Redistributed

When an industry shapes the rules meant to govern it, the result is regulation that protects incumbents rather than the public. This is regulatory capture — and AI policy is increasingly exhibiting its symptoms.

Understanding capture matters because regulation is where redistribution happens. Rules determine who can build AI systems, who bears liability when they fail, what counts as "safe," and whether competition can challenge established players. If those rules are written by the industry itself, the outcome is predictable: value flows upward, risk flows downward, and the appearance of accountability substitutes for the real thing.

What Regulatory Capture Means

Economist George Stigler formalized capture theory in his 1971 paper "The Theory of Economic Regulation." His core insight was counterintuitive: regulation often benefits the regulated industry, not the public. This happens not through corruption but through structure. Regulated firms have concentrated interests and resources; the public's interest is diffuse. Firms show up to every hearing, file every comment, fund every study. The public does not.

Stigler identified the mechanisms: industries seek regulations that restrict entry (protecting incumbents from competition), control substitutes (preventing alternatives from emerging), fix prices (ensuring profits), and provide subsidies (direct transfers from public to private). The prediction is that regulation will, over time, bend toward the interests of those with the most persistent engagement — which is almost always the industry itself.

This doesn't require bad actors — and recognizing this is important. Many regulators and industry participants are acting in good faith. Regulators depend on industry for information about how systems actually work. They recruit from industry and eventually return to it. They develop relationships with the people they oversee. They share assumptions about what's technically possible and economically viable. The capture happens through information asymmetry, personnel flow, and cognitive alignment — not conspiracy.

Susan Crawford's work on telecommunications provides the clearest modern example. The cable and telecom industries achieved near-total capture of the FCC, resulting in decades of policy that protected monopoly rents, prevented municipal broadband, and allowed the US to fall behind peer nations in internet access and pricing. The industry didn't bribe regulators; it shaped the reality regulators operated in.

Classic Capture Mechanisms

Three mechanisms drive capture across industries. Each operates in AI governance.

The Revolving Door

Personnel move between regulatory agencies and the firms they regulate. This creates three problems. First, current regulators may soften enforcement to preserve future employment prospects. Second, former industry employees bring industry assumptions into government. Third, former regulators bring insider knowledge and relationships back to industry.

The revolving door is not inherently corrupt. Government agencies need people who understand the industries they regulate. Industry hires former regulators because they genuinely know how government works. But the cumulative effect is a shared worldview between regulator and regulated — one that tends to treat industry priorities as synonymous with public interest.

Lobbying and Information Asymmetry

Regulators need to understand the systems they're regulating. In complex technical domains, the primary source of information is the industry itself. This creates structural dependence.

Companies fund research, employ experts, and produce the data that policymakers rely on. They have armies of lawyers and lobbyists to interpret draft regulations and propose changes. Public interest groups, academic researchers, and affected communities have a fraction of these resources.

When a regulator asks "is this requirement technically feasible?" or "what would compliance cost?", the most detailed answers come from the companies that will bear the costs. Those answers are not neutral.

Standard-Setting as Capture

Technical standards — the specifications that define what counts as compliance — are often developed outside government, in industry consortia and standards bodies. This is presented as efficient: let the experts who understand the technology define the requirements.

But standards bodies are not neutral arenas. Participation requires resources: travel, fees, staff time. The organizations that can afford sustained engagement are large companies. The standards that emerge reflect their priorities, capabilities, and business models. Smaller competitors, civil society groups, and affected communities are structurally excluded.

Standards become regulation when governments adopt them as binding requirements. At that point, industry-developed specifications have the force of law — without having gone through democratic legislative processes.

How These Apply to AI

AI governance exhibits all three capture mechanisms, with additional features that make capture particularly likely.

Technical Complexity Amplifies Asymmetry

AI systems are opaque even to their creators. Regulators face extreme information asymmetry: they cannot inspect a model's training process, verify claims about capability, or predict emergent behaviors. They depend on the labs for model access, evaluation data, and interpretability research.

This is not like telecommunications, where regulators could understand the technology even if industry controlled deployment. In AI, the gap between what labs know and what regulators can know is fundamental. As Meredith Whittaker and the AI Now Institute have documented, this asymmetry enables companies to define the terms of the safety debate — what risks matter, what mitigations are adequate, what evaluation metrics count.

Concentrated Industry, Diffuse Impact

The companies building frontier AI systems can be counted on one hand: OpenAI, Anthropic, Google DeepMind, Meta, a few others. They have clear, concentrated interests in AI policy. The people affected by AI systems — workers displaced by automation, communities subject to algorithmic decisions, societies navigating synthetic media — are diffuse, unorganized, and mostly absent from policy discussions.

This matches Stigler's conditions for capture precisely. The industry has the resources and motivation for sustained engagement. The public interest is dispersed across millions of people with other concerns.

Expertise Scarcity Creates Dependency

The people who understand AI systems at the technical level mostly work for the companies building them. The talent pool is small and concentrated. When governments need to hire AI expertise, they recruit from industry. When they need technical advice, they consult with industry. The resulting relationships create cognitive capture even without intentional influence.

The UK's AI Safety Institute and NIST's AI work both depend heavily on personnel with industry backgrounds. This isn't a failure of hiring — it reflects where expertise lives. But it means the framings, priorities, and assumptions of industry are embedded in the regulatory apparatus from the start.

Case Study: The EU AI Act Negotiations

The EU AI Act, passed in 2024, is the world's first comprehensive AI regulation. Its negotiation history illustrates capture dynamics in action.

The original Commission proposal, published in April 2021, focused on application-layer risks: high-risk uses like hiring, credit, and law enforcement. It did not regulate foundation models directly.

By the time GPT-4 launched in March 2023, the gap was obvious. Parliament moved to add foundation model provisions. The result was a negotiation between three positions: Parliament (stronger regulation), Council (member state interests), and an intensive lobbying campaign by AI labs and their investors.

The lobbying was explicit. OpenAI's Sam Altman met with European officials and suggested the company might leave the EU if regulations were too strict. French and German governments, influenced by their domestic AI sectors, pushed for weaker foundation model rules. Venture capital firms argued that compliance costs would entrench US incumbents.

The final text reflects these pressures. Foundation model obligations were weakened during trilogues. The compute threshold for "systemic risk" models (10^25 floating-point operations) was set high enough to exempt most current models. Open-source provisions were softened after industry objections. The European AI Office, which regulates foundation models directly, has limited resources and enforcement powers compared to the companies it oversees.

None of this was hidden. The lobbying was public. The arguments were made openly: Europe must remain competitive; regulation must not stifle innovation; industry knows best what's feasible. These arguments may even be partly correct. But the pattern — industry influencing the rules that govern it, in the direction of lighter oversight — matches capture theory exactly.

Case Study: US AI Safety Frameworks

American AI governance is more fragmented than Europe's, but capture dynamics operate at every level.

The White House Executive Order (October 2023) requires companies to report large training runs to government. It relies on voluntary commitments made by major labs in July 2023. As Tim Wu has documented in his work on platform power, voluntary commitments are not regulation — they're promises that can be withdrawn when convenient.

The executive order's development involved extensive industry consultation. The resulting requirements track what companies were already willing to do. Reporting thresholds were set at levels that affect few current projects. Safety requirements defer to industry-developed best practices rather than imposing independent standards.

NIST's AI Risk Management Framework is explicitly industry-informed. The framework was developed through a "consensus-based" process that prioritized stakeholder input — and the stakeholders with resources to participate were overwhelmingly industry representatives. The resulting framework is useful as guidance but lacks the teeth of binding regulation.

Sector-specific agencies (FDA, SEC, DOT) have authority over AI in their domains but limited technical capacity. They tend to adopt industry-developed standards and rely on company self-certification. The expertise gap makes independent oversight difficult.

The cumulative picture is governance shaped by the governed. The companies building AI systems have been the primary architects of the frameworks meant to constrain them.

"Responsible AI" as Capture

Perhaps the subtlest form of capture is the framing of AI governance as "responsibility" rather than regulation.

Major AI labs have extensive responsible AI teams, publish research on safety, and advocate publicly for AI governance. Some of this work is genuinely valuable — model cards, datasheets for datasets, and safety evaluations represent real contributions to transparency. But these efforts can simultaneously function as a capture strategy.

When a company defines what "responsible AI" means, it sets the terms for regulation. Problems the company can solve — bias testing, content moderation, model documentation — become the focus of policy. Problems the company cannot or will not solve — concentration of compute, labor displacement, power asymmetry — are framed as outside the scope of AI governance.

"Responsible scaling policies" are an example. Companies commit to evaluating models for dangerous capabilities and pausing deployment if thresholds are exceeded. This sounds prudent. But the company defines the thresholds, conducts the evaluations, and decides what "pause" means. The appearance of safety discipline substitutes for external oversight.

Self-regulation also raises entry barriers. Compliance costs that frontier labs can absorb become obstacles for smaller competitors. "Best practices" codified into standards reflect the capabilities of well-resourced organizations. Safety requirements designed for GPT-4-scale systems may be inappropriate or impossible for research projects, open-source developers, or companies in other countries.

The result is regulation that entrenches incumbents. As AI Now's reports on corporate concentration have documented, the companies calling loudest for AI governance are the ones best positioned to comply with — and benefit from — the governance frameworks they're shaping.

Standards Bodies as Capture Sites

Technical standards turn policy into practice. In AI, the key bodies are NIST (US), ISO/IEC (international), and IEEE. Each is a potential capture site.

NIST develops AI standards through public-private partnerships. Industry participation is essential — NIST lacks the expertise to develop standards alone. But industry's participation is also influence. The AI Risk Management Framework emerged from a process dominated by tech company representatives.

ISO/IEC develops international AI standards through consensus among national bodies. Participation requires sustained engagement: sending delegates to meetings, commenting on drafts, building coalitions. Large companies and wealthy nations are structurally advantaged. Standards that emerge may not reflect the priorities of smaller companies, developing countries, or civil society.

IEEE develops ethical and technical standards for AI. Its processes are open in theory but require membership and participation fees. The organizations that engage most intensively are the ones with the most resources.

When governments adopt these standards — incorporating them into procurement requirements, referencing them in regulations, treating compliance as evidence of safety — industry-developed specifications acquire legal force. The democratic deficit is concealed by technical complexity.

Counter-Capture: Designing Resistant Institutions

Capture is structural, not moral. The solution is not better people but better structures. Several design principles can make institutions more resistant to capture.

Diversify Funding and Expertise

Capture thrives on dependency. Regulators who depend on industry for information and personnel will tend toward industry positions.

Counter-capture requires independent sources of expertise. This means funding public AI research institutions that can evaluate industry claims. It means creating pipelines from academia and civil society into regulatory roles. It means investing in technical capacity within government so regulators can conduct their own evaluations rather than accepting industry submissions.

The EU AI Office's limited resources are a choice, not an inevitability. Adequately funding regulatory bodies is anti-capture policy.

Expand Participation

Capture exploits the gap between concentrated industry interests and diffuse public interests. Closing that gap requires making participation easier for non-industry stakeholders.

This means funding civil society organizations to engage in regulatory processes. It means translating technical proceedings into accessible language. It means creating structured roles for affected communities — workers, minorities subject to algorithmic discrimination, users of AI systems — in standard-setting and regulatory consultation.

The computing access required to evaluate AI systems is itself a participation barrier. A regulator that cannot run its own experiments depends on company-provided results. Public compute infrastructure for research and regulatory purposes is an anti-capture investment.

Lengthen Cooling-Off Periods

The revolving door can be slowed by requiring longer gaps between government service and industry employment (and vice versa). Two-year cooling-off periods are common; five-year periods for senior roles would be more effective.

Restrictions should apply in both directions: former regulators should face limits on lobbying their former agencies, and former industry employees entering government should face recusal requirements for matters involving their previous employers.

Increase Transparency

Capture operates most effectively in obscurity. Transparency requirements can expose influence.

This means publishing lobbying contacts, requiring disclosure of industry funding for research cited in regulatory proceedings, and making standard-setting deliberations public. The EU's transparency register and lobbying disclosure rules are models, though imperfect ones.

For AI specifically, transparency could include mandatory disclosure of compute investments, training data sources, and capability evaluations. Information that companies must reveal cannot be selectively provided to shape regulatory perceptions.

Separate Standard-Setting from Adoption

When governments automatically adopt industry-developed standards, they outsource lawmaking to private bodies. Capture resistance requires maintaining distance between technical standards and legal requirements.

Governments can adopt standards selectively, with modification. They can require that standards undergo public review before incorporation into regulation. They can create their own evaluation processes rather than deferring to industry certifications.

The goal is not to exclude industry expertise but to ensure it passes through democratic accountability before acquiring the force of law.

The Redistribution Stakes

Regulatory capture is not abstract. It determines how AI's benefits and harms are distributed.

If safety requirements are set by frontier labs, they will be calibrated to what those labs can achieve — raising barriers for competitors and cementing market position. If liability rules are designed by the companies that might bear liability, they will minimize corporate exposure while leaving affected individuals without recourse. If standards are set by well-resourced incumbents, they will reflect incumbent capabilities and business models.

The communities with the least power are already most affected by AI systems — through algorithmic denial of benefits, predictive policing, automated hiring discrimination. These same communities have the least access to regulatory processes. Capture ensures that the rules are written without them.

Counter-capture is redistribution policy. Making regulatory institutions more resistant to industry influence means making them more responsive to the public interest. It means ensuring that when AI creates value, the rules governing that value don't automatically direct it toward the powerful.

This is not a problem that solves itself. Industries do not voluntarily cede influence over their own regulation. The question is whether democratic institutions can muster the resources, expertise, and political will to govern AI in the public interest — or whether we outsource that task to the industry itself and call the result "responsible AI."

Sources

Stigler, George. "The Theory of Economic Regulation." Bell Journal of Economics and Management Science, 1971. JSTOR
Crawford, Susan. Captive Audience: The Telecom Industry and Monopoly Power in the New Gilded Age. Yale University Press, 2013. Yale Press
Wu, Tim. The Curse of Bigness: Antitrust in the New Gilded Age. Columbia Global Reports, 2018. Columbia
Wu, Tim. The Attention Merchants: The Epic Scramble to Get Inside Our Heads. Knopf, 2016. Penguin
Whittaker, Meredith et al. "AI Now 2019 Report." AI Now Institute, 2019. AI Now
AI Now Institute. "Confronting Tech Power." 2021. AI Now
European Union. "Regulation (EU) 2024/1689 (AI Act)." Official Journal, 2024. EUR-Lex
Corporate Europe Observatory. "The Lobbying Behind the EU AI Act." 2023. CEO
White House. "Executive Order on AI." 2023. White House
NIST. "AI Risk Management Framework." 2023. NIST
Veale, Michael & Zuiderveen Borgesius, Frederik. "Demystifying the Draft EU AI Act." Computer Law Review International, 2021. arXiv
Schuett, Jonas et al. "Towards Best Practices in AGI Safety and Governance." 2023. arXiv

General

Regulatory Capture in AI