From 'Merchants of Doubt' to 'Captors of Regulation' — Reading Big AI's Regulatory Capture through 27 Mechanisms

On May 7, 2026, a research team from Trinity College Dublin, the University of Edinburgh, TU Delft, and Carnegie Mellon University — led by Abeba Birhane — posted a paper to arXiv. The title: Big AI's Regulatory Capture: Mapping Industry Interference and Government Complicity. The paper will be presented at FAccT '26 (June 25-28, 2026, Montreal).

The paper's core thesis is this: the economic, political, and societal influence the AI industry has accumulated over the past decade is generating dysfunctions in regulation and oversight. And the tactics being used to generate that dysfunction are essentially identical to the regulatory capture playbook that tobacco, oil, and pharmaceutical industries perfected in the twentieth century.

The research team extends the analytic reach of agnotology — long focused on the "Merchants of Doubt" tradition — to the regulatory lobbying and discourse manipulation of the AI sector. Domain experts hand-coded 100 news articles and extracted 249 instances of regulatory capture, producing a five-category, twenty-seven-mechanism taxonomy.

The Agnotology Lab (ISVD-LAB-002) analyzes how ignorance is produced and maintained, through multi-dimensional coding. This article, as the first piece in the lab's synthesis phase, examines how the "five steps of doubt manufacturing" we theorized in the sixty-year war of tobacco and climate are being almost faithfully repeated at the frontier of AI regulation, using the 249-case empirical dataset.

This article addresses three questions:

1. How can Birhane et al.'s 27 mechanisms be connected to the "mechanism" axis of our 7-axis coding framework?
2. Where do the tactics of tobacco/oil industries and Big AI structurally converge, and where do they diverge?
3. What perspective should we hold when applying this taxonomy to Japanese society?

The answers serve as a starting point for reinterpreting the coding of our 17 existing case studies in Birhane vocabulary, and will be integrated as Chapter 4 of the Zenodo Working Paper ISVD-WP-2026-001 scheduled for June 2026.

Big AI Regulatory Capture Taxonomy — 5 Categories × 27 Mechanisms (27 mechanisms)

Direct influence on policy

4

• •Lobbying
• •Private meetings
• •Political contributions
• •Economic coercion of government

Conflicting involvement

3

• •Revolving door
• •Direct involvement in rulemaking
• •Ownership/Stake in company

Market influence

4

• •Standard setting in consortia
• •Corralling SMEs/orgs to oppose
• •Standard setting via monopoly
• •Economic coercion of competition

Elusion of law

6

• •Disregard existing laws
• •Misinterpret laws
• •Relocation of development and labour
• •Exploit weak regulations/jurisdictions
• •Retaliation
• •Bribery

Epistemic & Discourse Influence

10

• •Corporate sponsorship of events
• •Funding/sponsor research & education
• •Public facing campaign
• •Hyping technologies
• •Playing victim
• •Undermining risks/harms
• •Speculative studies
• •Ethics washing
• •Government adopting industry framing
• •Conflation of public and private interest

Source: Birhane, A. et al. (2026) Big AI's Regulatory Capture — FAccT'26

From Proctor to Birhane — The Contemporary Reach of Agnotology

When Robert N. Proctor proposed agnotology in 2008, his analytic object was the tobacco industry. From hundreds of thousands of pages of internal documents, Proctor uncovered the industry's strategic dictum — "Doubt is our product" — and showed that ignorance is not an accidental byproduct but a product manufactured systematically.

Birhane et al.'s new research transposes this starting point to the AI industry. But it is not mere analogy. As they explicitly state, many of the tactics deployed by the AI industry "mirror strategies that have historically been applied by similar industries such as Big Tobacco, Big Pharma, and Big Oil" (Birhane et al., 2026, §5.1.1). The same lineage of human networks, the same rhetoric, the same forms of interference in legislative processes — these are being replayed with the object replaced.

Points of Contact with Existing ISVD Research

This lab has already analyzed the structure of ignorance production through multiple case studies. Birhane et al.'s research offers a unifying lens to reinterpret them.

• The Doubt Manufacturing Industry — Sixty-Year War of Tobacco and Climate: Theorizes the five-step model of doubt manufacturing. Directly corresponds to Birhane et al.'s "Public facing campaign," "Speculative studies," "Undermining risks/harms"
• Anatomy of Predatory Business — Information Asymmetry: Market exploitation of information asymmetry. Connects to "Conflation of public and private interest"
• Algorithmic Filter Bubbles and Ignorance: Platform companies' control over epistemic environments
• What Is Not Reported — Media Agenda Setting and Invisibility: Mechanisms of media capture

Birhane et al.'s 27-mechanism taxonomy functions as an academically systematized taxonomy running through these individual cases. We can expand our 7-axis coding framework — particularly the mechanisms axis — with the empirically grounded vocabulary of 249 cases.

Five Categories × Twenty-Seven Mechanisms — Big AI's Tactical Taxonomy

The 27 mechanisms Birhane et al. present are organized into five upper-level categories.

Category 1: Direct influence on policy — 4 mechanisms

Direct approaches to officials and regulators. The classic group of lobbying tactics.

• Lobbying: Communications intended to influence the decisions or favourable positions of public officials
• Private meetings: Meetings outside lobbying and regulated channels
• Political contributions: Donations by persons or companies to political entities
• Economic coercion of government: Inducing policy through investment withdrawal or threats of withdrawal

The dataset records cases such as Amazon, Google, Meta, and Apple each contributing $1 million to Donald Trump's 2024 US presidential campaign, while Elon Musk donated $250 million to pro-Trump groups (Birhane et al., 2026, §1).

Category 2: Conflicting involvement — 3 mechanisms

Structural problems where public duties and corporate interests collide within the same agent.

• Revolving door: Public officials taking up conflicting roles in private entities or vice-versa
• Direct involvement in rulemaking: Involvement in developing law or policy without legal mandate or authorisation
• Ownership/Stake in company: Direct/indirect ownership or stake in regulated organisations by public officials

Specific examples: Former French Secretary of State for Digital Transition Cédric O becoming a shareholder and advisor for Mistral. Former UK Deputy Prime Minister Nick Clegg joining Meta. The influence wielded by Matt Clifford over UK AI policy. Of the 249 cases, 10 high-profile revolving door cases were identified: 6 US, 3 UK, 1 France.

Category 3: Market influence — 4 mechanisms

Establishing dominant position through coordination and coercion of market actors.

• Standard setting in consortia: Industry-led standards excluding broader stakeholders
• Corralling SMEs/orgs to oppose: Forming a unified industry front
• Standard setting via monopoly: Imposing de facto standards through monopolistic position
• Economic coercion of competition: Anti-competitive pricing and similar tactics

Category 4: Elusion of law — 6 mechanisms

Direct or indirect violations of the spirit or letter of existing laws.

• Disregard existing laws: Simply ignoring legal requirements or processes
• Misinterpret laws: Intentional misrepresentation of laws
• Relocation of development and labour: Moving operations to jurisdictions with weaker rules
• Exploit weak regulations/jurisdictions: Regulatory arbitrage
• Retaliation: Retaliating against whistleblowers or regulators
• Bribery: Payments or favours in exchange for official actions

In the dataset, systematic violations of copyright and privacy laws are detected. Disregard existing laws (17%) and Misinterpret laws (14%) emerge as high-frequency mechanisms — second only to Discourse & Epistemic Influence.

Category 5: Epistemic & Discourse Influence — 10 mechanisms

This is the heartland of agnotology. Of the 249 capture instances, this is the most frequent category, containing ten sub-mechanisms.

• Corporate sponsorship of events: Agenda setting via industry-sponsored conferences
• Funding/sponsor research & education: Steering the direction of academic research
• Public facing campaign: Media exposure and targeted advertising
• Hyping technologies: Promoting AI capabilities without empirical evidence
• Playing victim: Claiming unfairness as a regulatory target
• Undermining risks/harms: Denial or minimisation of harms
• Speculative studies: Publishing studies lacking scientific rigour
• Ethics washing: Performative ethics serving as a substitute for binding accountability
• Government adopting industry framing: Public bodies adopting industry vocabulary and premises
• Conflation of public and private interest: Private interest presented as public benefit

Each of these resonates multiply with our existing case analyses. Funding research connects to the knowledge-production distortion of Strategic Ignorance and EBPM; Public facing campaign to the Brandolini amplification; Ethics washing to Authority and the Reproduction of Ignorance.

Integration into the ISVD 7-Axis Coding

Our 7-axis coding framework has designed the mechanisms field as free tags (string[]). Birhane et al.'s 27 mechanisms can be incorporated as an empirically grounded vocabulary library for this free-tag field.

As the frontmatter of this article shows, we newly introduce seven mechanism tags:

These will be back-applied to the coding of our existing 17 case studies, increasing the precision of cross-case pattern extraction. For example, recoding case-tobacco-climate-doubt-industry's mechanisms to read ["doubt-manufacturing", "speculative-studies", "front-organization", "regulatory-capture"] in Birhane vocabulary will establish comparability with the new synthesis article.

Eleven Dominant Narratives that Justify Regulatory Capture

Birhane et al. extracted 11 dominant narratives from 49 of 100 articles that justify regulatory capture. Ranked by frequency (Birhane et al., 2026, Fig.3):

What is notable: the top four narratives (Regulation stifles innovation / Red tape / National interest / Competitiveness) are repeatedly adopted not only by the regulated but by the regulators themselves. Birhane et al. document the official call for "deregulation" by EU Commission President von der Leyen, the drafting of the General Purpose AI Code of Practice with industry participation and the optionalization of human-rights protections over successive stages, the delay of the UK AI Bill, and German authorities' discussions of withdrawing laws to maintain "attractive[ness] to tech companies" (Birhane et al., 2026, §1).

This is why "government complicity" is in the paper's title. Capture is described not as a one-sided industry attack but as a bidirectional process completed when governments themselves actively adopt the narratives.

The paper's scope is limited to the AI regulatory processes of the EU, US, UK, Korea, and France; Japanese cases are not included. But what matters for ISVD readers is to hold the lens by which we can check whether the same structures are creeping into Japan.

The following points may arise in Japan's AI strategy discussions (this article offers an overview only; deeper analysis will follow in a discourse-phase article):

• The representation of Big Tech in expert councils at METI and MIC (verifiable from public records)
• The presence of organized mobilization in public comment processes (aggregatable from open-government data)
• Discourse analysis of the "Japan is falling behind in AI regulation" narrative (Narrative 4: Competitiveness)
• Flows of industry funding into universities and the independence of AI ethics research
• The presence/absence of civil society representation at AI strategy councils

These are not, at present, grounds to declare that capture is occurring. What Birhane et al. provide is a framework for judgment. Applying this framework empirically to Japan is a forthcoming task for this lab.

In §5.1.2, Birhane et al. propose eight directions for resisting Big AI capture (Birhane et al., 2026, §5.1.2):

1. Civil society organizations advising on regulatory implementation (standard setting, strategic litigation)
2. Building shared narratives and bottom-up agendas (real-world examples like the AI Now Institute People's AI Action Plan)
3. Exposing the influences of lobbying and deregulation (Corporate Europe Observatory activities, etc.)
4. Continuous verification through independent investigative journalism
5. Conducting independent audits and circulating their results
6. Reinvigorating labor movements (organized responses to generative-AI-driven labor substitution)
7. Applying existing climate and environmental commitments to the AI sector
8. Rights protection (within and outside the boundaries of regulatory concerns)

Three concrete actions for ISVD readers:

Action 1: Submit public comments to Japan's AI policy consultations under your real name

Submit citizen and practitioner opinions to public comments by METI, MIC, and the Personal Information Protection Commission. This is a direct counter to the organized mobilization on the "capture" side. This lab will continue calling for public comments on the ISVD official channels.

Action 2: Track the membership composition of Japan's AI policy expert councils from public sources

Continuously monitor three points: members' affiliations, past industry funding received, and conflict-of-interest disclosures. This lab plans to publish quarterly observation reports (as part of the discourse phase).

Action 3: Develop the eyes to detect Ethics washing in your own organization

Develop the habit of asking what companies and organizations that declare "we have adopted AI ethics" actually do (funding flows, personnel, regulatory response). Our Authority and the Reproduction of Ignorance and Strategic Ignorance and EBPM provide the foundational concepts for that analysis.

The main factual claims in this article are based on a full reading of the body PDF of Birhane et al. (2026) arXiv:2605.06806. The paper adopts the following methodology.

Methodology: Design Science Research (DSR). After 10 months of expert meetings and literature review, the taxonomy was iteratively refined. 100 articles were independently annotated by two domain experts, with final classifications determined through consensus formation. The analysis centers on Reuters articles published around the EU AI Act negotiations and three global AI summits (UK 2023, South Korea 2024, France 2025), with a PRISMA flow narrowing 24,629 records to 100.

Limitations: Birhane et al. explicitly state that this study is descriptive and interpretive, and does not measure the frequency or causation of regulatory capture. The aim is "to discover what mechanisms are used and to taxonomize them." Sampling is Reuters-centered, so it is not free from reporting-style biases. Cases from Japan, Latin America, and Africa are not included.

Limitations of this lab's interpretation: This article introduces and synthesizes the paper for Japanese readers; empirical work specific to Japan is a forthcoming task. The "Implications for Japan's AI Regulatory Discussion" section presents hypothetical points of investigation, not assertions.

Data sources: Cross-checked against Birhane et al. (2026) arXiv:2605.06806 (full body PDF), The Register (2026-05-18), Tech Policy Press (2026-05), Mirage News (Edinburgh press-release feed), and the AI Now Institute official site as primary materials. All URLs in this article have been verified for HTTP status and page-title match by automated check (as of 2026-05-20).

This article is the first in the Agnotology Lab's synthesis phase. Starting from this article, the following synthesis-phase articles will follow:

• Integrated structural analysis of epistemic injustice (synthesis #2, planned for June 2026)
• Institutional design for the post-truth era (synthesis #3, planned for June 2026)
• Agnotological re-reading of Japanese social theory (synthesis #4, planned for July 2026)

The findings of these synthesis articles will be consolidated in the publication phase into Zenodo Working Paper ISVD-WP-2026-001 "An Inductive Research Framework for Agnotology" (scheduled for publication in June 2026).

Related Articles

• Research framework of the Agnotology Lab — Overview of the 7-axis coding framework
• The Doubt Manufacturing Industry — Theoretical foundation of this article
• Algorithmic Filter Bubbles and Ignorance — Structures of platform dominance
• Authority and the Reproduction of Ignorance — Cognitive basis of ethics washing
• Strategic Ignorance and EBPM — Mechanisms of knowledge-production distortion

Reference Books