Center for Security and Emerging Technology (CSET)

Stated Theory of Change

CSET's stated theory of change operates through two mechanisms:

Mechanism 1: Information provision. Produce data-driven, nonpartisan research that helps policymakers make better decisions about AI and emerging technology. The causal chain: CSET research -> better-informed policymakers -> better government decisions about AI -> reduced risk from AI. Open Philanthropy's grant writeup makes this explicit: ensuring "high-quality and well-informed advice to policymakers over the long run is one of the most promising ways to increase the benefits and reduce the risks from advanced AI."

Mechanism 2: Talent pipeline. Train people at the intersection of technology and national security, who then take government positions where they shape policy from the inside. The causal chain: CSET trains researchers -> alumni enter government -> alumni carry CSET's analytical frameworks into policy roles -> better decisions.

The stated ultimate goal connects to OP's focus area: reducing "potential risks from advanced artificial intelligence." The link is: if policymakers are well-informed about AI capabilities and risks, they are less likely to make catastrophic mistakes (arms race dynamics, premature deployment, inadequate safety requirements, miscalculation in US-China competition).

Revealed Theory of Change

What CSET's actions reveal about its actual priorities:

Revealed priority 1: US-China tech competition. The overwhelming majority of CSET's output concerns how the US should position itself relative to China on AI and semiconductors. The talent pipeline has primarily placed people in positions focused on export controls, semiconductor policy, and China-facing intelligence. Ben Buchanan's "AI Triad" framework explicitly argues for controlling compute as a strategic lever. This is not "reducing AI risk" in the abstract -- it is "ensuring US advantage in AI" as a means of security.

Revealed priority 2: Near-term policy influence over long-term risk reduction. Toner in 2019: "We are focused on producing work that is going to be relevant and useful to decisions that are coming up in the near term. It's not really something that's in our wheelhouse" (referring to general AI/AGI scenarios). While Toner is now pivoting toward frontier AI topics, for most of its existence CSET has deliberately avoided the kind of AI safety research that OP's stated concern (advanced AI risk) would most directly address.

Revealed priority 3: Institutional prestige and access. CSET's decisions about location (Georgetown, near Capitol Hill), hiring (intelligence community backgrounds), research topics (what policymakers ask about), and communication style (nonpartisan, data-driven) all optimize for credibility with the national security establishment. This is rational if your theory of change runs through government influence, but it creates distance from the AI safety community that funded the org.

The divergence: OP funded CSET to reduce risks from advanced AI. CSET has primarily worked on near-term technology competition with China. These overlap -- if you believe US-China miscalculation is an AI risk vector -- but they are not the same thing. The vast majority of CSET's output would be identical even if OP's stated concern about advanced AI risk did not exist. Export control analysis, semiconductor supply chain mapping, and PLA procurement tracking are intrinsically valuable national security research regardless of whether you believe AGI is coming.

Key Assumptions

Assumption 1: Policy influence is tractable and beneficial.

• Evidence for: Multiple CSET alumni in senior government roles. Toner's concrete examples of policy influence (outbound investment controls, DoD AI principles). Think tanks demonstrably shape DC discourse.
• Evidence against: Attribution is nearly impossible. The export controls would likely have happened with or without CSET. No independent impact evaluation exists. Founders Pledge notes "limited track record."
• Testable? Partially -- you can trace personnel flows and policy language, but counterfactual impact is inherently hard to measure.
• If wrong: CSET's $112M produced research that policymakers would have reached anyway through other channels, and the talent pipeline merely branded people who would have entered government regardless.

Assumption 2: US advantage in AI reduces catastrophic risk.

• Evidence for: If the US maintains AI leadership, Western democratic norms have more influence on AI governance. US labs (currently) invest more in safety research than Chinese counterparts. US government has stronger safety regulatory infrastructure.
• Evidence against: The Jacobin critique -- aggressive export controls may accelerate Chinese self-sufficiency, reduce cooperation potential, and create arms race dynamics. Katja Grace (MIRI): "The best individual action could be to move slowly and cautiously." Export controls could push China toward less safe, less transparent AI development.
• Testable? Only over decades. The counterfactual (what would Chinese AI development look like without export controls?) is unknowable.
• If wrong: CSET's most celebrated policy achievements (export controls) may have increased rather than decreased catastrophic AI risk by intensifying great-power competition.

Assumption 3: The national security frame is the right lens for AI risk.

• Evidence for: Governments control the regulatory environment. Military applications of AI are among the highest-risk use cases. The national security establishment has resources and urgency.
• Evidence against: Framing AI as a national security issue may securitize it in ways that reduce transparency, increase arms race dynamics, and exclude civil society voices. The AI safety community's concerns (alignment, loss of control) don't map cleanly onto national security frameworks.
• If wrong: CSET's entire institutional positioning -- at a foreign service school, staffed by intelligence community alumni, focused on competition with China -- would be an elaborate misdirection from the actual sources of AI risk.

Assumption 4: A single-funder dependency doesn't distort research.

• Evidence for: CSET's published research covers topics OP doesn't specifically prioritize (immigration, DoD procurement, cybersecurity). Baker's statement about "independent discretion." Research quality is generally respected.
• Evidence against: 94% funding from one source is extreme by any standard. OP's broader network has structural ties to AI labs (Anthropic especially). The organizational incentive to maintain OP funding may create subtle biases even without explicit direction.
• If wrong: CSET's research agenda is shaped by OP's worldview in ways that are invisible to outsiders, and the structural conflicts (OP-Anthropic ties) compromise CSET's ability to scrutinize the network it's embedded in.

Strengths

Genuine analytical depth. CSET's research is substantively strong. The Huawei chip analysis, PLA procurement data, and entity resolution work demonstrate technical sophistication that justifies the think tank model. ChinaTalk's "Think Tank of the Year" award reflects real peer recognition.

Extraordinary talent pipeline. No other AI governance organization has placed as many alumni in senior government positions. The CSET-to-government revolving door is the clearest mechanism through which the organization has real-world impact.

Data-driven differentiation. The in-house data science team genuinely distinguishes CSET from opinion-driven think tanks. The ETO platform creates public goods that other researchers use.

Helen Toner's unique positioning. Toner's combination of AI safety community background, OP experience, OpenAI board membership, national security establishment credibility, and direct experience with corporate AI governance failure makes her arguably the most credibly positioned person to lead an AI policy think tank in 2025-2026.

Institutional prestige and access. Georgetown affiliation, advisory board composition, and alumni network give CSET access to decision-makers that newer or less established organizations cannot match.

Weaknesses and Risks

Funding cliff. OP grants expired in 2025 with no public renewal. The apparent 70% staff reduction suggests this is already biting. If OP does not renew, CSET's existence is threatened. Diversification efforts (Google.org, Hewlett, donations) are orders of magnitude too small to replace OP funding.

Structural conflicts in the funder network. The OP-Anthropic-CSET relationships are not smoke-without-fire: the same people appear repeatedly across the network (Muehlhauser on Anthropic board and as CSET grant investigator; Matheny on Anthropic LTBT and as CSET founder; Meyer on CSET advisory board and at Anthropic). This doesn't prove corruption, but it creates the appearance and structural possibility of coordinated influence.

The near-term/long-term gap. $105M from a funder concerned about advanced AI risk produced an organization focused overwhelmingly on near-term technology competition. The AI safety community -- the people most concerned about the risks OP cited -- largely ignores CSET. This disconnect is striking and may indicate that OP's theory of change (policy influence reduces AI risk) is disconnected from the technical AI safety community's theory of change (alignment research reduces AI risk).

China framing risks. If the hawkish critics are right that export controls accelerate arms race dynamics and reduce cooperation, CSET's most celebrated policy achievements may be net negative for reducing catastrophic AI risk. This is the single hardest question about CSET's impact.

Georgetown institutional constraints. The BGI incident demonstrated that Georgetown's institutional risk tolerance can constrain CSET's research. A think tank that cannot defend its researchers' publications on politically sensitive topics has a fundamental credibility problem.

Opacity of impact. CSET claims significant policy influence but has no independent evaluation. The attribution problem is real -- you cannot prove that export controls would have been different without CSET -- but the absence of any rigorous impact assessment after $112M in funding is notable.

Cross-References

GovAI (Oxford): The closest peer organization. More focused on long-term AI governance theory, less on near-term US policy. Toner came to CSET via GovAI as a research affiliate. Complementary rather than competing.

RAND Corporation: CSET was explicitly modeled on early RAND. Matheny now leads RAND. RAND also receives OP funding for AI risk research. The organizations are now in the same network, with the same founder leading RAND and sitting on Anthropic's LTBT.

IAPS (Princeton): Newer peer institution. No direct comparison found.

Anthropic: Not a peer but a structural partner through OP. CSET advisory board member Joel Meyer works at Anthropic. CSET's recent "Beyond Corporate Promises" publication references Anthropic's RSP directly. The relationship is not adversarial.

METR: Referenced in CSET's recent publications as a third-party evaluation organization. Complementary -- METR does technical evaluations, CSET does policy framing.

Open Philanthropy: The funder-grantee relationship dominates CSET's financial structure. OP's broader AI portfolio (Anthropic investment, RAND grants, CNAS funding, Horizon fellowship) creates a network that CSET is embedded within.

What Would Change This Assessment

• OP renewal announcement with diversified funding targets: Would resolve the existential funding question and demonstrate continued funder confidence.
• Independent impact evaluation: A rigorous assessment of whether CSET's policy influence was net positive would dramatically clarify the theory of change.
• Chinese perspective on export controls: Evidence that export controls either (a) significantly slowed Chinese AI capabilities or (b) accelerated Chinese self-sufficiency would shift the assessment of CSET's signature policy achievement.
• Toner's frontier AI pivot producing substantive output: If CSET publishes research that the technical AI safety community finds valuable, the near-term/long-term gap narrows.
• Major staff departure with public criticism: Would reveal internal dynamics currently invisible.
• Evidence that OP directed research priorities: Would confirm the structural conflict critique.

Self-Critique

Weakest element of this analysis: The assessment of whether CSET's policy influence has been net positive for reducing AI risk. I have strong evidence on funding, structure, and personnel, but the core question -- did the export controls reduce or increase catastrophic AI risk? -- requires geopolitical analysis far beyond what the evidence base supports. My assessment that this is "uncertain" may be too agnostic; a more opinionated analyst might argue the answer is clearly yes or clearly no.

Potential bias: I may be over-indexing on the structural conflict critique (OP-Anthropic ties) because it makes for compelling analysis. In reality, many DC think tanks have tangled funding relationships, and the question is whether CSET's research quality is compromised, not just whether the relationships exist. The research quality appears genuinely high.

What a thoughtful disagreer would say: "CSET is exactly the kind of organization the AI safety community should be celebrating. It took AI risk concerns out of the ivory tower and into the rooms where real decisions are made. The export controls weren't perfect, but they bought time. The talent pipeline is the most concrete AI governance success story in existence. The structural conflicts are standard DC practice and don't compromise research quality. Critics like Politico are holding CSET to a standard that no think tank meets."

What I may have missed: CSET's classified or non-public work. Think tanks often have their most important influence through private briefings, not published papers. The evidence base is necessarily biased toward public outputs, and the most impactful work may be invisible.

Information that would most change my view: A rigorous, independent assessment of whether CSET alumni in government positions made materially different decisions than non-CSET counterfactuals would have. If the answer is yes, CSET's theory of change is validated. If the answer is no, $112M was spent on an elaborate credentialing exercise.