ControlAI

Theory of Change

ControlAI operates from what they call the "extinction doctrine": the belief that superintelligence, if developed, will escape human control and likely end human civilization. Their theory of change is that the problem is political, not technical: "Even if one company could figure out a way to make some systems controllable, nothing stops another one to rush ahead with systems that we cannot control. Ultimately, this is a deeply political problem that will be solved with policy solutions."

Their strategy is the Direct Institutional Plan (DIP): design policies that prevent ASI development, then "inform every relevant actor in the democratic process" and convince them to take a stance. The DIP is defined by four properties: strategic (addresses the full problem end-to-end), public (transparent about beliefs), scalable (compounds with more resources), and democratic (strengthens democratic institutions rather than concentrating power).

They explicitly frame this against other AI safety approaches. On evaluation organizations like METR: "Detection without enforcement is insufficient. There is nothing in this plan that forbids an AI company from ignoring the report." On Open Philanthropy's insider strategy: "By its very nature, the overall Open Philanthropy plan for AGI is undemocratic: it centralizes control over the main AGI projects in the hands of a small group of 'trusted' actors, where Open Philanthropy leadership is the sole arbiter of what constitutes 'trustworthiness.'"

The causal chain: ControlAI informs lawmakers and the public about ASI risks -> lawmakers publicly acknowledge the threat -> critical mass of political will forms -> national legislation bans ASI development -> international agreement enforces the ban, modeled on nuclear non-proliferation.

What They Do

ControlAI is fundamentally a legislative advocacy organization. Their primary output is briefing lawmakers and building parliamentary coalitions.

UK parliamentary campaign (flagship). Since September 2024, 140+ introductory briefings delivered to UK parliamentarians. 110+ now support their campaign statement acknowledging superintelligence as a national security threat. Roughly 1 in 2 briefed lawmakers become supporters. Two House of Lords debates on superintelligence resulted. The campaign reached a tipping point around 40-50 supporters, after which growth became non-linear and organic -- parliamentarians began referring colleagues without ControlAI prompting. An external analysis by Ryan Mizzen found the coalition larger than the combined membership of the UK's two climate APPGs, calling the conversion rate "staggering" given ControlAI had "no pre-existing contacts" in Parliament.

International expansion. In Canada: registered as lobbyist (Oct 2025), multiple parliamentary hearings, testimonies by Miotti, Buteau, and Leahy. Scaled from ~50 to 100+ lawmakers briefed in 2 months. In Germany: consulting program officer (Benjamin Balde), similarly rapid growth. In the US: 90+ congressional offices briefed, $42,500 federal lobbying spend in 2024, Connor Leahy moving to DC as Director of ControlAI US.

Ban Superintelligence statement (Oct 2025, co-organized with FLI). 850+ signatories including Hinton, Bengio, Wozniak, Prince Harry, Steve Bannon, Bernie Sanders. Covered by TIME, CNBC, CBS, Guardian. Broadly bipartisan and international.

Public engagement infrastructure. 160,000+ messages sent to lawmakers through their contact tools. 185,000+ signatures on public statement. 14 content creator partnerships totaling 20M+ subscribers and 150M+ views (including Rational Animations, 1.4M views on one video).

Earlier campaigns. Blimp over AI Safety Summit (Oct 2023), campaign against RSP endorsement, EU AI Act foundation model exemptions (Nov-Dec 2023), deepfakes policy campaign (Dec 2023-Jun 2024).

Research and publications. "A Narrow Path" (3-phase policy framework, red-teamed by Apart Research). "Three Main Doctrines on the Future of AI" (academic paper on SSRN). "How Middle Powers May Prevent ASI Development" (geopolitical strategy paper). The Compendium (co-authored with Conjecture team). Weekly Substack newsletter and podcast.

Key People

Andrea Miotti -- Founder and CEO. Italian, born 1996. Previously at Conjecture. Young (29) but has achieved remarkable political access: briefed UK PM's office, testified before Canadian Parliament, appeared on BBC, CNN, TIME, Guardian, Breaking Points.

Connor Leahy -- Advisor, incoming Director of ControlAI US (2026). Co-founded EleutherAI, was CEO of Conjecture (winding down). Describes having "stopped believing in technical AI safety" and that "humanity is bottlenecked far more on institutional malaise... than it is on technical research." Major media profile. Controversial: called EA "transhumanist weirdo cultists."

Leticia Garcia Martinez -- UK Parliamentary Engagement Lead. Economist, MA in Philosophy and Public Policy from LSE. Personally delivered 140+ parliamentary briefings and is responsible for the operational methodology that produced the 110+ supporter coalition.

The team is ~15 people, with Conjecture alumni in multiple positions (Miotti, Shimi, and now Leahy). Gabriel Alfour (CTO Conjecture) is an advisor and key intellectual contributor.

Money and Incentives

Total budget: Unknown. This is the most significant gap in the evidence. ControlAI has zero disclosed funding sources.

Revenue breakdown: Entirely opaque. No Coefficient Giving/Open Phil grants. No documented SFF grants. No 990 filings (501(c)(4) entities have minimal disclosure requirements). First UK accounts not due until September 2026. The 501(c)(4) structure in the US allows political lobbying but does not require donor disclosure. In the UK, the company limited by guarantee structure similarly offers limited transparency.

What we can estimate: A 15-person team with London salaries (one job listing shows 60-80K GBP range) and offices 5 minutes from London Bridge, catered lunches, pension contributions, visa sponsorship for exceptional candidates, and operations across 4 countries suggests an annual budget of at minimum 1-1.5M GBP. The $42,500 US lobbying spend is a data point but likely a small fraction of total spending.

Conjecture connection. Miotti was previously at Conjecture. Alfour "helped Andrea found ControlAI." Multiple Conjecture staff have moved to ControlAI. Conjecture "raised significant amounts of capital." Whether Conjecture's investors also fund ControlAI is unknown but the deep personnel overlap makes some connection plausible.

Business model. Appears to be individual donations and possibly foundation grants, but specifics are unknown. The impact report ends with "If you are a donor or partner who wants to help build the coalition that keeps humanity in control, please get in touch at partners@controlai.com."

Incentive analysis. ControlAI has explicitly criticized Open Philanthropy (the largest AI safety funder), likely forfeiting access to that funding. They receive no documented lab funding and are adversarial toward all frontier labs. They have no product revenue. They don't accept compute credits. Their financial incentives appear aligned with their stated mission (no lab money warping their views), but the opacity makes this impossible to verify. The 501(c)(4) structure, which sacrifices tax-deductibility for donors in exchange for political freedom and privacy, suggests they prioritize lobbying flexibility over fundraising ease.

What Others Say

Dean Ball (former White House AI policy advisor): The ban is "counterproductive and silly." You cannot define superintelligence in statute without banning beneficial systems. "How do you prove safety without building it?" Any ban creates a government monopoly on research. International enforcement is unrealistic. His P(doom) is 0.01%.

Max Tegmark (responding to Ball): You don't need to define superintelligence. You define the harms -- "Can it overthrow the US government? Can it make bioweapons?" -- and the company proves their system doesn't cause them, like the FDA for drugs. "Right now there are more regulations on sandwiches than superintelligence in the US."

Reboot Democracy: ASI is "not a scientifically defined, empirically measurable, or universally agreed-upon benchmark" so it cannot be governed. Doomerism "distorts public understanding" and "serves Big Tech's political and financial interests" by distracting from corporate accountability.

Zvi Mowshowitz: Generally supportive of the ban statement as "creation of common knowledge around common sense thinking." But warns the 2023 FLI pause letter backfired and was used to mock anyone concerned about AI risks. The risk of negative polarization is real.

Ryan Mizzen (external observer): "110 parliamentarians from 140 briefings is a fantastic conversion rate." ControlAI has achieved more parliamentary engagement in 18 months than climate groups have in decades.

Washington vibe shift (MIT Tech Review): David Sacks: "The Doomer narratives were wrong." GPT-5 disappointment has put safety advocates "on the back foot." Helen Toner warns about "boy-who-cried-wolf" credibility risks. The political environment in the US is hostile to ControlAI's message.

What's Absent

Funding transparency. For an organization that criticizes Open Philanthropy for undemocratic power concentration and AI labs for opacity, ControlAI's own financial transparency is essentially zero. We cannot assess funding independence, donor influence, or conflicts of interest.

Governance structure. No disclosed board, advisory council with oversight power, conflict-of-interest policies, or accountability mechanisms. Miotti is the sole listed Director. The organization that demands democratic accountability of AI developers has no visible democratic accountability of its own.

Legislative outcomes. The 279 lawmakers briefed is impressive but remains an output metric. No legislation has been passed. The closest legislative outcome is a UK cybersecurity bill amendment on superintelligence submitted by a coalition member. The transition from "awareness" to "legislation" is the hardest and most important step ahead.

Response to the definition problem. Dean Ball's most potent criticism -- that you cannot define superintelligence in law without either banning beneficial AI or creating a toothless definition -- remains unaddressed in ControlAI's published work. Tegmark's "define harms not capabilities" response is plausible but ControlAI themselves have not published their version.

US strategy specifics. The UK playbook is detailed and proven. The US strategy is much less developed, and the political environment (Sacks/Krishnan, pro-AI super PACs spending $200M) is far more hostile.