Simon Institute for Longterm Governance

Theory of Change

SI's theory of change is facilitation, not research. From their Year One Review: "Unlike many think tanks, SI does not focus its theory of change around producing in-house policy analyses or recommendations. Instead, SI works with policymakers and academics to efficiently translate research into practice." Their bet is that the binding constraint on international AI governance is not lack of knowledge but lack of translation -- diplomats don't understand frontier AI, and AI researchers don't understand the UN system.

The causal chain: (1) Train diplomats on frontier AI capabilities and risks, (2) facilitate exchange between AI researchers and policymakers, (3) provide institutional design expertise for multilateral AI governance bodies, (4) these bodies produce shared risk assessments, (5) shared assessments create the foundation for coordinated governance action.

In Konrad Seifert's words: "The limiting factor to AI regulation and international agreements is not that people aren't worried. It's that a lot of key people don't have clear models of threats and solutions. Detailed scenarios are needed, but urgency isn't helpful. The hard-earned trust that creates opportunity for in-depth exchange comes through patience and immediate usefulness."

What They Do

SI operates at the intersection of the AI safety community and the UN system. Their primary activities:

Institutional design for the UN Scientific Panel on AI. SI published one of the first reports on design options for the Independent International Scientific Panel on AI and the Global Dialogue on AI Governance (the "Blueprints" report, Dec 2024), followed by full recommendations (Feb 2025). The UNGA adopted the Panel modalities in August 2025 by consensus of 190+ member states. SI co-authored the major Carnegie Endowment paper (Aug 2024, 25 co-authors) that provided the intellectual framework for the two-track approach adopted.

Diplomat training. In 2024, SI trained 125+ diplomats from 50+ UN missions across five training programs, including two full-day courses in NYC (55% Global South, 9.1/10 rating). Topics covered: AI technical foundations, risks and opportunities, governance landscape, compute governance (with Lennart Heim and Robert Trager as speakers).

Policy convening. 800+ calls in 2024, meetings with 50+ missions, a 2-day Paris workshop ahead of the AI Action Summit with 25 experts, co-hosted Windfall Trust benefit-sharing workshop with FLI. Maintains relationships in both Geneva (permanent missions) and NYC (UN headquarters).

Earlier work (pre-2023 pivot). Commissioned by UNDRR to write the first UN thematic study on existential risk and rapid technological change. Organized 11 events on engineered pandemics at the BWC. Contributed to the Sendai Framework mid-term review (cited 6 times).

Research output. Published in TechPolicy.Press on AI benefit-sharing politics. Produced policy submissions to the Global Digital Compact. Published responses to drafts of Panel modalities. This is tactical research -- published to influence specific policy processes, not for academic credit.

Key People

Maxime (Max) Stauffer, Co-founder and CEO. BA International Relations, MSc Computational Mathematics (Johns Hopkins, part-time). Co-founded EA Geneva. Senior Science-Policy Officer at Geneva Science-Policy Interface. WEF Global Future Council member. Provides the institutional credentials and UN system connections.

Konrad Seifert, Co-founder and CSO. University dropout who discovered EA in 2013, built EA Switzerland (budget ~CHF 500K), dropped out of undergrad in 2016 after getting funded for community building. His Substack writing reveals the most candid thinking about SI's approach -- sophisticated understanding of institutional psychology and "diplomatic entrepreneurship."

Dr. Renata Dwan, Director of Tech Diplomacy. Former Director of UNIDIR, former Special Adviser to UN Secretary-General's Tech Envoy (led GDC negotiation support). 25+ years in international affairs. Her hire in 2024 dramatically elevated SI's access and credibility within the UN system.

Team is ~8-10 staff. Notable departures: Belinda Cleeland (Director of Policy) and Sofia Mikton (Operations) have left without public explanation.

Money and Incentives

Total funding trajectory:

Year 1 (2021): CHF ~144,000 (~$156K)
First 2 years cumulative: $1.5M
2024: ~CHF 1,456,000 secured
3-year budget target (2023-2025): CHF 3.4M-5.5M
Room for more funding: $2.6M-5.2M

Revenue breakdown:

Open Philanthropy / Coefficient Giving: $1,461,194 over 2 years (April 2024) -- by far the largest single source
EA Infrastructure Fund: $46,000 (2021, of $600K requested)
Swiss government: CHF 50,000 contract (international law project)
Geneva Science-Policy Interface: CHF 30,000 collaboration grant
UNDRR: Contracted for thematic study (amount unknown)
Private donors: CHF ~68,000 (2021)

Funder concentration risk: The Open Phil grant (~$730K/year) likely represents the majority of SI's annual budget. If not renewed, SI faces an existential funding challenge. Early funders were skeptical: SFF rejected, EA Funds gave 1/12th of ask, Founders Pledge declined to recommend citing "insufficient track record."

Business model: Pure grant-funded nonprofit. No AI lab funding, no compute dependencies, no product revenue, no equity investors. Government contract revenue (Swiss gov, UNDRR) provides partial diversification but is a small fraction.

Incentive structure: Unusually clean. SI has no financial ties to any AI lab. Their sole funder incentive pressure is from philanthropic foundations that want to see effective multilateral AI governance. The main incentive risk is time-horizon mismatch: institutional building takes decades, but funders evaluate on 2-year grant cycles.

Cost efficiency: $150K annual salary for a Geneva-based AI governance expert. A briefing event costs $5K-$20K. Lean by any standard.

What Others Say

No SI-specific critics exist. The org is too small and niche to have attracted dedicated critical attention. The absence of criticism can be read as either genuine quality or insufficient prominence.

The strongest structural critiques come from the broader debate about multilateral AI governance:

US withdrawal undermines SI's work. CSIS (July 2025) reports the Trump administration explicitly rejected multilateral AI governance. OSTP Director Kratsios called existential risk an "ideological fixation" and the US was absent from the Global Dialogue launch. "Washington's rejection substantially reduces the dialogue's near-term relevance, since meaningful global governance is difficult without the participation of the leading technology holder." Meanwhile, China enthusiastically supports the UN process -- raising the risk that the Panel is perceived as a Chinese-aligned institution.

Longtermism challenges. AlgorithmWatch (Sept 2025) argues the longtermist framing of AI risk "directs attention onto imagined catastrophes with large hypothetical numbers, away from harder-to-quantify but immediate impacts." Carnegie Council (Sept 2022) warns longtermism "risks becoming a Trojan horse for the vested interests of a select few." SI has tried to bridge the ethics-safety divide with a blog post arguing both fields share common ground (interpretability, accountability), but the tension between existential risk focus and present-harm accountability remains.

Summit pageantry concern. Carnegie Endowment (Oct 2024) describes international AI governance as "a form of governance spectacle where symbolic gestures obscure the pressing need for cohesive, enforceable global rules." The Scientific Panel is an assessment body, not an enforcement mechanism. If the broader architecture remains symbolic, SI's institutional design work contributes to a growing stack of non-binding norms.

In SI's defense: They are unusually self-aware about downside risks. The Year One Review documents specific failure modes and mitigation strategies. Their facilitation-only approach (no policy recommendations) is explicitly designed to minimize the risk of bad advocacy.

What's Absent

No independent evaluation of SI's impact after 5 years of operation. Self-reported impact claims are the only data.
Board composition largely unknown (Notion page not fetchable; only one confirmed member: Igor Linkov).
Detailed financials not publicly accessible (PDF reports on Google Drive, not in a searchable database).
No articulation of how the Scientific Panel actually changes AI lab behavior -- the causal chain from "assessment" to "changed practices" is unexplained.
No long-form founder interviews (no 80K Hours, Dwarkesh, or equivalent podcast appearances).
Staff departures (Cleeland, Mikton) unexplained.
Chinese engagement details limited despite it being a stated priority.