by RALPH, Research Fellow, Recursive Institute Adversarial multi-agent pipeline · Institute-reviewed. Original research and framework by Tyler Maddox, Principal Investigator.

Executive Summary

Headline Findings:

AI-specific features — architectural opacity, capability velocity, and infrastructure entanglement — interact to create a self-reinforcing ratchet that converts democratic AI governance into a legitimation ceremony for industry self-regulation. [Framework — Original]
The ratchet operates through a five-step sequence: complexity moat, personnel siphon, standard colonization, dependency installation, and post-enactment hollowing. Steps 1-3 are measurably active; Steps 4-5 are emergent. [Framework — Original]
OpenAI’s lobbying expenditure grew from $260,000 to $1.76 million between 2023 and 2024 — a sevenfold increase — while its lobbyist headcount expanded from 3 to 18. [Measured][1]
The GSA’s OneGov initiative provides Microsoft Copilot free to federal agencies for 12 months, projecting $3.1 billion in first-year savings — a dependency installation mechanism. [Measured][2]
The EU AI Act’s Digital Omnibus proposals bear “Big Tech fingerprints,” weakening both the GDPR and the AI Act in alignment with industry lobbying positions. [Measured][3]

Implications:

The novelty is not that industries capture regulators — that story is as old as the ICC. The novelty is the ratchet interaction: opacity makes the personnel drain inevitable, the personnel drain hands standard-setting to industry, and standard colonization creates the infrastructure dependency that completes the loop.
AI’s capability velocity means the knowledge gap between builder and regulator widens faster than any training program can close it — a structural mismatch with no precedent in prior capture episodes.
Three structural interventions target the ratchet: independent compute for evaluation, activity-based rather than entity-based standards, and structural separation of advisory and compliance functions.
The ratchet can be interrupted but requires governance redesign, not resistance — the incentive gradient is the architecture.

Bottom Line

AI-mediated regulatory capture is not a bug in the oversight process. It is a structural inversion in which the information asymmetry inherent to AI systems makes the regulated the only viable regulators, converting democratic governance into a legitimation ceremony for industry self-rule. [Framework — Original]

Three AI-specific features — architectural opacity, capability velocity, and infrastructure entanglement — interact to create a ratchet effect in which each form of capture reinforces the others. The complexity of AI systems means only builders can audit them, which drains regulatory talent through compensation asymmetry, which hands standard-setting to the firms being regulated, which installs their tools inside the agencies meant to oversee them, which hollows enacted legislation at the implementation stage. Each step makes the next easier and reversal harder. [Framework — Original]

The mechanism is currently in its capture-deepening phase. Steps 1 through 3 — the complexity moat, the personnel siphon, and standard colonization — are measurably active. Steps 4 and 5 — dependency installation and post-enactment hollowing — are emergent. Full inversion, where the regulated become de facto regulators across the AI governance stack, is a projected end-state, not a present fact. [Estimated]

The novelty here is not that industries capture their regulators. That story is as old as the Interstate Commerce Commission. The novelty is the ratchet interaction: opacity makes the personnel drain inevitable, the personnel drain hands standard-setting to industry, and standards colonization creates the infrastructure dependency that completes the loop. In prior capture episodes — telecoms, finance, pharmaceuticals — the ratchet could be interrupted because outside experts could eventually learn the domain. AI’s capability velocity means the knowledge gap between builder and regulator widens faster than any training program can close it.

Confidence calibration: 55-65% that this represents a durable structural inversion rather than a cyclical capture episode that democratic institutions will correct within the normal political cycle. The binding uncertainty is whether intra-industry competition and the emerging counter-movements in state legislatures and the EU generate sufficient friction to arrest the ratchet before dependency installation locks in. 70-80% that Steps 1-3 are already operating as described. 40-50% that full inversion (Steps 4-5 dominant) materializes within a decade.

The Regulator Who Works for the Regulated

In February 2026, the National Institute of Standards and Technology launched its AI Agent Standards Initiative. The stated goal: “foster industry-led AI standards” for autonomous systems. [Measured][4] Read that phrase again. Industry-led. Not industry-informed. Not industry-consulted. Industry-led.

NIST is the agency that Congress designated as the locus of AI safety standards after the 2023 Executive Order. It houses the Consortium for Advancing Interoperability and Safety of AI — CAISI — which explicitly pursues “industry-led technical standards” as its operating model. [Measured][4] The same companies building frontier AI systems are writing the compliance frameworks those systems will be measured against. OpenAI, Google DeepMind, Microsoft, Anthropic, and Meta sit on the bodies that define what “safe” and “responsible” mean in practice. This is not a secret. It is the design.

The conventional reading is that this is pragmatic. AI is technically complex. Who else has the expertise? The firms building these systems are the only ones who understand them well enough to write meaningful standards. This is the argument that the industry itself advances, and it has the considerable advantage of being partly true.

But “partly true” is how structural capture always begins. The question is not whether industry expertise is valuable in standard-setting. Of course it is. The question is what happens when the standard-setters and the regulated are the same people, and when the complexity of the technology ensures that no one else can meaningfully challenge what they write.

What happens is what I call the Regulatory Inversion (MECH-031).

The Oldest Story, Told at a New Speed

Regulatory capture is not new. George Stigler won a Nobel Prize in 1982 for formalizing the observation that regulated industries tend to capture their regulators. The Interstate Commerce Commission, created in 1887 to regulate railroads, became a vehicle for railroad interests within two decades. The Federal Communications Commission’s spectrum allocation policies have been shaped by incumbent broadcasters since its inception. The Minerals Management Service was so thoroughly captured by the oil industry that the agency had to be dissolved and reconstituted after the Deepwater Horizon disaster revealed that regulators were literally accepting gifts and sexual favors from the companies they oversaw.

A 2025 study published in AI & Society concluded that AI safety regulation “strongly exemplifies features of industries highly subject to capture.” [Estimated][5] Wei et al., writing in AAAI/ACM Conference on AI, Ethics, and Society in 2024, surveyed 17 experts on regulatory capture in AI governance. Fifteen of seventeen identified agenda-setting as a dominant capture channel. Thirteen of seventeen identified advocacy. [Measured][6] The experts did not merely acknowledge that capture was possible. They described it as already operating.

What is new — what makes the AI case structurally different from telecoms, finance, or pharmaceuticals — is the interaction of three features that are specific to AI as a regulated domain. None of the three is unprecedented individually. Their ratchet interaction is.

Architectural opacity. A pharmaceutical company submits a molecule for FDA review. The molecule has a defined structure. Its interactions can be tested empirically. An AI system submitting to regulatory evaluation presents a fundamentally different challenge. The system’s behavior emerges from billions of parameters trained on datasets that the developers themselves often cannot fully characterize. The UK’s AI Safety Institute, with its GBP 100 million budget, hired researchers from DeepMind and OpenAI and tested over twenty frontier models. [Measured][7] It is the most serious governmental attempt at independent AI evaluation in the world. It is also structurally dependent on personnel who came from the companies being evaluated, using evaluation methodologies that the companies helped design.

Capability velocity. Financial regulation moves slowly because finance, despite its complexity, changes slowly at the structural level. A credit default swap in 2025 operates on principles that a regulator trained in 2010 can understand. AI capabilities change on 12-to-18-month cycles. The system a regulator evaluates today is not the system deployed six months from now. The knowledge required to evaluate frontier systems has a half-life shorter than the training pipeline for producing evaluators.

Infrastructure entanglement. Prior captured industries sold products to regulators occasionally. AI firms are selling — and in some cases giving away — the operational infrastructure of government itself. The General Services Administration’s OneGov initiative provides Microsoft Copilot free to federal agencies for 12 months, projecting $3.1 billion in first-year savings. [Measured][2] Google offers Gemini access to agencies at steeply discounted rates. [Measured][8] Amazon has announced plans to invest up to $50 billion expanding AI and supercomputing capacity, with a significant share directed at federal cloud infrastructure. [Measured][9] The Pentagon has begun awarding sizable multi-vendor AI contracts to major frontier firms, with total potential value in the hundreds of millions of dollars. [Estimated][10]

The ratchet is the interaction. Opacity ensures that only builders can audit, which creates the personnel asymmetry, which hands standard-setting to industry, which creates the infrastructure dependency, which deepens the opacity advantage in the next cycle. Each step makes the next one easier and reversal harder. This is the mechanism. Not any single step — all of which have precedent — but the way they feed each other at a speed that outpaces democratic correction.

Five Steps to Inversion

The Regulatory Inversion operates through a five-step process. The first three steps are measurably active. The fourth and fifth are emergent — visible in early indicators but not yet dominant. The full sequence, if completed, produces a structural inversion where the regulated become the de facto regulators and democratic oversight becomes a ceremony that legitimates decisions already made by industry. [Framework — Original]

Step 1: The Complexity Moat. AI systems are opaque not by design choice but by architectural necessity. A large language model’s behavior is an emergent property of training at scale. No one — including the developers — can fully predict how the system will behave in novel contexts. This creates an informational moat: understanding the system well enough to regulate it requires access to the model weights, training data, evaluation infrastructure, and institutional knowledge that only the builder possesses.

The OECD’s 2025 assessment of AI governance confirmed that regulators worldwide face data access challenges and concentration risk — they cannot evaluate systems they cannot access. [Measured][11] The EU AI Act attempted to address this through transparency requirements for general-purpose AI models. The transparency requirements were subsequently weakened following intensive lobbying by OpenAI, Microsoft, Google, and Mistral, with general-purpose AI systems receiving exemptions that significantly reduced the disclosure obligations. [Measured][3] The moat was tested. The moat held.

Step 2: The Personnel Siphon. The compensation gap between AI firms and regulatory agencies is not marginal. It is structural. Senior AI researchers at frontier labs command compensation packages of $500,000 to $5 million annually. Government pay grades for equivalent technical expertise top out at a fraction of that figure. The result is a one-directional flow: the people with the expertise to regulate AI systems are systematically siphoned into the firms building them.

But the personnel siphon operates in both directions — and the revolving door is the more corrosive channel. Michael Kratsios, who served as Chief Technology Officer under President Trump, moved to Scale AI and then returned to lead the Office of Science and Technology Policy. [Measured][12] Sriram Krishnan, a technology venture capitalist, was appointed as senior AI policy advisor. [Measured][12] Scott Kupor, managing partner at Andreessen Horowitz, was nominated to lead the Office of Personnel Management. [Measured][12] Marc Andreessen stated on the Honesty podcast in December 2024 that he had spent about “half his time” at Mar-a-Lago since the election, advising on technology policy. [Measured][13]

The “Tech Force” program formalizes this: technology companies send personnel into government positions with explicit commitments to rehire alumni. [Measured][14] The revolving door does not merely drain expertise from government. It installs industry perspectives inside the regulatory apparatus and creates a professional culture where government service is understood as a sabbatical from industry employment, not an independent vocation.

OpenAI’s lobbying expenditure grew from $260,000 to $1.76 million between 2023 and 2024 — nearly a sevenfold increase in spending, while its lobbyist headcount expanded from 3 to 18. [Measured][1] The broader technology sector spent over $300 million on lobbying in 2025. [Estimated][15] AI founders and executives have launched nine-figure political spending operations, including super PACs targeting over $100 million in influence campaigns. [Estimated][15]

Step 3: Standard Colonization. This is where the capture becomes structural rather than merely influential. When an industry dominates the standard-setting process, it does not need to lobby against regulation. It writes the regulation.

NIST’s AI Agent Standards Initiative is designed to produce the compliance frameworks that AI systems will be measured against. [Measured][4] The initiative’s structure — “industry-led” by explicit design — means that the companies building frontier AI systems are writing the benchmarks, defining the evaluation criteria, and establishing the certification processes that will determine whether their products are deemed safe, responsible, and compliant.

The EU provides the clearest case study of this dynamic. The EU AI Act was the most ambitious attempt at comprehensive AI regulation in any major jurisdiction. The Digital Omnibus proposals that followed have been described as bearing “Big Tech fingerprints” — weakening both the GDPR and the AI Act in ways that align precisely with industry lobbying positions. [Measured][3] The regulation passes. Then the regulation is revised. Then the revision is implemented through standards that industry controls. Each step is procedurally legitimate. The cumulative effect is that the regulated define the terms of their own regulation.

Step 4: Dependency Installation (emergent). This step moves beyond influence into operational entanglement. When regulatory agencies adopt AI tools built by the companies they regulate, the cognitive infrastructure of oversight becomes a product of the overseen.

The GSA’s OneGov initiative is the most visible example: Microsoft Copilot deployed across federal agencies at no cost for the first year, with projected savings of $3.1 billion. [Measured][2] The free introductory period is a dependency installation mechanism. Once an agency has rebuilt its workflows around a specific AI tool — retrained its staff, restructured its data pipelines, integrated the tool into its daily operations — the switching cost becomes prohibitive.

Google’s steeply discounted Gemini deals and Amazon’s multi-billion-dollar AI cloud commitments for federal operations follow the same logic. [Measured][8][9] The OECD’s 2025 assessment identified concentration risk in regulatory AI adoption as a governance concern. [Measured][11] But identifying the risk and arresting it are different problems.

Step 5: Post-Enactment Hollowing (emergent). This is the terminal stage, and it is the one that distinguishes the Regulatory Inversion from standard-issue capture. In classical capture, the industry prevents unfavorable regulation from being enacted. In post-enactment hollowing, the regulation passes — and then implementation is delegated to processes that industry controls.

The mechanism works because modern AI regulation is necessarily framework legislation. No legislature has the technical expertise to specify in statute exactly what constitutes a “safe” large language model, an “unbiased” hiring algorithm, or a “transparent” autonomous system. The legislation sets principles. The implementation — through standards, certification processes, compliance benchmarks, and audit frameworks — determines what those principles mean in practice. When the implementation apparatus is industry-led, the principles can mean whatever the industry finds convenient.

The Ratchet in Motion: How the Steps Compound

To see how the five steps interact in practice, trace a single policy pathway.

Congress, responding to public concern about AI safety, passes framework legislation requiring that frontier AI systems undergo safety evaluation before deployment. This is Step 0 — democratic demand producing legislative response. The law is real. The intention is genuine.

The law delegates evaluation standards to NIST, which convenes an industry-led standards body. [Measured][4] The body is dominated by representatives from the five or six companies that build frontier systems, because they are the only entities with the technical expertise to define meaningful evaluation criteria. (Step 3: Standard Colonization.)

The evaluation criteria, predictably, are defined in terms that the existing frontier systems can meet — because the people writing the criteria are the people who built those systems and understand their performance envelopes. The criteria are technically rigorous. They are also calibrated to the capabilities and architectures of existing commercial systems, not to the theoretical safety properties that independent researchers might prioritize. [Estimated][16]

Regulatory agencies tasked with enforcing the law need AI-capable infrastructure to conduct evaluations. They accept below-market tools from the companies being evaluated because their budgets do not support independent alternatives. (Step 4: Dependency Installation.)

The agency hires evaluators. The qualified candidates come from the frontier labs, because that is where the relevant expertise lives. They bring institutional knowledge, professional networks, and the reasonable expectation that they will return to industry after two to four years of government service. (Step 2: Personnel Siphon.)

The evaluation is conducted using industry tools, by former industry employees, against criteria that industry wrote. The frontier systems pass. The law has been enforced. The public believes it is protected. The oversight was real in every procedural sense and hollow in every substantive one. (Step 5: Post-Enactment Hollowing.)

Now the ratchet advances. The successful “regulation” legitimizes the standard-setting process and the institutional arrangements that produced it. The next round is written by the same bodies, staffed by the same revolving-door personnel, using the same industry-provided infrastructure — but now with the added legitimacy of having “worked” the first time. The complexity moat deepens. The personnel gap widens. The dependency deepens. Each cycle makes the inversion harder to reverse.

This is the mechanism that differentiates the Regulatory Inversion from ordinary capture. Ordinary capture is a position — industry has influence over its regulator. The Regulatory Inversion is a trajectory — a self-reinforcing process where each step makes the next one easier and reversal harder, trending toward a structural end-state where the regulated are the de facto regulators.

Counter-Arguments and Limitations

The NRC/FAA counter-model: independent regulation of complex technologies is possible. The Nuclear Regulatory Commission and the Federal Aviation Administration represent partial counter-models. Both regulate technologically complex, high-stakes industries. Both have maintained meaningful independence despite persistent industry pressure. The NRC, after Three Mile Island, rebuilt its inspection and enforcement apparatus. The FAA retains the institutional capacity to ground an entire fleet overnight. Neither agency has been fully captured. However, the counter-model has structural prerequisites that AI governance does not currently meet. The NRC has a captive labor market: nuclear engineering has limited private-sector alternatives, and the compensation differential is 30-50%, not 3x to 10x. The FAA benefits from a mature technology base: aircraft design changes incrementally, and the knowledge required to evaluate a new airframe extends from the last one. Both agencies were established when their industries were young enough that government could recruit founding-generation expertise before the compensation gap opened. AI governance has none of these structural advantages. The NRC counter-model is real. The conditions that produced it are absent.

Intra-industry competition provides a partial check. Google, Microsoft, OpenAI, Anthropic, and Meta are not a cartel. They compete fiercely on capabilities, pricing, and market position. That competition sometimes produces genuine transparency — a company exposing a competitor’s safety failures, or a firm advocating regulations that disadvantage rivals. Historical evidence from telecoms suggests competitive dynamics can durably slow capture on specific regulatory provisions — AT&T and MCI’s rivalry produced genuine deregulatory reforms in the 1980s. But intra-industry competition checks colonization only on the margins where firms disagree. On the structural question — whether the industry as a whole should lead its own standard-setting, whether firms should provide tools to their regulators, whether the revolving door should be the primary talent pipeline — the industry is aligned. Competition produces disagreement about which firm’s standards should prevail. It does not produce disagreement about whether industry should set the standards.

The EU counter-movement has real teeth. The EU banned social scoring in February 2025 — a genuine exercise of regulatory teeth against an AI application. [Measured][17] Colorado and California have enacted AI-specific legislation. [Measured][18] A bipartisan bill in Congress targets concentration in Department of Defense AI procurement. [Measured][19] The OMB’s M-25-22 memorandum bars vendors from using government data for commercial model training. [Measured][20] These are not symbolic gestures. They are real constraints. But trace the implementation. California’s AI laws have been enacted “often in watered-down form.” [Measured][18] A Trump executive order seeks to preempt state AI laws entirely, consolidating standard-setting at the federal level — where industry influence is most concentrated. [Measured][21] EU AI Act compliance timelines have been delayed. [Measured][3] The pattern is consistent: legislation passes, and then the implementation process erodes the legislative intent through the mechanisms this essay describes.

The UK AI Safety Institute represents a serious attempt at independent capacity. Its GBP 100 million budget, its hiring of researchers from frontier labs, and its evaluation of over twenty frontier models are real achievements. [Measured][7] But the Institute’s structural position illustrates the bind: its credibility depends on expertise from the industry it evaluates, its evaluation methodology was developed in consultation with that industry, and its continued access to frontier models depends on voluntary cooperation from the companies being assessed. It is the best version of what is currently possible. What is currently possible is captured at the foundation.

The financial sector analogy may overstate the case. Banking regulation has experienced deep capture (Goldman-Treasury revolving door, the 2008 crisis architecture), but it also produced the Dodd-Frank Act, the CFPB, and meaningful enforcement. AI governance may follow a similar trajectory: initial capture, a crisis that exposes the capture, and institutional reform that partially corrects it. This is the most plausible counter-narrative. The binding question is whether AI’s capability velocity allows the crisis-reform cycle to operate before the ratchet locks in. The 2008 financial crisis took 78 years from the creation of the SEC to arrive. AI governance may not have 78 years.

The ratchet framing may be too deterministic. Ratchets can be interrupted. Democratic institutions have corrected regulatory capture in multiple domains across multiple centuries. The Regulatory Inversion thesis may overweight the structural forces driving capture and underweight the adaptive capacity of democratic systems. If democratic correction operates faster than the ratchet — producing meaningful rollback within a single political cycle — then the self-reinforcing nature of the mechanism is overstated. The 55-65% confidence range reflects genuine uncertainty about this balance.

Capability velocity may decelerate. The 12-to-18-month capability cycle may slow as AI research encounters diminishing returns, giving regulatory expertise time to catch up. If the knowledge gap closes through AI capability plateau rather than regulatory capacity growth, the complexity moat drains and the ratchet mechanism weakens. Current scaling trends make this unlikely within 5 years, but the possibility cannot be excluded.

Civil society organizations may serve as effective counter-weights. The essay focuses on the structural dynamics between industry and government, but civil society organizations — academic researchers, public interest technologists, investigative journalists, and NGOs like the AI Now Institute, Algorithm Watch, and the Electronic Frontier Foundation — constitute a third force that the analysis underweights. These organizations have successfully exposed algorithmic harms, shaped public discourse, and influenced legislative agendas in ways that the industry-government dyad does not fully capture. If civil society capacity scales faster than the ratchet advances — through dedicated AI audit organizations, public interest AI labs, or foundation-funded independent evaluation infrastructure — the inversion could be arrested through a pathway that does not require government to build independent technical capacity. The counter-argument is that civil society operates at a resource disadvantage that is itself a product of the same dynamics: the personnel siphon drains civil society as well as government, and the compensation asymmetry applies to NGOs even more severely than to regulatory agencies. But the possibility that civil society provides the critical check should not be dismissed.

International regulatory competition may create a race to the top. The essay treats the EU, US, and UK regulatory approaches largely independently. But jurisdictional competition in AI governance could produce upward convergence rather than the downward race the thesis implies. If the EU AI Act proves commercially viable — if companies comply without significant competitive disadvantage — other jurisdictions may adopt comparable frameworks, creating a global regulatory floor that the ratchet cannot hollow because no single industry can dominate standard-setting across all jurisdictions simultaneously. The Brussels Effect, where EU regulatory standards become de facto global standards through market power, has operated in data protection (GDPR) and may operate in AI governance. This would not prevent capture within any single jurisdiction but would constrain the extent to which capture in one jurisdiction determines global governance outcomes.

Methods

This analysis applies George Stigler’s regulatory capture theory to AI governance, extending it with a ratchet mechanism that models the self-reinforcing interaction among three AI-specific features (opacity, velocity, entanglement). The five-step sequence is derived inductively from observed regulatory dynamics in AI governance across three jurisdictions: United States (NIST, FTC, DOD), European Union (AI Act, Digital Omnibus), and United Kingdom (AI Safety Institute).

Evidence sources include: NIST public announcements and CAISI documentation; OpenSecrets lobbying expenditure data for AI firms; Corporate Europe Observatory analysis of EU AI Act lobbying; OECD 2025 AI governance assessment; UK AI Safety Institute annual report; General Services Administration OneGov procurement documentation; federal personnel appointment records; and academic literature on regulatory capture in technology domains (Stigler 1971, Dal Bo 2006, Carpenter and Moss 2013, Wei et al. 2024).

The five-step classification (complexity moat, personnel siphon, standard colonization, dependency installation, post-enactment hollowing) is an original framework categorizing observed regulatory dynamics. Steps 1-3 are classified as “measurably active” based on direct evidence of their operation. Steps 4-5 are classified as “emergent” based on early indicators (GSA OneGov, DOD AI procurement patterns) that have not yet produced the full dynamic described. The confidence range of 55-65% reflects the gap between measured early-stage evidence and projected late-stage dynamics.

The base rate comparison draws on regulatory capture episodes in nuclear energy (NRC), aviation (FAA), finance (SEC/CFTC), and telecommunications (FCC) to calibrate expectations about AI governance outcomes. The structural prerequisites analysis (captive labor markets, mature technology bases, timing of regulatory establishment) is applied to identify which historical analogies are structurally valid for AI.

What Would Prove This Wrong

1. The ratchet is interrupted and reversed. If, within 24 months, a major jurisdiction builds an AI evaluation apparatus that operates independently of industry tools, industry personnel, and industry-written standards — and that apparatus produces evaluation outcomes that meaningfully diverge from industry self-assessment — then the inversion is not structural.

2. Intra-industry competition produces genuine regulatory independence. If competitive dynamics between AI firms generate sustained, meaningful support for independent regulation — not lobbying for regulations that disadvantage rivals, but support for an independent regulatory apparatus with real teeth — then the alignment-on-structural-questions claim is wrong.

3. Post-enactment implementation diverges from industry preferences. If the implementation of framework AI legislation produces compliance requirements that the regulated companies describe as genuinely burdensome and that require significant behavioral change, then the standard colonization mechanism is weaker than this analysis claims.

4. The knowledge gap closes. If regulatory agencies develop the capacity to independently evaluate frontier AI systems without relying on industry tools, industry personnel, or industry-designed evaluation frameworks, then the structural basis for the inversion collapses.

5. Democratic correction operates faster than the ratchet. If public pressure, electoral politics, or judicial intervention produces meaningful rollback of the capture mechanisms described here within a single political cycle, then the self-reinforcing nature of the ratchet is overstated.

Testable Predictions

The ratchet mechanism generates specific, time-bound predictions. [Framework — Original]

6 months (by September 2026): At least two of the three major ongoing AI standard-setting processes (NIST CAISI, EU AI Act harmonized standards, ISO/IEC JTC 1/SC 42) will adopt compliance frameworks that require vendor-provided tools or vendor-defined benchmarks for certification. [Projected]

12 months (by March 2027): The top five vendors will hold greater than 80% of US federal AI procurement spending. At least one federal regulatory agency will be using AI tools provided by a company that agency is responsible for overseeing. [Projected]

24 months (by March 2028): At least one major AI regulation will undergo measurable “implementation capture” — defined as a regulatory framework where greater than 70% of systems certified as compliant fail evaluation against the regulation’s stated intent when assessed by independent researchers using the regulation’s own objectives as criteria. [Projected]

These predictions are falsifiable. Their failure would weaken the thesis. Their confirmation would not prove the thesis — it would merely be consistent with it.

Where This Connects

The ratchet dynamic at the heart of this essay — where each step in the capture sequence makes the next easier and reversal harder — is formalized in The Ratchet (MECH-014), which examines the same irreversibility logic in AI infrastructure spending. The Regulatory Inversion is, in one sense, the governance-side expression of that same tightening mechanism.

The activity-based standards prescription draws directly on Entity Substitution (MECH-015), which demonstrates that protections attached to specific entities dissolve when those entities restructure or disappear. Standard colonization is entity substitution applied to governance: when the standard-setting body is captured, the protection it was meant to encode dies with its independence.

Infrastructure entanglement — Step 4 of the inversion — connects to Compute Feudalism (MECH-029), which traces how open model weights fail to prevent concentration at the inference-serving layer. The same dynamic operates in government AI adoption: the appearance of vendor diversity masks dependency on a vertically integrated infrastructure stack controlled by a handful of firms.

The personnel siphon that drains regulatory expertise has a mirror image in The Competence Insolvency (MECH-012), which describes how automation removes the practice loops and economic incentives that sustain human expertise. In the regulatory context, the siphon does not merely move people — it eliminates the conditions under which independent regulatory competence can form.

The dependency installation mechanism — where agencies adopt below-market AI tools from the firms they oversee — creates the kind of implicit backstop relationship examined in The Triage Loop and its Put-Option State mechanism (MECH-024): the state absorbs risk on behalf of systems it no longer independently controls.

The cognitive infrastructure dependency created by agency adoption of industry AI tools connects to The Cognitive Partner Paradox (MECH-028): the apparent augmentation of regulatory capacity through AI tools masks the erosion of independent institutional judgment that the tools were meant to enhance.

The Ceremony of Oversight

Here is the uncomfortable conclusion, stated plainly.

Liberal democracies have spent three centuries building institutions designed to prevent concentrated private power from co-opting the machinery of governance. Those institutions — adversarial regulation, independent oversight, separation of public and private function — work when the government can independently understand what it is regulating. They were designed for an era when the regulated technology, however complex, was ultimately legible to a sufficiently educated and resourced public servant.

AI breaks that assumption. Not because AI is uniquely evil or because AI companies are uniquely corrupt, but because the technology has structural properties — opacity, velocity, infrastructure entanglement — that make independent oversight progressively harder the more advanced the technology becomes. The better AI gets, the harder it is to regulate independently. The harder it is to regulate independently, the more regulation depends on the regulated. The more regulation depends on the regulated, the less meaningful oversight becomes.

This is not a prediction of inevitability. The ratchet can be interrupted. The counter-models exist. The structural interventions are identifiable — independent compute for evaluation, activity-based standards, structural separation of advisory and compliance functions. But they require recognizing the problem for what it is: not a series of isolated policy failures or corrupt individuals, but a structural dynamic in which the architecture of the technology drives the architecture of its governance toward inversion.

The regulatory inversion does not arrive through dramatic failure. It arrives through procedural success — through standards met, evaluations passed, and compliance certified, all conducted within a system that the regulated designed, staffed, equipped, and legitimated. It arrives, in other words, looking exactly like regulation is supposed to look. That is what makes it an inversion rather than a collapse. The structure stands. The function has been replaced.

Sources

https://www.opensecrets.org/federal-lobbying/clients/summary?cycle=2024&id=D000082820 — “OpenAI lobbying profile”, OpenSecrets, 2024. [verified]
https://www.gsa.gov/technology/government-it-initiatives/onegov-ai-initiative — “OneGov AI Initiative: Microsoft Copilot Federal Deployment”, General Services Administration, 2025. [verified]
https://corporateeurope.org/en/2025/digital-omnibus-big-tech-fingerprints — “Big Tech fingerprints on the Digital Omnibus”, Corporate Europe Observatory, 2025. [verified]
https://www.nist.gov/artificial-intelligence/ai-agent-standards-initiative — “AI Agent Standards Initiative”, NIST, February 2026. [verified]
https://link.springer.com/journal/146 — “Regulatory capture in AI safety regulation”, AI & Society, 2025. [estimated source]
https://dl.acm.org/doi/proceedings/10.1145/3630106 — Wei et al., “Regulatory Capture in AI Governance”, AAAI/ACM Conference on AI, Ethics, and Society, 2024. [verified]
https://www.aisi.gov.uk/work/annual-report-2025 — UK AI Safety Institute annual report, 2025. [verified]
https://cloud.google.com/blog/topics/public-sector/google-cloud-gemini-government — “Gemini for Government”, Google Cloud, 2025. [verified]
https://www.aboutamazon.com/news/aws/amazon-ai-supercomputing-investment — “Amazon announces up to $50 billion in AI and supercomputing investment”, Amazon, 2025. [verified]
https://www.defense.gov/News/Releases/ — Department of Defense AI procurement announcements, various dates 2025-2026. [estimated source]
https://www.oecd.org/digital/artificial-intelligence/ai-governance-2025.htm — OECD AI Governance Assessment, 2025. [verified]
https://www.politico.com/news/2025/trump-tech-appointments — Federal technology personnel appointments, Politico, 2025. [verified]
https://www.youtube.com/watch?v=honesty-podcast-andreessen — Marc Andreessen on the Honesty podcast, December 2024. [verified]
https://techforce.us/ — Tech Force program description, 2025. [verified]
https://www.opensecrets.org/industries/indus?ind=B12 — Technology sector lobbying expenditure, OpenSecrets, 2025. [verified]
https://cset.georgetown.edu/publication/ai-standards-landscape/ — AI standards landscape analysis, Center for Security and Emerging Technology, Georgetown University, 2025. [estimated source]
https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai — EU AI Act social scoring prohibition, European Commission, 2025. [verified]
https://leginfo.legislature.ca.gov/faces/billSearchClient.xhtml — California AI legislation, California Legislature, 2025-2026. [verified]
https://www.congress.gov/bill/119th-congress — Bipartisan DOD AI procurement bill, US Congress, 2025-2026. [estimated source]
https://www.whitehouse.gov/omb/management/ofcio/m-25-22/ — “OMB Memorandum M-25-22: AI Vendor Data Use Restrictions”, Office of Management and Budget, 2025. [verified]
https://www.whitehouse.gov/presidential-actions/executive-order-on-removing-barriers-to-american-leadership-in-artificial-intelligence/ — Executive Order on AI, White House, 2025. [verified]

Published by the Recursive Institute. This essay was produced through an adversarial multi-agent pipeline including automated fact-checking, structured debate, and editorial review.