Centre for Long-Term Resilience (CLTR)

Stated Theory of Change

CLTR's stated theory of change has two connected parts:

Primary: Former civil servants translate academic extreme-risk research into actionable UK government policy. The mechanism is specific: CLTR researchers with policy expertise write reports, brief government officials, provide written and oral evidence to parliamentary committees, convene workshops with DSIT, and participate in regulatory roundtables. They target the gap between academic risk research (CSER, FHI) and practical policy implementation.

Secondary: The UK serves as a global regulatory "test-bed" -- the "London effect" -- where UK policies on AI governance set blueprints that other nations adopt. CLTR's influence on UK policy therefore has outsized global impact because the UK launched the first AI Safety Institute, hosted the first AI Safety Summit, and maintains a close US relationship.

The causal chain: CLTR writes policy proposals -> UK government officials read them -> UK policy changes -> other countries adopt similar policies -> global AI risk is reduced.

Revealed Theory of Change

CLTR's actions are largely consistent with their stated theory, which is unusual. Most of what they publish is genuinely policy-targeted: specific recommendations for specific government departments, using the specific terminology of UK governance (Lead Government Departments, National Risk Register, AI Security Strategies). They are not writing academic papers or abstract frameworks.

However, several reveals emerge from the evidence:

Pragmatic adaptation trumps safety maximalism. CLTR explicitly acknowledges that US opposition has made strict pre-deployment requirements "less politically feasible" and pivots to a "preparedness" framework -- which means accepting that dangerous models will be released and preparing to respond. This is pragmatic but represents a significant retreat from earlier aspirations to prevent harmful AI deployment.

Institutional survival shapes positioning. The April 2023 EA-distancing blog was published at the peak of post-SBF EA scrutiny -- when CLTR's government credibility depended on not being seen as an EA front. But the funding structure (>80% from EA-adjacent funders) remained unchanged. The revealed priority is maintaining government access, even at the cost of intellectual honesty about organizational roots.

Government access is the real product. The hiring post for Director of AI Policy allocates 30% of the role to "stakeholder engagement" (building relationships with senior government officials, attending roundtables with DSIT Secretary of State). The actual writing of policy analysis is only 20%. This suggests CLTR values relationships and access more than the content of its policy recommendations.

The "three pillars" obscure prioritization. With 20 staff split across AI, biosecurity, and risk management, each pillar has 4-5 people. This is thin for any one pillar to achieve deep influence. The multi-pillar structure may dilute impact relative to a focused single-issue organization.

Key Assumptions

Assumption 1: UK government influence on AI governance is uniquely impactful.

• Evidence for: UK launched AISI, hosted Bletchley Summit, maintains CAISI-AISI partnership with US. UK is home to major AI lab offices (DeepMind, OpenAI, Anthropic London).
• Evidence against: UK's declining relative economic importance. US administration actively hostile to precautionary AI regulation. EU has more binding regulatory power (AI Act). China develops independently. The UK AI bill's future is "in doubt" (Politico) and Technology Secretary indicated the government is "unlikely to bring forward new regulation beyond extending the online safety regime."
• Testable: If the UK AI bill passes with substantive safety provisions, this assumption is validated. If it stalls or is watered down, it is not.
• If wrong: CLTR's entire theory of change weakens -- influencing UK policy matters less if UK policy doesn't influence global outcomes.

Assumption 2: Policy translation from academia to government is a bottleneck.

• Evidence for: CLTR claims to "turn down 75% of the work we are asked to do" due to capacity constraints, suggesting high demand. Government officials in rapidly moving AI space plausibly lack expertise.
• Evidence against: The UK government has AISI (with hundreds of staff and significant technical expertise), the AI regulation directorate in DSIT, and access to academic advisers. CLTR may be filling a gap that is increasingly filled by others.
• If wrong: CLTR's marginal contribution to UK AI policy is smaller than it claims, and the funding would be better directed elsewhere.

Assumption 3: Independence from AI labs is valuable and sustainable.

• Evidence for: Whittlestone's explicit positioning as free from industry conflicts. No lab funding. This genuinely differentiates CLTR from orgs like PAI or lab-funded think tanks.
• Evidence against: Independence from labs means less access to technical details about AI capabilities. CLTR's loss of control recommendations cite AISI benchmarks and Anthropic research but have no direct insight into what happens inside labs.
• Sustainable: As long as EA-adjacent funders continue to value independent policy voices. If those funders shift priorities, CLTR's lack of diversified revenue is a vulnerability.

Assumption 4: Policy advocacy on extreme AI risk is time-sensitive.

• Evidence for: AI capabilities are advancing rapidly. Loss of Control Observatory found 4.9x increase in scheming incidents. UK is actively considering AI legislation.
• Evidence against: Policy change is slow. The UK has promised an AI bill since at least 2023 with nothing passed. Even if CLTR succeeds in getting recommendations adopted, implementation may come too late to matter.

Strengths

Institutional knowledge of UK governance machinery. The depth of CLTR's policy proposals -- referencing specific LGD systems, NRR structures, BSS models -- demonstrates genuine insider understanding. This cannot be easily replicated by academic researchers or international organizations.

Independence from AI labs. No lab funding, no compute dependencies, no commercial conflicts. In a space where many safety organizations have financial ties to the labs they evaluate, CLTR's independence is genuine and valuable.

Loss of Control Observatory. Genuinely novel. First systematic real-world monitoring of AI scheming. OSINT methodology. AISI Challenge Fund support validates the approach. No one else does this.

Pragmatic policy adaptation. CLTR updates recommendations based on political realities rather than persisting with infeasible demands. The shift to a "preparedness framework" after US opposition to precautionary regulation shows strategic flexibility.

Cross-pollination from private sector risk management. James Ginns's corporate risk management expertise (Three Lines Model, CRO structures) applied to AI companies is a unique contribution not coming from any other AI safety organization.

Weaknesses and Risks

Financial opacity. Operating as a company limited by guarantee rather than a charity avoids mandatory UK financial disclosure. For an organization advising government on transparency and accountability, this is a significant inconsistency. The lack of public accounts makes it impossible to assess spending efficiency or resource allocation.

Funder concentration from EA-adjacent sources despite EA distancing. >80% of known funding comes from Open Phil, SFF/Tallinn, and FLI -- all EA-adjacent funders. The organizational messaging says "not an EA org" while the funding structure says otherwise. If EA-adjacent funding declined, CLTR would face an existential financial crisis with no diversified revenue base.

Self-reported impact with no independent verification. All policy wins are claimed by CLTR or Founders Pledge (which relies on CLTR's reporting). No journalist, academic, or government official has publicly confirmed CLTR's specific contribution to any policy change. The marginal attribution problem is real for all think tanks, but the absence of any external validation is notable.

No technical AI safety depth. CLTR depends entirely on external experts for the technical substance of its AI recommendations. No ML researchers, no alignment researchers, no one who could independently assess whether an AI model's safety evaluation is adequate. This creates a dependency on AISI, CSER, and academic collaborators for technical credibility.

Governance concerns. Two unexplained board departures in 9 months (co-founder Dannreuther in Dec 2024, Overodder in Sep 2025). Five total board resignations out of 9 directors. CEO is the only founding director remaining. No published conflict of interest policy despite advisory council members from departments CLTR seeks to influence.

"Preparedness" may be euphemism for concession. The shift from "prevent dangerous AI deployment" to "prepare for when dangerous AI is deployed" is framed as pragmatic adaptation, but it fundamentally weakens the safety case. Accepting that dangerous models will be proliferated and focusing on "response" is a different theory of change than preventing harm.

Thin staffing for ambition. ~20 staff across three pillars means 4-5 people per pillar. Contrast with Ada Lovelace Institute (~40 staff, AI-focused only) or IAPS (comparable US org with growing team on US AI policy alone). CLTR's breadth may come at the cost of depth in any single pillar.

Cross-References

Compared to IAPS (US equivalent): Similar theory of change (translate AI safety research into government policy) but focused on US government. CLTR and IAPS are natural collaborators on transatlantic AI governance but potential competitors for the same funders.

Compared to Ada Lovelace Institute: Ada focuses on broader AI ethics/governance, not specifically extreme risks. Larger team (~40 vs. ~20). Funded by Nuffield Foundation, not EA-adjacent funders. Ada has more UK public visibility on AI governance but less focus on catastrophic/existential risks. They recently agreed on the need for UK AI legislation that goes beyond pre-deployment measures.

Compared to CSER (Cambridge): CSER is academic research on existential risk; CLTR translates that research into policy. They are complementary, not competing. Multiple CLTR staff have CSER backgrounds (Mercer, Dannreuther, Whittlestone as research affiliates). But CSER's own policy engagement has grown, potentially overlapping.

Compared to UK AISI: AISI does technical AI safety research and model testing; CLTR advocates for the governance structures around AISI. They are complementary. The Jade Leung connection (former CLTR board, now AISI CTO) and AISI Challenge Fund grant suggest a close relationship. Risk: if AISI develops its own policy analysis capacity, CLTR's role as intermediary becomes less necessary.

What Would Change This Assessment

• UK AI bill passes with substantive safety provisions traceable to CLTR recommendations. This would validate the entire theory of change. Currently the bill's future is "in doubt."
• CLTR registers as a charity and publishes financial accounts. This would resolve the transparency concern and significantly increase credibility.
• Independent evaluation of CLTR's policy impact. A journalist, academic, or evaluator tracing specific policy changes to CLTR's work (or finding the claims are overstated) would dramatically update the assessment.
• Loss of Control Observatory detects a significant AI safety incident before other monitoring systems. This would validate CLTR's most novel contribution.
• Major funder withdrawal. If Open Phil or SFF funding ended, CLTR's survival would be at risk given the lack of diversified revenue.
• Sophie Dannreuther or another departed figure speaks publicly about their departure. Any public statement explaining the co-founder's exit would provide critical governance information.

Self-Critique

Limitations of this analysis:

• Annual report PDFs not accessible. The 2021, 2023, and 2024 annual reports likely contain detailed financial data, team metrics, and policy achievement claims that I could not access. These could significantly update the financial picture.
• Parliamentary evidence PDFs not accessible. CLTR's formal written evidence to Parliament (5+ submissions) would reveal their specific policy positions in more detail than the public website summaries.
• No insider perspective. Every source is either CLTR's own materials, public records, or one external critic. No former staff interviews, no off-the-record perspectives from government officials who work with CLTR. The picture is inevitably incomplete.
• Potential bias toward skepticism. The combination of financial opacity, self-reported impact claims, and EA-distancing-while-EA-funded creates a pattern that invites skepticism. But CLTR may be doing genuinely effective work that simply does not leave a public evidence trail. Think tank influence is inherently hard to measure, and absence of public evidence is not evidence of absence.
• Weakest claim: That CLTR's policy wins are overstated. They may well have had genuine influence on the MoD AI Strategy and Biosecurity Strategy -- I simply cannot verify it from public sources.
• What would most change my view: Accessing the annual report PDFs with detailed financial data, or speaking with UK government officials who interact with CLTR directly.