RAND Corporation (Global and Emerging Risks)

Stated Theory of Change

RAND's stated AI safety theory of change is a nuclear technology analogy: develop governance guardrails for AI comparable to those built for nuclear weapons, enabling humanity to "get the benefit and reduce the risk." The mechanism runs through RAND's unique institutional position:

1. Produce rigorous, nonpartisan analysis of AI risks (model weight security, compute governance, biological misuse, dangerous capabilities)
2. Convene senior policymakers in classified wargaming exercises (Day After AGI) to surface capability gaps
3. Shape federal regulation by providing analytical foundations for executive orders, NIST standards, and legislation
4. Train the next generation of tech policy analysts via fellowships and the RAND School
5. Evaluate frontier AI systems through the Canary project (with METR) before public release

The implicit claim is that RAND's 78-year institutional credibility, government trust, FFRDC access, and ability to work at classified levels give it policy leverage that pure AI safety organizations cannot match.

Revealed Theory of Change

RAND's actions largely match its stated theory, but with several meaningful divergences:

What actions reveal:

• The AI safety pivot is disproportionately driven by one person (Matheny) and one funder (Open Phil). Most Open Phil grants are "at Matheny's discretion." If either actor departed, the pivot could reverse.
• AI safety is a small initiative within a massive defense institution. RAND produces ~1,000 reports per year across dozens of policy areas. The AI safety fraction is unknown but likely small.
• The work focuses on misuse prevention and governance (model theft, bioweapons, compute controls, evaluations), not technical alignment. This is a deliberate scope choice reflecting RAND's institutional capabilities.
• RAND's AI safety framing is consistently US-national-security-centric. The Geopolitics of AGI Initiative asks "How should the US government respond?" not "How should humanity respond?"
• RAND has hired from the EA/AI safety community (Heim from GovAI, Nevo from Google, fellows via Horizon/Open Phil pipeline) but does not appear to coordinate with traditional AI safety research orgs beyond the METR collaboration.

Where actions diverge from stated theory:

• RAND claims nonpartisan objectivity, but its AI safety work is funded by a small number of EA-aligned sources and staffed by people with EA ties, producing research aligned with EA priorities. This doesn't mean the research is wrong, but it complicates the "nonpartisan" claim.
• Matheny sat on Anthropic's LTBT while RAND was being funded by Anthropic's largest backer (Open Phil) and while RAND personnel were drafting regulations affecting Anthropic. He only stepped down after the conflict was publicly exposed.
• RAND's bio-risk study (LLMs don't help with bioweapons) is genuinely independent-minded -- it pushes back against EA/AI safety community consensus. This is evidence that the nonpartisan instinct still functions.

Key Assumptions

Assumption 1: US national security and global AI safety are aligned.

• Evidence for: Many x-risks (bioweapons, rogue AI, compute proliferation) genuinely threaten the US and the world simultaneously. US policy leadership could set global norms.
• Evidence against: US-China AI competition could lead to "safety race to the bottom." Export controls may accelerate Chinese domestic chip development. US-centric framing excludes global coordination scenarios.
• Testable: Yes -- does US AI regulation actually reduce global risk, or does it just shift capability development to less safety-conscious actors?
• If wrong: RAND's entire approach could produce governance that protects US interests but increases global risk.

Assumption 2: Policy interventions are the most impactful lever for AI safety.

• Evidence for: Compute governance is genuinely tractable (chips are detectable, excludable, quantifiable). Evaluations could catch dangerous capabilities before deployment. RAND's wargaming reveals real preparedness gaps.
• Evidence against: Policy is fragile to political shifts (Biden AI Diffusion Rule rescinded by Trump). Regulation follows capability, not the reverse. Technical alignment may matter more than governance if AI systems become genuinely autonomous.
• Testable: Partially -- the durability of RAND's policy contributions across administrations is already being tested.
• If wrong: RAND's entire portfolio is second-order -- necessary but insufficient without complementary technical safety work.

Assumption 3: RAND's institutional credibility transfers to AI safety topics.

• Evidence for: RAND can convene government officials that no AI safety org can. The Day After AGI exercises involved senior DoD/State/IC personnel. RAND's analysis shaped the Biden EO.
• Evidence against: The EA controversy has damaged credibility with some policymakers. Congressional scrutiny of RAND-Open Phil ties suggests trust erosion. Under a Trump-style administration, RAND's Biden-era AI safety work may be viewed as partisan.
• Testable: Yes -- does the next administration use RAND's AI safety analysis?
• If wrong: RAND's policy influence is administration-dependent, not institutional.

Assumption 4: Open Phil funding doesn't compromise RAND's objectivity.

• Evidence for: RAND's bio-risk study contradicted some EA/AI safety community claims. Open Phil grants are <2% of RAND's total revenue. RAND's transparency about funding is above average for think tanks.
• Evidence against: Internal employee concern about EA agenda. "At Matheny's discretion" grant structure. Personnel overlap with Anthropic/OpenAI governance. The grants fund the specific work they're meant to influence.
• Testable: Difficult -- would RAND produce the same analysis if Open Phil funding disappeared?
• If wrong: RAND's AI safety research carries an institutional bias toward EA priorities.

Strengths

1. Unique convening power. No other AI safety organization can run classified wargaming exercises with former NSC principals. The Day After AGI exercises surface real preparedness gaps that no amount of academic analysis can identify.

2. Scale and institutional depth. 1,850 employees, $467M revenue, 78-year reputation. RAND can absorb political attacks that would destroy smaller organizations. The AI safety work benefits from being embedded in a large, diversified institution.

3. Technical credibility in security domains. The model weights report (38 attack vectors, 5 security levels) demonstrates serious security expertise. The bio-risk red-team study shows willingness to publish nuanced findings that contradict hype.

4. Policy translation capacity. RAND can translate AI safety concepts into the language of national security policy. Compute governance, model evaluations, and wargaming are policy-relevant outputs, not just academic papers.

5. Matheny's personal network. FHI, IARPA, CSET, White House, National Security Commission on AI -- Matheny has a unique network spanning the AI safety community and the national security establishment.

Weaknesses and Risks

1. Single-point-of-failure in Matheny. The AI safety pivot is driven by one CEO who controls most of the EA-aligned funding. Board composition (national security establishment, no AI safety expertise) suggests limited institutional commitment independent of the CEO.

2. Structural conflicts of interest. The RAND-Open Phil-Anthropic triangle is a genuine problem. Matheny's LTBT membership while RAND was shaping regulation affecting Anthropic, McCauley's dual RAND-OpenAI role, and the "at discretion" grant structure all undermine the nonpartisan claim that is RAND's core institutional asset.

3. National security tunnel vision. AI safety framed entirely through US national security may miss or worsen global coordination problems. RAND's institutional DNA favors the US perspective, which may be insufficient for a global risk.

4. No technical alignment work. RAND's portfolio covers governance and misuse but not technical alignment. If AI risk comes primarily from alignment failure rather than misuse, RAND's contributions are peripheral.

5. Political fragility. The Biden AI EO was closely associated with RAND. Under a different administration, this association is a liability. The Trump admin already rescinded RAND-influenced export controls. Policy work that doesn't survive political transitions has limited long-term value.

6. AI safety as a small island in a defense ocean. AI safety is a tiny fraction of RAND's work. The institution's primary incentive is to serve the federal government on defense, education, health, and dozens of other topics. AI safety could be deprioritized without affecting RAND's core business.

7. Lennart Heim departure. The loss of RAND's compute governance lead, without a public successor announcement, raises questions about the sustainability of this research line.

Cross-References

Complements: METR (joint Canary project for AI evaluations), CSET Georgetown (Matheny's prior org, similar tech policy focus), GovAI Oxford (Heim's prior org, compute governance overlap).

Overlaps with: Centre for the Governance of AI (compute governance), CSET (emerging technology policy), Center for AI Safety (evaluations), UK AISI (model evaluations).

Distinct from: MIRI/ARC (technical alignment -- RAND doesn't do this), Anthropic/OpenAI safety teams (lab-internal safety research), PauseAI (activist advocacy).

Tension with: RAND's defense establishment position may conflict with organizations advocating for international AI governance (which could constrain US capabilities). RAND's national security frame implicitly treats AI safety as a US competitive advantage issue.

What Would Change This Assessment

• Evidence that GER division budget exceeds $50M annually: Would indicate substantive institutional commitment beyond Matheny's personal project.
• RAND producing AI safety analysis that contradicts Open Phil priorities: Would demonstrate genuine independence. The bio-risk study partially does this.
• Matheny departure and AI safety work continuing: Would demonstrate institutional commitment beyond one person.
• RAND publishing work on international AI governance that constrains US capabilities: Would demonstrate willingness to prioritize global safety over US interests.
• AI safety community engagement with RAND's work on LW/EA Forum: Would indicate RAND's analysis is taken seriously by the broader community.
• A post-Trump administration adopting RAND's AI governance framework: Would demonstrate durability of RAND's policy contributions.

Self-Critique

Weakest claim: My assessment of RAND's AI safety work as "a small island in a defense ocean" may understate the impact of Matheny's leadership. A visionary CEO can redirect a large institution in ways that are not proportional to budget allocation. The Open Phil funding, though small relative to RAND's total, may have catalytic effects that create much larger shifts in RAND's research agenda.

Potential bias: This analysis may be overly influenced by the conflict-of-interest narrative, which is well-documented but may not accurately predict research quality. RAND researchers may produce genuinely excellent work regardless of funding source, just as academic researchers produce important work despite being funded by interested parties.

Missing sources: I could not access rand.org (blocked), the Politico article (paywalled), or Matheny's Senate testimony (binary PDF). RAND's own descriptions of its AI safety work, research output lists, and staff pages are all missing from the evidence. The analysis is therefore biased toward external perceptions of RAND rather than RAND's own articulation.

What a thoughtful disagreer would say: "RAND is uniquely positioned to bring AI safety into the national security mainstream. The EA funding is a feature, not a bug -- it allows RAND to work on long-term risks that short-term government contracts won't fund. The conflicts of interest are manageable and RAND's bio-risk study proves the research can be independent. Dismissing RAND's AI safety work because it's 'only' a fraction of a $467M institution misses the point -- that fraction has access to policymakers no one else can reach."

Single weakest claim: That RAND's AI safety work is heavily dependent on Matheny personally. This may underestimate the degree to which the GER division has developed its own institutional momentum, hiring, and culture independent of the CEO.

What would most change my view: Seeing quantitative data on GER division headcount, budget, and growth over time. If the division is 200+ people and growing fast with dedicated institutional budget lines, the "small island" critique weakens substantially.