Ada Lovelace Institute

Theory of Change

Ada's theory of change is fundamentally about making AI governance work like pharmaceutical regulation: mandatory pre-market testing, statutory enforcement powers, and independent oversight. The causal chain is: rigorous evidence + public deliberation + policy influence = binding regulation = AI systems that are safe, fair, and democratically legitimate.

In their own words: "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." They compare the current state of AI governance to "testing new drugs according to standards set solely by a pharmaceutical company, with regulators having no input or powers to deny approval." They estimate effective AI safety governance would cost GBP 100m+/year, comparable to other sector regulators (the Civil Aviation Authority costs ~GBP 140m/year).

The 2025-28 strategy explicitly frames the "global AI arms race narrative" as an obstacle: "Major nation-states are seeking national security and economic advantage through AI development, framed as a zero-sum competition between nations." Ada argues this narrative leads to deregulation and concentration of power in a small number of tech companies, which they identify as the central problem rather than any intrinsic danger from AI capability.

What They Do

Over 100 reports published since 2018 across biometrics governance, AI regulation, public sector AI, public attitudes, and foundation model evaluation. Key concrete outputs:

EU AI Act: Submitted 18 recommendations; over half adopted in some form. Specific wins include inclusion of "affected persons" as a legal category, establishment of the AI Office, and fundamental rights impact assessments.
UK AISI Critique: "Safety First?" blog (May 2024) reported that 3 of 4 major labs failed to provide pre-release model access to AISI, called for statutory enforcement powers.
"Under the Radar?" report: Most comprehensive public analysis of foundation model evaluation limitations. Key finding: "evaluations alone are not sufficient for safety."
Ryder Review: Independent legal review of biometric data governance (63,000 words), led by Matthew Ryder KC. Recommended new legislation and suspension of live facial recognition.
NHS AIA Pilot: Developed and piloted an algorithmic impact assessment framework -- described as a "world first."
Public attitudes research: Nationally representative surveys with Alan Turing Institute. 2025 data: 89% want independent AI regulation, 89% say safety over speed, 84% believe government prioritizes tech companies over public interest.
COVID contact tracing: Rapid evidence review (April 2020) found "no evidence to support immediate deployment" -- cautionary stance vindicated by subsequent trust problems.

GitHub presence: 6 repositories, all survey datasets. No technical tools or evaluation code.

Key People

Gaia Marcus (Director since June 2024): Career entirely in government data policy and charity sector. Deputy Director DLUHC, Deputy Director ONS, Head of National Data Strategy at DCMS, Head of Civil Service Reform at Cabinet Office. MA Human Rights. No tech industry experience. Her appointment signals a more insider-government engagement approach compared to predecessor Carly Kind.

Carly Kind (founding Director, 2019-Feb 2024): Human rights lawyer, formerly Legal Director of Privacy International. Left to become Australian Privacy Commissioner -- a move consistent with Ada's theory of change (advocacy to regulation pipeline).

Andrew Strait (Associate Director to March 2025): Previously Ethics & Policy Researcher at Google DeepMind. Left for UK AI Safety Institute as Head of Societal Resilience. His Man Group podcast (Sept 2023) is the most candid source on Ada's worldview: argued Bletchley Summit's narrow x-risk focus was "missing a whole suite of AI systems that are already in use."

Team is ~28 staff, described as "legal, policy, social science, philosophy, economy, culture, technology and participation experts." Growth from "tiny beginnings" to current size over 7 years. Heavy social science/law composition; no ML researchers.

Money and Incentives

Primary funder: Nuffield Foundation (UK charitable trust, est. 1943). Initial GBP 5m over 5 years (2018-2023). GBP 3.2m in 2024 alone (per Nuffield annual report). Nuffield committed GBP 150m over 5 years (2025-2030) across all its work; Ada is one of three hosted centres. Estimated Ada share: GBP 3-5m/year.

Secondary funders: Luminate, MacArthur Foundation, Minderoo, Network of European Foundations (European AI Fund), Omidyar, Open Societies Foundation. Specific amounts unknown.

Additional: AHRC Collaborating Partner for "Enabling a responsible AI ecosystem" programme (2022-2026), providing research funding.

Total estimated budget: GBP 3.5-5m per year.

Business model: Pure grants. No products, no consulting, no contracts with AI companies. Revenue entirely from philanthropic foundations and government research councils.

Legal structure: Not a separate legal entity. Ada is a hosted programme within the Nuffield Foundation. Benefits from in-kind admin/finance/HR support. Board appointed by Nuffield. Director reports to Nuffield Foundation.

EA-adjacent funding (not previously known): Cross-validation of grants data reveals ~$706K from Survival and Flourishing Fund and ~$276K from Future of Life Institute. This complicates the "no EA-aligned funders" narrative — while these amounts are small relative to Nuffield's support, they indicate Ada is not entirely outside the EA funding ecosystem.

Key incentive analysis: Ada's funding is entirely from entities that benefit from stronger AI governance -- philanthropic foundations, academic research councils. There is no structural incentive to soften critique of AI companies. The main incentive risk runs the other direction: financial dependence on Nuffield Foundation means if Nuffield's priorities shifted away from AI, Ada would be existentially vulnerable. A secondary risk is that the AHRC partnership gives Ada influence over which AI ethics research gets funded in the UK, creating a potential for self-reinforcing academic consensus.

What Others Say

No substantive published criticism of Ada Lovelace Institute was found despite 37+ targeted search queries. This is itself a significant finding. Possible explanations:

Ada's positions are mainstream in UK policy circles
Ada operates entirely outside the EA/rationalist ecosystem where most AI governance debates happen (0 LessWrong/EA Forum posts, 0 Coefficient Giving grants)
Nuffield Foundation's 80-year institutional reputation provides a credibility shield
Ada is small enough (~28 staff, GBP ~4m/year) that no one considers it worth attacking
Industry actors who might disagree don't engage in public criticism of small think tanks

The strongest implicit counter-argument to Ada's approach comes from the broader "present harms vs existential risk" debate. Those who believe the most important AI risk is loss of control or catastrophic misalignment would argue that present-governance advocacy, while valuable, addresses the wrong timescale. Ada's response (from Strait's podcast): present harms are not hypothetical, they are well-documented, and governance frameworks that work for present harms will also help with future risks.

External media consistently treats Ada as credible. TechCrunch described their UK regulation report as making "pretty awkward reading for ministers." BBC News, The Times, The Guardian, and New Statesman have covered Ada's work. No hostile or dismissive coverage found.

AISI convergence: Even the UK AISI's own December 2025 report validated some of Ada's critique, finding that "safeguards of every system they tested can be broken" and "more capable models do not necessarily have better safeguards."

What's Absent

Financial transparency: Ada does not publish annual accounts or detailed financial reports, despite advocating for AI transparency. The register of board interests is maintained but not publicly available.
No Wikipedia article after 7 years, suggesting limited international name recognition outside UK policy circles.
No engagement with EA/AI safety ecosystem: Complete separation from LessWrong, EA Forum, Alignment Forum, technical AI safety organizations. Two communities working on overlapping problems with zero discursive overlap.
No technical AI safety research: No alignment papers, evaluation methodologies, or safety engineering. All contributions are governance, legal, and social-science focused.
No analysis of frontier labs' RSPs or ASL frameworks: Ada critiques voluntary commitments generally but hasn't published detailed analysis of specific self-governance frameworks.
No engagement with compute governance: Hardware-level safety interventions (chip controls, compute monitoring) are absent from Ada's work.