Global Catastrophic Risk Institute

Stated Theory of Change

GCRI claims two interlocking mechanisms for reducing GCR:

1. Cross-risk integration: By studying all GCRs together, GCRI identifies synergies (actions that reduce multiple risks) and tradeoffs (actions that reduce one risk while increasing another) that siloed orgs miss. The canonical example: nuclear explosives for asteroid deflection could inadvertently increase nuclear war risk -- a tradeoff only visible to someone working across both risks.

2. Win-win solutions: By designing interventions that serve decision-makers' existing priorities alongside GCR reduction, GCRI makes implementation more likely. The canonical example: housing density policy that reduces costs (popular) and emissions (GCR-relevant).

On AI specifically, the mechanism is: apply professional risk analysis methods to AI catastrophe scenarios, identify governance mechanisms (corporate governance, collective action, certification) that could reduce risk while addressing other AI concerns (near-term harms, ethics), and push these into policy processes (NIST, IEEE, OSTP).

Revealed Theory of Change

What GCRI's actions suggest differs from the stated theory in several ways:

The cross-risk agenda lost to funding gravity. GCRI's stated raison d'etre is cross-risk analysis, but from 2019-2022, AI consumed the majority of research effort because that's where the money was (Irlam donations). Baum acknowledges: "our funding has mainly been for targeted work on specific risks." The cross-risk intellectual framework exists in papers and topic pages but has not been the primary research activity for years.

The real output is Seth Baum's scholarship. GCRI is, functionally, a research vehicle for one highly productive academic with unusually broad interests. The org structure provides a platform (website, fiscal sponsorship, advisor network) that legitimizes Baum's solo work as institutional output. This is not inherently bad -- many small think tanks are built around one person -- but it means GCRI's impact is bounded by one person's capacity.

The pivot from research to public commentary suggests a bet on influence-per-dollar. With funding collapsed, Baum cannot sustain the staffed research model. The 2025-2026 pivot to YouTube, symposia, and timely commentary (Iran war, Ukraine nuclear risk) appears to be a rational adaptation: public intellectual work requires less money than peer-reviewed research and could have higher marginal impact if it reaches the right audiences.

Policy work was Barrett's domain, and it left with him. Barrett drove the NIST, IEEE, and government engagement. His departure to DHS CISA is both a success (GCRI produced a government policymaker) and a loss (no one at GCRI now does policy work). The theory of change around policy influence has a broken link.

Key Assumptions

Assumption 1: Cross-risk analysis produces insights that single-risk analysis cannot.

• Evidence for: The asteroid-nuclear tradeoff example is genuinely instructive. Climate change interacting with food supply catastrophes. GCRI's COVID origin survey drawing on biosecurity methods.
• Evidence against: Most cross-risk "insights" at the GCR level are relatively obvious (wars worsen pandemics, climate change worsens conflict). The deep technical work happens within single risks.
• Testable: Are there GCRI cross-risk findings that changed a specific policy or research agenda?
• If wrong: GCRI's differentiator is illusory, and they are just a small generalist org competing poorly with larger specialists.

Assumption 2: Governance and risk-analysis approaches to AI are valuable for reducing AI catastrophe risk.

• Evidence for: Corporate governance paper cited by Jonas Schuett (now leading AI governance research at GovAI). NIST framework engagement. The LLM takeover assessment introduces a practical monitoring framework.
• Evidence against: AI governance work is abundant and GCRI's contributions are small relative to GovAI, CSET, CAIP, FHI's governance program. Risk analysis of AI is valuable in principle but GCRI lacks the technical depth to assess actual AI systems.
• If wrong: GCRI's AI work is intellectually sound but practically redundant given larger, better-funded governance orgs.

Assumption 3: One person can meaningfully reduce GCR.

• Evidence for: Baum has published in top journals, been cited by Brookings scholars, placed a co-founder in DHS CISA, and advised 50+ people annually on GCR careers. Not nothing.
• Evidence against: The marginal impact of one person's commentary and research, in a field with thousands of researchers and billions of dollars at stake, is extremely small. GCRI's claim to institutional significance rests on being "one of the oldest" GCR orgs, not on demonstrated impact.
• If wrong: GCRI should merge with a larger institution or Baum should seek a position at CSER/GovAI/a university where his research would have institutional backing.

Assumption 4: Baum's heterodox AI views (LLMs as bags of heuristics, products not gods) are closer to correct than the AI safety mainstream.

• Evidence for: Scaling laws did hit diminishing returns in 2024. Reasoning models face questions. The "products not gods" framing aligns with growing industry sentiment post-hype.
• Evidence against: Capabilities continue to surprise. "Bags of heuristics" may be too dismissive of what large collections of heuristics can do. The mainstream AI safety community focuses on catastrophic tail risk even if most likely outcomes are modest.
• Testable: If LLM-based or reasoning-model-based systems achieve dramatically new capabilities in 2026-2027, this weakens Baum's position significantly.
• If wrong: GCRI is providing false reassurance about AI risk at the worst possible time.

Strengths

1. Independence. Zero AI lab funding, zero corporate influence, no board conflicts. GCRI may be the only AI safety org with genuinely clean incentives. When Baum says LLMs are bags of heuristics, he has no financial reason to say otherwise.

2. Intellectual range. Baum is one of very few people who can write credibly about nuclear probability modeling, AI governance, asteroid risk tradeoffs, climate ethics, and social choice theory. This breadth is rare and valuable for the cross-risk mission.

3. Longevity and institutional memory. Founded 2011, GCRI predates most of the current AI safety ecosystem. Baum has been working on GCR since before "AI safety" was a career. This experience provides perspective that newer orgs lack.

4. Barrett's government placement. Getting a co-founder into DHS CISA is a concrete policy outcome that most research orgs would envy.

5. Publication in quality venues. Risk Analysis, IEEE Transactions, Technology in Society, Science -- these are not vanity publications. GCRI's work meets peer review standards in multiple disciplines.

Weaknesses and Risks

1. Organizational near-death. GCRI is a one-person operation that considered disbanding in 2024. The org has no financial reserves (post-2022), no staff pipeline, no succession plan, and no visible path to sustainability. The risk of GCRI ceasing to exist in the next 2-3 years is high.

2. Failure to scale over 14 years. GCRI peaked at 5 staff and $400K budget. The $1.5M scaling goal was never approached. Every hire eventually left. This pattern suggests something structural -- whether it's the cross-risk framing being unfundable, Baum's management, compensation constraints, or career growth limitations for junior staff.

3. No demonstrated real-world impact. After 14 years and 60+ publications, GCRI cannot point to: a policy change they caused, a specific disaster they helped prevent, a quantifiable influence on AI governance, or even a widely-cited paper that shifted the field's thinking. The closest is Barrett's government career.

4. Invisible to the AI safety community. Zero EA forum presence, zero CG/OpenPhil funding, near-zero citations in the AI safety discourse. GCRI publishes for academic risk analysis audiences, not for the community that drives AI safety funding, careers, and policy. This means GCRI's work is unlikely to influence the organizations or people most directly working on AI catastrophe prevention.

5. Baum's AI views may be too dismissive. The "bags of heuristics" framing and skepticism of imminent AGI could lead to complacency if wrong. While intellectual independence is a strength, being systematically wrong about the central question in your field is disabling.

Cross-References

Compared to CSER (Cambridge): CSER is the natural comparison -- academic, cross-risk, credible researchers. CSER has university backing, larger team, more funding, and Sean O hEigeartaigh (GCRI advisor) as a key figure. GCRI is essentially a miniature, independent version of what CSER does with institutional support.

Compared to GovAI: GCRI's AI governance work overlaps with GovAI (Jonas Schuett, a GCRI co-author, is now at GovAI). GovAI has more resources, more researchers, and institutional home at Oxford. GCRI's advantage is independence from any university's politics.

Compared to FHI (now closed): FHI's closure in 2024 leaves a gap in foundational GCR/ER research. GCRI claims the "oldest active GCR institute" mantle. However, FHI's output in its prime dwarfed GCRI's.

Compared to ALLFED: Interesting comparison -- ALLFED's David Denkenberger is the other most prolific GCR author per the bibliometric review. Both are small, focused GCR orgs. ALLFED is more applied (food security), GCRI is more theoretical (risk analysis, governance).

What Would Change This Assessment

• New significant funder (>$200K/year) willing to support cross-risk work or GCRI's public outreach pivot. Would signal external validation and enable re-hiring.
• Baum's "bags of heuristics" thesis vindicated by continued LLM scaling stalls and no dramatic new capabilities. Would increase the value of his independent analysis.
• Policy outcome traceable to GCRI research -- e.g., a government adopting their monitoring framework from the LLM takeover paper, or the all-hazards approach being implemented.
• YouTube/public outreach gaining significant traction -- e.g., 10K+ subscribers, media invitations increasing. Would validate the pivot.
• Conversely: If Baum cannot find new funding by 2027, GCRI will likely cease to exist and this would confirm the org never found product-market fit.

Self-Critique

What sources should I have checked but didn't:

• Baum's 2015 and 2021 EA Forum AMAs (blocked domain) would provide candid Q&A about GCRI's strategy and views.
• Larks' annual alignment reviews (LessWrong, blocked) would provide yearly external assessments.
• Google Scholar citation analysis would reveal which GCRI papers have actually been cited by other researchers.

Where is this analysis potentially biased:

• I may be too harsh on organizational viability because the financial data makes GCRI look more dire than it may be (Baum may have private funding or personal resources not reflected in tracked donations).
• I may underweight the long-term influence of publications -- a paper's impact often takes years to manifest through citation chains and policy adoption.
• The zero-EA-forum-presence finding may reflect more about the EA forum ecosystem than about GCRI's actual influence in academic and policy circles.

What would a thoughtful person who disagrees say: "GCRI is a lean, independent research institute that has produced a remarkably prolific body of work on $2M total over 14 years. Baum is one of the most knowledgeable cross-risk scholars in the world, his work is cited by Brookings and published in top journals, and his co-founder is now at DHS CISA. The fact that EA funders haven't funded GCRI says more about EA funder myopia than about GCRI's quality."

My single weakest claim: That GCRI has "no demonstrated real-world impact." Barrett's DHS appointment, Baum's media commentary during nuclear crises, the NIST/IEEE policy engagement, and the advising of 200+ people on GCR careers may collectively represent meaningful impact that is simply hard to measure. I may be applying an unfairly high evidentiary bar.

What information would most change my view: Evidence that a specific government policy, corporate governance change, or institutional decision was directly influenced by GCRI research. Even one clear case would significantly upgrade my assessment of their theory of change.