Ada Lovelace Institute

Stated Theory of Change

Ada's stated theory of change is that AI should be governed with the same rigor as pharmaceuticals, aviation, and financial services. The mechanism is: (1) build rigorous evidence on how AI affects people and society, (2) surface public attitudes and democratic demands for regulation, (3) influence policymakers with evidence-based recommendations, (4) achieve binding statutory regulation with enforcement powers. They explicitly reject voluntary, self-regulatory approaches.

The core premise: "Those best able to manage risks and harms at each point in the AI value chain should be credibly incentivised and empowered to do so." The operative word is "incentivised" -- Ada's worldview is that without statutory requirements and enforcement, companies will not adequately manage AI risks because commercial incentives push the other direction.

They estimate effective AI safety governance would cost GBP 100m+/year (comparable to existing sector regulators) and propose industry levies to fund it, as in pharmaceuticals and finance.

Revealed Theory of Change

Ada's actions align closely with their stated theory. Their publications consistently argue for binding regulation over voluntary commitments. Their EU AI Act engagement produced measurable legislative outcomes (over half of 18 recommendations adopted). Their public polling provides the empirical ammunition for regulation arguments. Their AISI critique was specific and factual (3 of 4 labs non-compliant with access commitments).

Where stated and revealed theory diverge slightly:

1. Incrementalism vs. ambition. The strategy language is sweeping ("redistribution of power," "challenge stories told about technology"), but the actual policy recommendations are incremental: extend existing regulatory powers, create new statutory duties, establish industry levies. This is appropriate pragmatism for a small org working within the policy system, but it means Ada is optimizing for achievable reform rather than transformative change.

2. Present harms vs. future risks. Ada's stated position acknowledges both: "There is potential for unprecedented and likely uneven acceleration in AI capabilities... which could lead to sweeping changes in society." But every specific policy recommendation and research project addresses present or near-term governance problems. The x-risk acknowledgment functions as throat-clearing rather than a driver of research priorities.

3. Independence claims vs. structural ties. Ada claims independence from government and industry, and this is largely true financially. But the career pipeline (Strait to AISI, Kind to Australian regulator, Marcus from civil service) reveals deep structural embeddedness in the government ecosystem. This is not necessarily a weakness -- insider access enables policy influence -- but it complicates the "independent voice" framing.

Key Assumptions

Assumption 1: Regulatory analogy is valid. Ada's entire approach rests on the premise that AI can be regulated like pharmaceuticals or aviation. This assumes: (a) risks are identifiable before deployment, (b) pre-market testing is feasible and meaningful, (c) regulatory bodies can keep pace with technical development, (d) enforcement is practical across borders.

• Evidence for: Regulatory models work in other high-risk domains. Ada's own research documents specific mechanisms that could transfer.
• Evidence against: AI systems are more general-purpose than drugs or aircraft. AI development cycles are faster than regulatory cycles. Global companies can deploy from jurisdictions with lighter regulation. The "bait and switch" problem Ada identifies (testing one model version, deploying another) undermines the entire pre-market approval concept.
• Testable: Yes -- the EU AI Act's implementation over 2025-2027 will test whether risk-based AI regulation is enforceable at scale.
• If wrong: Ada's entire theory of change fails. Present-governance approaches would need to be supplemented with fundamentally different mechanisms.

Assumption 2: Public opinion matters. Ada invests heavily in demonstrating public demand for regulation (89% want independent regulation, etc.). This assumes democratic accountability translates into policy action.

• Evidence for: Public polling has been cited by policymakers. UK government included Ada's research in white paper.
• Evidence against: The gap Ada documents -- between public expectations and government action -- is itself evidence that public opinion is not currently driving policy. Governments are pursuing deregulation despite 84% of the public believing government favors tech companies. Industry lobbying and economic competition pressures may dominate.
• If wrong: Ada's public participation work becomes academically interesting but politically irrelevant. The "present harms" framing would need to find leverage points beyond democratic accountability.

Assumption 3: Incremental policy change is sufficient. Ada's approach assumes the pace of governance reform can keep up with the pace of AI development.

• Evidence for: EU AI Act passed in 2024. UK legislation is being debated. International coordination (AISI network) is expanding.
• Evidence against: Ada's own December 2025 blog admits the UK AI Bill has stalled. Deregulatory headwinds are strengthening globally. The gap between AI capabilities and governance capacity appears to be widening, not narrowing.
• If wrong: Ada's modest incremental wins could be rendered irrelevant by rapid capability gains. A "fast takeoff" scenario would make present-governance frameworks obsolete before they are fully implemented.

Assumption 4: Present harms and future risks are not in tension. Ada implicitly assumes that governance frameworks built for present harms will also help with future, more extreme risks.

• Evidence for: Regulatory infrastructure (enforcement powers, transparency requirements, evaluation capacity) is useful regardless of the specific risk level. A 2025 PNAS study found that concern about existential risk does not diminish concern about present harms.
• Evidence against: Some future risks (recursive self-improvement, instrumental convergence, deceptive alignment) may require fundamentally different governance mechanisms than present harms. Regulatory capacity built for evaluating bias in hiring algorithms may not transfer to evaluating the alignment of superhuman systems.
• If wrong: Ada's work remains valuable for present-governance challenges but misses the highest-stakes scenario.

Strengths

1. Genuine independence from AI industry funding. Unlike many AI governance organizations, Ada takes no money from tech companies, frontier labs, or their associated philanthropic vehicles. The funder base (Nuffield Foundation, Luminate, MacArthur) has no structural interest in weakening AI regulation. This is a rare and valuable position.

2. Measurable policy impact. The EU AI Act influence (half of 18 recommendations adopted) is one of the most concrete policy outcomes in the AI governance space. Most AI safety orgs cannot point to specific legislative text they influenced.

3. Public legitimacy evidence base. Ada's nationally representative polling provides empirical ammunition that no other AI governance org matches. The 89% support for independent regulation is a powerful political fact.

4. Institutional stability. Nuffield Foundation's 80-year history and GBP 150m five-year commitment provides unusual financial security for a think tank. Ada doesn't face the year-to-year fundraising pressure that distorts many organizations' priorities.

5. Practical tools. The NHS AIA pilot, the procurement framework for local government, and the EU Code of Practice working group participation show Ada can translate theory into practice.

6. Cross-disciplinary team. The legal, social science, philosophy, and policy mix enables analysis that purely technical organizations miss. The Ryder Review (commissioning a KC for independent legal analysis) demonstrates serious methodological rigor.

Weaknesses and Risks

1. Single-point-of-failure funding. GBP 3.2m from Nuffield in 2024 out of an estimated GBP 3.5-5m total budget means ~65-90% funding concentration in a single source. If Nuffield Foundation's priorities shift, Ada is existentially threatened.

2. No technical depth. Ada cannot evaluate AI systems, propose technical safety measures, or engage with the alignment research community on a technical level. Their "Under the Radar?" report critiques evaluation methods but cannot propose better ones. In a field where technical credibility matters, this limits Ada's influence with the people building the systems.

3. The incrementalism trap. Ada's theory of change requires governments to implement binding regulation. But Ada's own December 2025 assessment admits: the UK AI Bill has stalled, deregulatory headwinds are strengthening, and the gap between AI capabilities and governance is widening. If the policy window closes before Ada's recommendations are implemented, their entire approach may prove too slow.

4. Ecosystem isolation. Zero engagement with the EA/AI safety community means Ada misses insights about tail risks, alignment challenges, and capability dynamics. The EA community misses Ada's strengths in democratic legitimacy, legal analysis, and institutional governance. Both communities are poorer for the lack of cross-pollination.

5. Governance opacity. Ada advocates for AI transparency but does not publish its own financial accounts, board interest register, or detailed budget. The Shakir Mohamed (DeepMind) board appointment creates a perceived conflict that the unpublished interest register cannot address.

6. UK-centricity. Despite EU engagement, Ada's work is predominantly UK-focused. The UK is not where the most important AI governance decisions will be made -- those happen in the US, EU, and China. Ada's influence on the global debate is mediated through UK policy channels that may themselves be marginal.

7. The Carly Kind gap. The departure of both the founding Director (Kind) and the most technically credible Associate Director (Strait) within 14 months creates leadership continuity risk. Marcus brings government experience but less international profile and less tech-sector credibility than her predecessor.

Cross-References

• AI Now Institute (US): Most direct comparison. AI Now also focuses on present harms, power concentration, and governance. Both are independent of AI industry funding. AI Now is more confrontational and US-focused; Ada is more establishment-embedded and UK/EU-focused.
• UK AISI: Ada's most important interlocutor. Ada critiques AISI's voluntary approach and narrow frontier focus; AISI hired Ada's former Associate Director (Strait). The relationship is constructive tension.
• Centre for the Study of Existential Risk (CSER): Founding board member Huw Price connects the organizations intellectually, but their research programs do not overlap. CSER works on x-risk; Ada works on present governance.
• Centre for Data Ethics and Innovation (CDEI): Government-created body with overlapping UK AI governance mandate. Ada provides independent civil society perspective that CDEI, as a government entity, cannot.
• Partnership on AI (PAI): Ada is a PAI partner. PAI is multi-stakeholder (including industry); Ada is independent. Complementary but different governance models.
• Coefficient Giving / Open Philanthropy / EA Funders: Zero overlap. Ada is funded by entirely different philanthropic networks. The complete separation between Ada's funder network and the EA funder network reflects the deeper intellectual separation between present-governance and x-risk approaches.

What Would Change This Assessment

• UK AI Bill passes with Ada's recommendations. If the stalled UK legislation moves forward incorporating Ada's proposals for statutory enforcement powers and pre-market approval, that would validate their theory of change in the jurisdiction where they have the most influence.
• EU AI Act implementation fails. If the Act proves unenforceable, overly burdensome, or gamed by companies (as Ada's own "bait and switch" analysis warns is possible), it would undermine the regulatory analogy at the center of Ada's approach.
• Rapid capability gains outpace governance. If AI systems achieve significantly superhuman performance before governance frameworks are in place, Ada's incremental approach would look insufficient regardless of its quality.
• Ada publishes financial accounts. If Ada addressed its own transparency gap by publishing detailed annual accounts and board interest registers, it would strengthen their credibility and remove one of the few valid process criticisms.
• External criticism emerges. If substantive critics make specific arguments about why Ada's approach is wrong (as opposed to simply ignoring Ada), that would test whether Ada's ideas hold up under adversarial scrutiny.
• Nuffield Foundation reprioritizes. If Nuffield shifted funding away from AI governance, Ada's survival would depend on secondary funders who currently provide only a fraction of the budget.

Self-Critique

What sources should I have checked but didn't? The Nuffield Foundation's full 2024 annual report (PDF was referenced but not fully fetched) would provide more detailed financial data. The Actuary interview with Carly Kind (fetch failed) might contain candid material. Parliamentary committee evidence from Ada (GAI0086, PDF inaccessible) would show how they present to government.

Where is this analysis potentially biased? I may be underweighting Ada's impact because I'm evaluating them partly against an x-risk framework they explicitly reject. From Ada's own perspective, reducing present harms and establishing governance infrastructure IS the most important contribution to AI safety. If they're right that present governance frameworks are the best path to managing future risks (their assumption 4), then my critique about insufficient engagement with x-risk is beside the point.

What would a thoughtful person who disagrees say? Someone sympathetic to Ada might argue: "The AI safety community has been warning about AGI risk for 20 years and has produced zero binding regulation. Ada has influenced the EU AI Act in 3. Who is actually reducing risk?" This is a fair point -- measurable incremental policy wins may be more valuable than theoretical work on alignment problems that may never materialize in the way predicted.

What's my single weakest claim? The claim that Ada's ecosystem isolation from the EA/AI safety community is a "weakness." It could equally be a strength -- maintaining intellectual independence from a community with its own biases and blind spots. Ada's focus on present harms and democratic legitimacy addresses real problems that the AI safety community tends to under-prioritize.

What information would most change my view? Evidence that the EU AI Act's implementation is either clearly succeeding or clearly failing at managing real AI risks. This would either validate Ada's regulatory analogy (their core theory) or undermine it. Right now, it's too early to tell.