ControlAI

Stated Theory of Change

ControlAI claims the path to reducing AI risk runs through democratic institutions: inform every relevant political actor about the extinction risk from superintelligence, build political will, pass national legislation banning ASI development, and ultimately establish an international non-proliferation regime.

The chain is: awareness -> political will -> national legislation -> international coordination -> enforced prohibition on ASI.

This is explicitly contrasted with three alternatives they consider inadequate: (1) technical AI safety research ("alignment is way too hard, maybe impossible" -- Leahy), (2) model evaluations without enforcement ("detection without enforcement is insufficient"), and (3) insider influence strategies ("undemocratic... centralizes control in the hands of a small group").

The key intellectual move is treating ASI development as a collective action problem where defection doesn't benefit anyone (unlike a true prisoner's dilemma). If building ASI kills even the builder, then the problem is informational: people don't understand the risks, rather than motivational: people understand but have incentives to defect. This justifies focusing on education/awareness rather than incentive design.

Revealed Theory of Change

ControlAI's actions strongly match their stated theory. They spend virtually all their resources on:

• Briefing lawmakers (279 and counting across 4 countries)
• Public engagement tools (160K+ messages to lawmakers)
• Media and content creator partnerships (150M+ views)
• Academic publications framing the debate (Three Doctrines, Middle Powers papers)
• Podcast amplifying insider whistleblowers (Adler ex-OpenAI episode)

There is no meaningful gap between stated and revealed theory of change. They say they do advocacy and they do advocacy. The closest thing to a divergence is their earlier campaigns (blimp stunt at AI Safety Summit, deepfakes campaign) which were more attention-grabbing and less institutional -- suggesting the DIP represents a genuine strategic maturation.

One subtle tension: they position themselves as democratic and transparent, but their own organizational governance is opaque. They criticize Open Philanthropy for unaccountable power concentration while their own board structure, funding sources, and decision-making processes are undisclosed.

Key Assumptions

1. Superintelligence, if built, will escape human control and likely cause extinction.

• Evidence for: CAIS statement signed by lab CEOs, P(doom) estimates from Hinton (50%), Hendrycks (80%), Yudkowsky (95%+). ControlAI's own Three Doctrines paper surveys the expert landscape.
• Evidence against: Dean Ball (0.01%), replacement doctrine proponents. No empirical evidence of AI systems causing catastrophic harm. GPT-5 underperformance suggests the curve may flatten.
• Testable? Partially -- if systems demonstrate concerning autonomous capabilities, this supports the assumption. But the assumption is about a system that doesn't yet exist.
• What changes if wrong: If ASI can be built safely, ControlAI's entire raison d'etre dissolves. If P(doom) is genuinely near zero, their campaign is harmful by slowing beneficial AI development.

2. The problem is primarily political, not technical.

• Evidence for: Voluntary commitments have failed (Anthropic dropping RSP pledges). Companies race despite acknowledging risks. No binding regulation exists.
• Evidence against: Technical alignment research continues to advance (circuit breakers, oversight methods). Some argue regulation is premature without better understanding of what to regulate.
• Testable? If a major alignment breakthrough occurs that demonstrably solves control, this weakens the case.
• What changes if wrong: If technical solutions are sufficient, the advocacy approach is at best unnecessary and at worst harmful distraction.

3. Democratic institutions can act fast enough.

• Evidence for: Nuclear non-proliferation was achieved. 110+ UK parliamentarians engaged in 18 months. Polls show 73% UK public favor halting fast ASI development.
• Evidence against: AI development timelines may be 3-5 years. Legislative processes take years. The US political environment is hostile (Sacks, Krishnan). Tech lobbying vastly outguns ControlAI ($200M super PAC vs. $42.5K lobbying spend).
• Testable? If AGI arrives before any binding legislation passes, this assumption is falsified.
• What changes if wrong: If democratic processes are too slow, more extreme interventions (compute controls, international enforcement mechanisms) become necessary -- or the cause is simply lost.

4. Banning ASI development is both achievable and beneficial.

• Evidence for: We ban biological weapons, restrict nuclear technology, regulate gain-of-function research. Polls show strong public support.
• Evidence against: Software is harder to regulate than physical materials. Defining ASI in law may be impossible without over-breadth or under-breadth. Enforcement requires unprecedented international coordination. A ban could drive development underground or to less accountable actors.
• Testable? Only by attempting it. Partial tests: compute governance proposals, chip export controls.
• What changes if wrong: If a ban is unenforceable, ControlAI's entire strategy collapses and resources would be better spent on alignment research or other approaches.

Strengths

Demonstrated political access. 279 lawmakers across 4 countries is not a vanity metric -- it represents real institutional contact at a scale no other AI safety org has achieved in legislative engagement. The 110+ UK parliamentary coalition is unprecedented.

Operational competence. The Leticia Garcia Martinez accounts reveal a sophisticated understanding of legislative engagement: cold outreach, non-linear growth dynamics, bipartisan framing, tactical messaging. The conversion rate (1 in 2 briefed lawmakers become supporters) suggests genuine persuasive capacity.

Strategic clarity. ControlAI knows what it is and isn't. They don't try to do technical research, evaluations, AND advocacy. The DIP is a focused strategy with clear metrics.

Bipartisan and international credibility. The ban statement's signatories span from Steve Bannon to Bernie Sanders, from Prince Harry to Steve Wozniak. This makes it very hard to dismiss as partisan.

Intellectual honesty about worldview. The Three Doctrines paper, the DIP design document, and Leahy's retrospective are remarkably transparent about their beliefs and how they arrived at them. They don't hide behind technicality -- they say "ban superintelligence" plainly.

Speed and efficiency. A 15-person team achieving these results in 2.5 years is exceptional by nonprofit standards.

Weaknesses and Risks

Funding opacity is a serious problem. An organization demanding democratic accountability from AI labs while having zero financial transparency faces a legitimate hypocrisy charge. We cannot assess whether their advocacy is influenced by donors with competing interests.

The definition problem is unresolved. Dean Ball's core objection -- "how do you define superintelligence in law without either banning beneficial AI or creating a toothless law" -- is substantive and ControlAI has not published a compelling response. Tegmark's "define harms not capabilities" is promising but ControlAI hasn't developed it into statutory language.

The US political environment is hostile. The Washington "vibe shift" (Sacks/Krishnan/super PACs) means ControlAI's US expansion faces dramatically harder conditions than the UK campaign. Congressional engagement may require different tactics than parliamentary briefings.

Conjecture monoculture. Nearly all intellectual leadership comes from Conjecture alumni (Miotti, Leahy, Alfour, Shimi). This creates groupthink risk. Leahy's anti-EA rhetoric and his management track record at Conjecture (failed company, "fire more" as top lesson) are liabilities.

No legislative outcomes yet. 279 lawmakers briefed is impressive advocacy output, but zero laws have been passed. The gap between "awareness" and "legislation" is where most advocacy campaigns fail.

Credibility risk from timeline claims. If AGI does not materialize on predicted timelines, ControlAI's urgency narrative loses force. Helen Toner's "boy-who-cried-wolf" warning applies directly.

No contingency plan. If the ban strategy fails -- if ASI development proceeds despite advocacy -- ControlAI has invested everything in one approach. There is no Plan B.

Cross-References

Complementary to: FLI (co-organized ban statement), MIRI (Leahy and Bourgon both testified in Canada), PauseAI (grassroots arm vs. ControlAI's insider game). The ecosystem needs both grassroots and institutional advocacy.

Competing with: Open Philanthropy / Coefficient Giving (explicitly criticized), METR / Apollo Research (evals approach attacked), GovAI and other governance research organizations (dismissed as insufficient without enforcement).

Differentiated from: CAIS (statement but no advocacy machinery), FLI (broader mandate, less focused), PauseAI (grassroots vs. institutional), Stop AI (radical/confrontational vs. professional).

Interesting comparison with PauseAI: Both want to slow/stop AI development. PauseAI is grassroots, protest-oriented, volunteer-driven. ControlAI is professional, insider-oriented, and institutional. Transformer News describes ControlAI as "the most professionalized" of activist groups. The strategies are complementary -- PauseAI creates bottom-up pressure while ControlAI works top-down.

What Would Change This Assessment

• Disclosure of funding sources revealing major conflicts of interest (e.g., funded by a competing AI lab, or by individuals with strong financial positions affected by regulation) would significantly downgrade assessment.
• Passage of binding legislation in any G7 country influenced by their advocacy would significantly upgrade assessment. This is the key test.
• AGI/ASI development proceeding despite or before their advocacy achieves legislative results would reveal whether the approach was too slow.
• A major alignment breakthrough that demonstrably solves the control problem would undermine the core premise that technical solutions are insufficient.
• US political environment shift -- if the next administration is more receptive to regulation, ControlAI's US strategy becomes much more viable.
• Internal governance reforms -- publishing a board, financial reports, and conflict-of-interest policies would address the transparency criticism.

Self-Critique

What sources should I have checked but didn't?

• The full text of "A Narrow Path" (only the landing page was available, not the book itself).
• The Hansard transcripts of the two House of Lords debates (blocked by Hansard).
• The UK Parliament written evidence (RAI0031) from Miotti and Adler (403 error).
• The NITRD RFI response text (PDF was unreadable).
• More detail on the Apart Research red-team results.

Where is this analysis potentially biased?

• I may be too impressed by the output metrics (lawmakers briefed) without adequately weighting the lack of legislative outcomes. Impressive advocacy output does not equal impact.
• The funding opacity should perhaps be weighted more heavily as a red flag rather than simply noted as a gap.
• I may underweight the "AI ethics" critique that extinction risk framing distracts from present-day harms.

What would a thoughtful person who disagrees say? "ControlAI is a slick lobbying operation built on unfalsifiable doom claims, funded by unknown donors, with zero legislative achievements, and led by Conjecture refugees who failed at technical AI safety and pivoted to policy advocacy. They're selling fear to lawmakers who don't know enough to push back. Their nuclear analogy is fatally flawed because you can build nuclear weapons in secret but you need massive visible compute infrastructure for AI. And their own governance is less transparent than the companies they attack."

What's my single weakest claim? That ControlAI fills a "genuine gap" in the ecosystem. If the political pathway to banning ASI is infeasible (because of definitional problems, enforcement challenges, or political hostility), then the gap doesn't need filling -- the resources would be better spent elsewhere.

What information would most change my view? Disclosure of ControlAI's funding sources. If they are funded by one or two wealthy individuals with strong views, the organization's apparent independence is illusory. If they are funded by a broad base of small donors, their credibility increases substantially.