PauseAI

Stated Theory of Change

PauseAI's stated theory of change is unusually explicit: grassroots activism (protests, lobbying, public education) shifts the Overton window, builds political pressure, and ultimately produces an international treaty pausing frontier AI development. Enforcement relies on the centralized AI chip supply chain -- ASML, TSMC, Nvidia are chokepoints that could halt training runs globally.

The mechanism has two pathways: (1) traditional treaty-building at international summits (analogous to Montreal Protocol on CFCs, Protocol on Blinding Laser Weapons), or (2) unilateral supply-chain intervention where ASML or the Dutch government restricts lithography machine sales to customers that don't commit to a moratorium.

The pause is indefinite -- "until we know how to do it safely and keep it under democratic control." Founder Joep Meindertsma has explicitly acknowledged he cannot define "provably safe AI" or specify end conditions.

Revealed Theory of Change

PauseAI's actions reveal a theory of change that is subtly different from the stated one. Their most effective work has been targeted accountability campaigns (the 60-MP DeepMind letter) and direct political engagement (25 Congressional offices, Capitol Days of Action), not the protests that generate their media coverage. The protests function primarily as movement-building and recruitment tools -- they bring people in -- while the political lobbying does the actual policy work.

Maxime Fournes, PauseAI's CEO, candidly admitted at the London protest: "I don't think that the pressure on companies will ever work." His actual strategic bets are on whistleblower protections and stigmatizing AI work. This is a revealing divergence -- the CEO does not believe the most visible tactic (protests at company offices) will achieve its stated goal. The real theory of change is about building a sufficiently large movement that politicians feel pressure to act, regardless of what companies think.

Holly Elmore's trajectory reveals another dimension: the adversarial theory of change. She has effectively given up on working within the AI safety establishment and is betting that external political pressure will force regulation that insiders cannot achieve through collaboration. Her prediction that "the only meaningful technical safety work is going to come after capabilities are paused, with actual external regulatory power" is a radical bet against the entire inside-game approach.

Key Assumptions

1. International cooperation is achievable on AI governance.

• Evidence for: Montreal Protocol, blinding laser weapons ban, nuclear nonproliferation (imperfect but meaningful). Scott Alexander's recent piece ("Every Debate On Pausing AI") notes that even pause opponents often argue against a strawman (unilateral pause) rather than the bilateral approach PauseAI actually advocates.
• Evidence against: AI is far more economically valuable than CFCs or blinding lasers. Nora Belrose argues military advantage from AGI will dwarf chemical weapons. China question is real but more nuanced than "China won't agree" -- China has shown some regulatory interest.
• Is it testable? Yes -- ongoing diplomatic engagement at AI summits provides evidence. Each failed summit is weak negative evidence.
• If wrong: PauseAI becomes an awareness organization without a viable policy pathway, which may still have Overton-window-shifting value.

2. Public support for AI regulation translates into political action.

• Evidence for: Polls consistently show 60-80% support for slowing AI development across US, UK, Australia. PauseAI cites these extensively.
• Evidence against: Spencer Greenberg's objection to Joep: this isn't a voting priority. AI lobbying budgets exceed $100M. Most protest attendees describe low-urgency motivations ("Sometimes you don't have that much to do on a Saturday").
• Is it testable? Yes -- election cycles and legislative votes provide evidence.
• If wrong: PauseAI may build a large but politically impotent movement. Environmental movement analogy: broad support, limited legislative achievement in many countries.

3. A pause would actually improve outcomes (not make things worse).

• Evidence for: Buying time for alignment research, democratic deliberation, and international coordination seems valuable on face.
• Evidence against: Belrose's three arguments (feedback loop, hardware overhang, underground research) are substantive and largely unaddressed by PauseAI. Historical evidence from the "last pause" (pre-GPT-3 era): MIRI's theoretical research "utterly failed" by their own admission.
• Is it testable? Only through natural experiments or modeling.
• If wrong: PauseAI's entire mission is counterproductive. This is the most load-bearing assumption and the one with the least robust defense.

4. The supply chain is a viable enforcement mechanism.

• Evidence for: ASML has already blocked some exports for national security. Remote shutdown capabilities exist. The chip supply chain is genuinely centralized.
• Evidence against: Hardware improves -- GPU price-performance doubles every 2.5 years. Algorithmic efficiency improves. Eventually training on consumer hardware becomes feasible. The enforcement window is temporary.
• Is it testable? Partially -- trends in hardware costs and algorithmic efficiency can be tracked.
• If wrong: Enforcement becomes progressively harder over time, potentially requiring "increasingly draconian international pressure" (per Belrose).

Strengths

Genuine independence. Zero economic ties to AI labs. No compute credits, no career pipeline, no consulting revenue. PauseAI is one of the only AI safety organizations that can criticize any lab without financial consequence. This matters more than it appears -- the evals org critique (Alfour/PranavG) documents how lab dependency corrupts ostensibly independent organizations.

Message clarity. "Pause AI" is a two-word proposition anyone can understand. Compare with "responsible scaling policies" or "frontier model evaluations." PauseAI wins the simplicity contest by a wide margin.

Grassroots legitimacy. 600+ volunteers, 30+ chapters, 13+ countries, mostly unpaid. This is real decentralized organizing, not astroturf. The movement building has genuine momentum -- 7 people in Brussels (2023) to ~500 in London (2026).

Demonstrated political access. 60 UK MPs on a letter. 25 Congressional offices visited. Capitol Days of Action. For an organization with a total budget smaller than a single senior researcher's salary at Anthropic, the political access is remarkable.

Concrete supply-chain enforcement story. The ASML/TSMC/Nvidia chokepoint argument is specific, factual, and strategically novel. Most AI governance discourse is abstract; PauseAI points to actual bottlenecks.

Lean operations. ~EUR 715K total funding over 2+ years with 6 paid staff and 600+ volunteers. Cost per volunteer-hour is almost nothing. If impact scales with volunteer activity, the cost-effectiveness could be extraordinary.

Weaknesses and Risks

No answer to Belrose. The strongest technical argument against PauseAI's core proposal -- that a pause would degrade alignment research, create hardware overhang, and push capabilities underground -- has no comprehensive rebuttal. PauseAI's counterarguments page handles common objections but does not engage at the required depth.

Definitional crisis. Scott Alexander's analysis is damning: even among sympathetic experts, "participants couldn't agree on basics of what they meant by 'pause', whether it was possible, or whether it would make things better or worse." PauseAI has partially addressed this by specifying compute thresholds (>10^25 FLOPs, >10^12 parameters), but the end-condition ambiguity remains.

Extreme donor concentration. FLI at 59% of total funding. One funder not renewing is an existential threat. Diversification attempts have yielded only small individual donors and a few small grants.

Coalition dilution. The London protest demonstrated the tension between movement scale and message discipline. Coalition partners brought generic anti-tech-CEO, environmental, and anti-monopoly messaging that drowned out the x-risk focus. This is the classic grassroots dilemma: the coalition broad enough to matter may be too diffuse to achieve the specific policy ask.

No success metrics. For an organization rooted in EA's measurement culture, the absence of defined KPIs or success benchmarks is striking. How would PauseAI know if their approach is failing? What evidence would cause a strategic pivot?

Holly's adversarial posture. Her public antagonism toward EA and OP -- including specific, unverified allegations about OP serving Anthropic's interests -- risks alienating potential allies and funders. Even if her analysis is correct, the delivery may be counterproductive.

No research capability. PauseAI cannot produce the kind of technical analysis that would strengthen their policy proposals or rebut technical critics. They are entirely dependent on external research (FLI, MIRI, academics) to support their claims.

Cross-References

vs. ControlAI: More professionalized "inside game" approach. Similar policy goals but different tactics (slick campaigns vs. grassroots protests). PauseAI UK collaborates with ControlAI on some events. ControlAI also criticizes Open Philanthropy. Both sit outside the lab-aligned safety ecosystem.

vs. FLI: PauseAI's primary funder and spiritual progenitor (the FLI pause letter inspired PauseAI's founding). FLI does AI safety index scoring, advocacy, and policy -- PauseAI is the grassroots complement to FLI's institutional approach.

vs. MIRI: Shares the "doom if we build it" framing. MIRI is technical research (with acknowledged failure), PauseAI is advocacy. Complementary missions -- PauseAI explicitly frames its role as "buying time for safety researchers."

vs. Evals orgs (Apollo, METR): PauseAI represents the opposite end of the independence spectrum. Evals orgs maintain lab relationships and API access; PauseAI has no lab ties at all. Alfour/PranavG's argument that evals are captured and PauseAI is the genuine alternative represents a real faultline in the AI safety community.

vs. AI labs' safety teams: Holly Elmore's critique -- that safety teams at labs have been "captured" and that "the only meaningful technical safety work is going to come after capabilities are paused" -- is PauseAI's most adversarial cross-reference. If she's right, internal safety work is largely theater. If she's wrong, PauseAI is undermining the people doing the most important work.

What Would Change This Assessment

Upward: A concrete policy win -- legislation passed in any major jurisdiction that PauseAI can credibly claim credit for. A successful international negotiation on AI compute governance. A major AI incident that validates the urgency argument and creates political window. Diversification of funding base beyond FLI dependence.

Downward: FLI defunding. A comprehensive empirical demonstration that alignment is on track (which would undermine the need for a pause). Evidence that protests are backfiring by associating x-risk concerns with fringe movements. Internal fracture between PauseAI Global and PauseAI US. An incident of violence or extremism attributed to PauseAI despite their non-violence commitment.

Fundamental question: Does grassroots advocacy work for highly technical policy questions? The environmental movement analogy is instructive -- broad public support, massive protests, limited legislative achievement over decades. PauseAI may be playing the same role for AI risk that early climate activists played in the 1990s: important for awareness, unclear on policy impact.

Self-Critique

Weakest claim: My assessment that PauseAI's political access is "remarkable for their budget" may overweight the 60-MP letter. Getting politicians to sign letters is far easier than getting them to pass legislation. The letter-to-legislation gap is enormous, and PauseAI has not yet demonstrated ability to cross it.

Potential bias: I may be giving too much weight to PauseAI's self-reported growth narrative. The 500-person London protest number comes from PauseAI/Pull the Plug; the MIT Tech Review reporter described "a couple of hundred." The actual attendance may be lower than claimed.

What a thoughtful critic would say: "PauseAI is solving for the wrong variable. The binding constraint on AI safety is not public opinion (which already favors regulation) or political awareness (which is growing organically). It's the technical question of how to make AI safe. Every dollar spent on protest banners is a dollar not spent on alignment research. And the pause itself, even if achieved, might make the technical problem harder by breaking the alignment feedback loop."

What I missed: I have limited insight into the internal dynamics of the Dutch board, the real nature of the FLI-PauseAI relationship, or whether there are unreported tensions between PauseAI Global and PauseAI US. Holly's public statements may not represent the median PauseAI member's views. I also lack evidence on whether PauseAI's messaging actually changes minds (vs. preaching to the converted), which is the key question for their theory of change.

Single weakest claim: That PauseAI's protest-to-policy pipeline has any evidence of working beyond Overton-window-shifting. The 60-MP letter is suggestive but not causal. No AI pause legislation exists anywhere because of PauseAI's work.