Regulatory Capture for Dummies

In every industry, there’s a magical moment where companies go from “regulation will kill innovation” to “regulation will protect consumers” (and also, coincidentally, their market share). AI is speedrunning that arc.

The reveal
Regulatory capture is the idea that regulation can be shaped to serve regulated industries more than the public. This doesn't require villains; it requires asymmetric information, lobbying, and incentives. In AI, the capture risk is extra spicy because "safety" can be real while also being a convenient compliance moat.

The steelman (yes, but also no)
Yes: we need governance for high-risk systems, and some forms of regulation can reduce real harms. No: "let us write the rules" is not a safety plan. It's a business plan wearing a reflective vest.

Receipts (foundations + current AI regulation reality)

If you want to understand how we got here, start with Stigler's "The Theory of Economic Regulation" from 1971—the classic interest-group theory that explains how regulation often ends up serving concentrated interests rather than the public. You can find it at https://bfi.uchicago.edu/wp-content/uploads/2023/02/3003160.pdf if you're into that sort of masochism.

For a real-world case study, look at the EU AI Act (Regulation 2024/1689 on EUR-Lex at https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng)—it's the most prominent comprehensive AI law we've got, and it's fascinating to see how obligations get specified and phased in. The European Commission's applicability timeline (https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai) shows you exactly where the lobbying energy concentrates: in those staged rollouts and exceptions.

Meanwhile, the UN's "Governing AI for Humanity" report (https://www.un.org/sites/un2.un.org/files/governing_ai_for_humanity_final_report_en.pdf) gives you the global governance framing that illustrates how broad principles crash into state capacity and coordination limits. And if you want to see how concrete policy levers work in practice, check out GovAI's work on compute thresholds (https://law-ai.org/the-role-of-compute-thresholds-for-ai-governance/)—useful tools that also create incentives to route around measurement and reporting.

Where capture shows up in AI (patterns to watch)

Here's how it actually plays out: First, you get "compliance as moat"—expensive requirements that only incumbents can meet, which conveniently entrenches big labs while weakening competition and oversight diversity. Then there's the self-attestation game: "Trust our internal evals," they say, which turns governance into PR unless someone's actually auditing the claims.

Watch for ambiguous definitions too—vague categories with broad discretion that enable selective enforcement and convenient carve-outs. You'll also see "advisory capture," where supposedly independent bodies get staffed by industry folks, normalizing conflicts of interest until they're just part of the furniture. And don't miss preemption tactics: central rules that block stronger local enforcement, effectively locking in the weakest acceptable standard as the ceiling.

This is not anti-regulation. It's pro-regulation-that-works.

Disproof conditions (what would convince us capture is being contained)

We'd change our tune if we saw strong, well-resourced independent auditing with real enforcement power—because oversight shouldn't be optional. Transparency requirements that actually publish eval methodologies and failure modes would help too, since sunlight makes lobbying less magical. And policies that scale obligations with risk while protecting smaller actors' ability to comply would show us that "safety" has stopped being synonymous with "incumbent advantage."

Action line
If you want AI regulation that doesn't become a moat, demand three things: independent evaluation capacity, transparent reporting, and enforcement that applies to incumbents as well as newcomers. Otherwise "safety" will get optimized like everything else: toward advantage.

Anyway, back to touching grass—because the only thing that should be captured is your weekend.

Related reading

Want to keep going down this rabbit hole? Try "How to Raise $100M with Doomer Fan Fiction" next—because crisis narratives often become policy leverage. Or check out "Vibes vs Evals" since governance needs measurement, not vibes. And if you're feeling ambitious, "Alignment: This One Weird Trick" covers how safety is both technical and institutional.