RAND Corporation (Global and Emerging Risks)

Theory of Change

RAND's AI safety theory of change, articulated by CEO Jason Matheny, rests on a nuclear technology analogy: build governance guardrails for AI comparable to those developed for nuclear weapons, so humanity can "get the benefit and reduce the risk." Matheny describes himself as an "apaco optimist" -- optimistic about humanity's trajectory if existential risks can be navigated -- and identifies AI and synthetic biology as the primary threats.

In his own words: "The only real threat to technological advancement is an apocalyptic catastrophe" (Business for Good podcast, 2023). He recommends Toby Ord's "The Precipice" as the most important book he's read. When asked what research pod he'd run at RAND, he chose "guardrails on AI and synthetic biology, comparable to the guardrails we put on nuclear technologies."

RAND's approach is specifically policy-focused and US-government-facing: produce technical analysis (model weight security, compute governance, AI evaluations, bio-risk assessments), conduct wargaming exercises with senior officials, train the next generation of tech policy analysts, and translate findings into federal regulation. The theory is that RAND's unique position -- trusted by the national security establishment, able to convene classified exercises, operating FFRDCs for DoD/DHS -- gives it policy influence that pure AI safety orgs lack.

What They Do

RAND's AI safety work is housed in the Global and Emerging Risks (GER) division, which contains several distinct centers:

RAND CAST (Center on AI, Security, and Technology): Founded by Sella Nevo. Produced the landmark report on securing AI model weights -- identifying 38 attack vectors and proposing 5 security levels. Lennart Heim led the compute governance team here until his recent departure.

Canary Project ($38M via Audacious Project + $10M from Open Phil): Collaboration with METR to develop AI evaluation methods and test systems for dangerous capabilities. Target: by 2027, enable rigorous safety evaluations for any new AI system before release.

Geopolitics of AGI Center (launched Dec 2023): Runs "Day After AGI" tabletop exercises simulating NSC responses to AI crises. The "Robot Insurgency" scenario tested US readiness for rogue AI cyberattacks with participation from former officials at DoD, State, Treasury, Commerce, and the IC. Conclusion: the US is "unprepared."

Technology and Security Policy Center (TASP): ~100 people, founded by Jeff Alstott. Broader emerging technology policy work, partly funded by Open Phil.

Meselson Center for Biodefense: Directed by Nevo. Biosecurity policy, including AI-bio intersection.

Notable research outputs: RAND's bio-risk red-team study (Jan 2024) found current LLMs do not meaningfully increase biological weapons risk compared to using the internet alone -- a nuanced empirical finding that pushes back against hype. Heim's compute governance work on hardware-enabled mechanisms and the AI Diffusion Framework influenced chip export controls (though the Trump administration rescinded the AI Diffusion Rule in May 2025).

Policy influence: RAND helped draft Section 4 of Biden's AI Executive Order (Oct 2023). RAND spokesperson confirmed involvement. All six policy recommendations from a senior RAND official's congressional testimony appeared in the EO.

Key People

Jason Matheny -- CEO since July 2022. The architect of RAND's AI safety pivot. Career arc: art history BA (Chicago), PhD applied economics (Johns Hopkins), research director at FHI Oxford, founded New Harvest (cultivated meat), directed IARPA, founded CSET Georgetown, Biden White House NSC/OSTP, then RAND CEO. Was an initial trustee of Anthropic's LTBT (stepped down Dec 2023 to preempt conflicts of interest). Deep EA/longtermist ties. Compensation: $1.04M (FY2024).

Sella Nevo -- Founding director of RAND CAST and Meselson Center. Previously founded Google's Karmel humanitarian ML group (70 people). ALLFED board member. Led the model weights security report. Brings genuine technical credibility plus x-risk orientation.

Lennart Heim (departed) -- Formerly led compute governance at RAND CAST. Previously at GovAI Oxford. Now independent researcher. His departure leaves a significant gap in RAND's compute governance portfolio.

Notable connections: Tasha McCauley is an adjunct RAND scientist who was on OpenAI's board (participated in Altman firing, left March 2024) and is an Effective Ventures UK trustee. Jeff Alstott (TASP founding director) worked with Matheny at the White House and has EA ties.

Money and Incentives

Total revenue: $467M (FY2023), $514M (FY2024). RAND is one of the world's largest policy research organizations.

Revenue breakdown: 89% from US federal government contracts. RAND operates 4 FFRDCs for DoD, Air Force, Army, and DHS. FY2024: $328M from federal government, $15.5M from program services, $7.5M from investments. Endowment: $330M. Net assets: $466M.

Business model: Contract research. RAND is paid to answer specific questions from government and foundation clients. 93% of revenue is classified as "contributions" (including government contracts).

Open Philanthropy grants: $65.5M across 13 grants (2014-2025). This is <2% of annual revenue but is strategically concentrated in the GER division. Most grants are explicitly "to be spent at the discretion of RAND president and CEO Jason Matheny." Largest: $16.5M biosecurity (2023), $10.5M emerging tech (2023), $10M Canary AI evaluation (2025), $6M TASP (2024).

Audacious Project: ~$38M for Canary (with METR), Oct 2024.

Defense funding context: Between 2014-2019, RAND received $1.029 billion from defense contractors and government agencies -- 95% of all such funding to the top 50 US think tanks (Center for International Policy data).

Incentive tensions:

RAND's primary financial incentive is to serve the US government, which may or may not align with global AI safety priorities.
Open Phil grants are small relative to RAND's budget but disproportionately fund the AI safety work. If Open Phil withdrew, the AI safety pivot could contract.
The "at Matheny's discretion" grant structure concentrates power and creates an appearance of personal patronage.
RAND has financial incentives to support the priorities of whoever funds it -- the same echo chamber critique applied to defense spending applies to EA-aligned spending.
If Matheny leaves, the AI safety pivot's survival is uncertain -- it depends heavily on him personally.

What Others Say

Internal dissent: At an October 2023 all-hands meeting, a RAND employee expressed concern that Open Phil ties "might undermine [RAND's] reputation for rigor and objectivity" and worried about prioritizing the "effective altruism agenda" over impartial policy development.

Congressional critique (Sen. Cruz, Sept 2024): "RAND apparently had an outsized role in drafting the AI Executive Order." Cruz lists RAND-Open Phil-Anthropic personnel connections and questions RAND's objectivity. The conflict-of-interest claims are substantive and largely confirmed.

Bipartisan congressional concern (House Science Committee, Dec 2023): Bipartisan letter to NIST warning about "self-referential" AI safety research from RAND and others, lacking "quality that comes from revision in response to critiques by subject matter experts."

Structural critique (Chalmers Johnson, Global Policy): RAND is "America's University of Imperialism." Historical argument that RAND has always served US military hegemony and was "super-cautious about speaking truth to power." The analytical errors in Vietnam and the Cold War stemmed from ideological commitment to American dominance.

Pentagon echo chamber (Inkstick, 2020): RAND received 95% of all defense-think tank funding. Congressional testimonies implicitly recommend more Pentagon spending. "There are enormous financial incentives for RAND to support the Pentagon."

Internal counterpoint (Marek Posard, RAND sociologist): Published commentary arguing philosophical debates about AI risks (EA vs e/acc) are a "distraction" from real AI policy work. Called both perspectives "not particularly helpful" -- "they're assumptions of what a small group thinks the world is."

Community visibility: RAND's AI safety work is essentially invisible on LessWrong and EA Forum (3 posts, all tangential). The AI safety community does not appear to discuss RAND's work substantively.

What's Absent

RAND does not publish data on how much of its $467M+ budget goes to AI safety work versus its dozens of other research areas. No public data exists on GER division headcount or growth trajectory. No external evaluation of RAND's AI safety portfolio exists despite $65.5M+ in Open Phil funding. RAND does not address how its US-national-security framing of AI safety relates to global AI safety priorities. RAND's AI safety work focuses entirely on governance and misuse (model theft, bioweapons, compute controls, evaluations) -- technical alignment is not discussed. No published conflict-of-interest policy specific to the Open Phil funding arrangement has been identified. Lennart Heim's departure occurred without a public explanation. No evidence of coordination with traditional AI safety research orgs (MIRI, ARC, Redwood) beyond the METR/Canary collaboration exists.