER's permanence. But the director turnover, brain drain, and funding fragility create an organization that may lack the institutional continuity needed to build a field.

• Testable: Does the MPhil attract strong students and does CSER retain leadership for >3 years?
• If wrong: CSER remains a revolving door that contributes individual careers but doesn't build lasting institutional capacity.

Strengths

1. Cambridge institutional platform: CSER has access to one of the world's most prestigious academic brands. This opens doors that no startup nonprofit can open -- Vatican workshops, UN tracks, parliamentary engagement, Harari visiting fellowships. The ITH integration deepens this embedding.

2. Policy engagement track record: The APPG, Bletchley involvement, DSIT secondments, and UN AI for Good summit participation demonstrate real policy access. These are not theoretical influence channels -- CSER researchers have been physically present in rooms where decisions are made.

3. Talent development: CSER alumni are in high-impact positions: Belfield at DeepMind's Frontier Planning, Avin leading UK AISI's Systemic Safety Fund, Whittlestone leading AI policy at CLTR. If you view CSER as a talent accelerator, it has been remarkably effective.

4. The Malicious Use report model: CSER demonstrated the ability to convene multi-institutional collaborations that produce high-impact policy-relevant outputs. This model is replicable and valuable.

5. Feinberg endowment + MPhil: These represent a genuine pivot toward permanence. If the endowment is $3-5M, it could fund leadership in perpetuity, and the MPhil creates institutional stickiness that pure research centers lack.

Weaknesses and Risks

1. Catastrophic brain drain: CSER lost nearly all AI-focused researchers in 2024-2025. Belfield, Avin, Beard, and Whittlestone collectively represented the vast majority of CSER's AI governance capacity. Replacing them with a new director (Amadae) whose expertise is nuclear security and game theory, not AI, represents a massive discontinuity.

2. Director revolving door: Four directors in four years is not a sign of organizational health. Each brings a different disciplinary lens. The FLI-funded position was meant to provide stability, but the first holder left before year 2. This pattern suggests either structural problems with the role, Cambridge institutional friction, or that CSER is a stepping stone rather than a destination.

3. Lack of focus: The 5/23 "aligned" ratio from Sempere, the "2+ orders of magnitude" value variance from EA Forum, and the acknowledged heavy non-AI research all point to the same conclusion: CSER's broad mandate means most of its resources are not deployed on AI safety. For someone evaluating CSER specifically as an AI safety organization, this is a critical weakness.

4. No technical AI safety work: CSER contributes nothing to alignment research, interpretability, model evaluations, or other technical safety work. Its entire contribution is governance/policy. While this is a legitimate niche, it means CSER depends entirely on others producing the technical solutions that governance must govern.

5. Financial fragility: ~$1.3M/year from fragmented sources, no anchor funder, grants that perpetually "begin to end." The Feinberg endowment may help, but CSER has been chronically underfunded relative to its ambitions.

6. Absence of Open Phil funding: The largest x-risk funder choosing not to fund CSER is a strong negative signal. Open Phil's evaluation processes are rigorous, and their omission of CSER suggests they see better marginal value elsewhere.

Cross-References

• vs. FHI (now closed): CSER was often called "the Cambridge counterpart to Oxford's FHI," but FHI was 4-7x larger, heavily Open Phil-funded, and produced both technical and governance work. FHI's closure didn't visibly benefit CSER.
• vs. GovAI: GovAI (originally Oxford, now independent) is the more focused AI governance research competitor. Sempere preferred FHI/GovAI over CSER for AI governance input.
• vs. CLTR/Alpenglow: Jess Whittlestone left CSER for CLTR, which is more focused on UK AI policy. CLTR/Alpenglow may be eating CSER's lunch on UK AI policy influence.
• vs. UK AISI: Avin's move to UK AISI suggests government has absorbed some of CSER's policy function. AISI has vastly more resources.
• vs. FLI: FLI co-founded by Tallinn (same as CSER), provides CSER's director funding. FLI is a funder/advocacy org while CSER is a research center -- complementary rather than competing.
• vs. BERI: BERI provided infrastructure support to CSER (multiple SFF grants via BERI-CSER collaboration). BERI's model of supporting university-affiliated x-risk groups was designed for organizations like CSER.

What Would Change This Assessment

1. Amadae publishes a clear strategic direction focusing CSER's resources on 2-3 high-priority areas with measurable goals. This would address the breadth/dilution concern.
2. Open Phil or another major funder makes a significant grant to CSER. This would signal that rigorous evaluators see value in CSER's approach.
3. CSER hires 2-3 strong AI governance researchers to replace Belfield/Avin. This would demonstrate the organization can rebuild AI capacity after brain drain.
4. The MPhil program produces graduates who go on to high-impact positions. This would validate the field-building theory of change.
5. Amadae stays for 3+ years. This would break the director revolving-door pattern and suggest the structural problems have been addressed.
6. CSER's director turnover is publicly explained. Any credible account of why four directors in four years would significantly update this assessment.

Conversely:

• Amadae leaving within 2 years would confirm the structural problem hypothesis.
• The MPhil failing to attract strong students would undermine the field-building pivot.
• Continued brain drain of remaining AI-relevant researchers would suggest terminal decline of CSER's AI safety relevance.

Self-Critique

Sources I should have checked but didn't:

• UK Charity Commission data for Cambridge University (CSER may appear in university-level filings)
• Specific Templeton, Grantham, and other foundation grant databases for exact amounts
• Cambridge University annual reports for ITH/CSER mentions
• Amadae's prior publications for clues about her strategic direction

Where this analysis potentially biased:

• I may be over-indexing on the AI safety community's views (Sempere, Larks) which evaluate CSER through a narrow alignment/longtermist lens. CSER's broader x-risk mandate may produce value that these evaluators systematically undercount.
• The brain drain narrative may be overstated -- staff leaving for high-impact positions could reflect CSER's success at talent development rather than organizational failure.
• I may be undervaluing the MPhil program and Feinberg endowment, which could represent a genuine inflection point that historical patterns don't capture.

What a thoughtful disagreer would say: "CSER's breadth is a feature, not a bug. The most dangerous risks are the ones that span categories -- AI-bio convergence, climate-nuclear cascades, AI-enabled bioweapons. An organization that studies these intersections is more valuable than one that only does AI alignment. You're judging CSER by the wrong metric."

Weakest claim: That CSER's absence of Open Phil funding reflects a considered negative judgment about CSER's value. It could equally reflect institutional barriers (Cambridge overhead), strategic disagreement about approach, or simply that CSER never applied.

What information would most change my view: A credible explanation for the director turnover pattern. If it's "Cambridge institutional friction makes the role frustrating" -- that's fixable and less concerning. If it's "each director realized the organization lacks strategic direction and left" -- that's much more concerning.