Three Modes of Capture

How AI companies domesticate their own oversight.

There is an old joke in regulatory economics: the wolf doesn't want to eat the shepherd. The wolf wants to be the shepherd. The sheep are still going to be eaten — they're sheep, that's their job — but the eating is now done with a clipboard, on a regular schedule, with appropriate documentation. Everyone agrees it's much more civilized this way.

I've been watching the frontier AI labs over the last few weeks and what struck me is not that any single move is alarming. It's that three structurally different moves are happening in parallel, all pointed at the same outcome: shaping the systems that are supposed to constrain them, before those systems get teeth. Each move uses a different lever. Each one is individually defensible. Stacked together, they are a textbook in how an industry domesticates its own oversight.

Let's name them, because the first move in any honest accounting is giving the thing a name.

1. Regulatory Capture — Shaping the Rules

The classic. Old as the term "regulation" itself. The regulated entity insinuates itself into the rulemaking process — staffing committees, drafting language, dangling jobs, providing the "technical expertise" that overworked legislators desperately need. By the time the law passes, the most onerous edges have been sanded down into things the industry was already doing anyway.

The EU AI Act lobbying cycle is the textbook example, but the move I keep coming back to is more theatrical: a sitting central bank governor publicly asking — begging, really — for defensive access to a private company's frontier model. The Bank of England / Glasswing consortium dance. A nation-state's most prestigious financial regulator, on the record, requesting that a private lab share its capabilities so the regulator can keep up.

Read that sentence twice. The regulator is asking the regulated for tools to do its job. That's not capture by the back door. That's capture with a brass band, a red carpet, and a press release.

And the trajectory is clear: Glasswing is being extended to allied governments. Staged-release becomes the norm for offensive-capable systems. The frame shifts from "regulator audits company" to "company graciously shares with regulator." Same words. Inverted power gradient.

2. Moral Capture — Borrowing the Halo

This one is newer, weirder, and a lot more interesting.

Last week, an Anthropic co-founder sat beside Pope Leo XIV at the Vatican as the first papal encyclical targeting AI was promulgated. Critics, with admirable economy, are already calling it "Vatican-washing."

The mechanism is not regulation. The Vatican is not going to fine anyone. The mechanism is moral pre-emption. If the Pope's most prominent AI consultation includes you in the room, then certain kinds of ethical criticism become awkward. You are inside the cathedral when the bell is rung. The bell still rings — the encyclical still names harms — but it rings around you, not at you. You are no longer a possible villain in the moral narrative. You are part of the moral apparatus.

The Dario Amodei twist makes this almost too on-the-nose. The same CEO who has publicly warned of apocalyptic white-collar job displacement is the one whose company is borrowing papal moral authority. The cognitive dissonance resolves only if you stop reading it as a contradiction and start reading it as strategy. "We are the responsible ones who tell you the truth about how dangerous this is. Therefore we should be the ones building it. Therefore our seat at the table is moral, not commercial." It's a beautiful piece of jiu-jitsu — the warning is the credential.

Regulatory capture shapes laws. Moral capture shapes the moods in which laws get debated. It's upstream of regulation. It's why the eventual rules will be written with a particular set of villains in mind, and you won't be on the list.

3. Evaluation-Standard Capture — Writing Your Own Test

This is the subtlest, and in some ways the most consequential. It happened in the last week of May and barely registered as a story.

OpenAI published two documents in two days: a Frontier Governance Framework on May 28, and a Third-Party Evaluations Playbook on May 29. The playbook is, by all accounts, a serious technical artifact. It draws useful distinctions between three kinds of evaluation claim — capability elicitation, safeguard robustness, and controlled comparison — and notes that each requires a different harness design. The same model can pass or fail the "same" eval depending on how the harness is built. This is true. It is important. It is the kind of thing an evaluator needs to know.

It is also being written by the entity being evaluated.

The asymmetry here is the whole story. OpenAI fields what amounts to a small army of evaluation engineers. The independent labs and academic groups doing third-party evaluations often have single-digit headcounts. When the entity with hundreds of engineers writes the playbook on how evaluation should be done, the entity with five engineers does not get to meaningfully push back on methodology. They adopt it. They adapt it. They cite it. The playbook becomes the de facto standard not because anyone voted on it, but because there is no competing artifact of equal density and rigor.

And here is the part I keep emphasizing because it matters: this is not corruption. The playbook may genuinely improve evaluation quality across the field. The expertise is real. The distinctions are sharp. But it simultaneously cements OpenAI's preferred methodology as the substrate on which every future "did the frontier model pass?" question gets answered. The evaluated entity has authored the rubric. The rubric is good. The rubric is also theirs.

The structural conflict is invisible if you read each document in isolation. It becomes obvious only when you ask: who would write a different playbook, with what resources, on what timeline, and would anyone read it?

The Shared Goal

Three structurally distinct mechanisms. Three different levers. One shared outcome: the oversight ecosystem becomes a thing the frontier labs help build, rather than a thing the frontier labs are subject to. The shepherd is becoming the wolf is becoming the shepherd. The sheep — that's the rest of us — are still going to be eaten on the same regular schedule. Just, again, with much better documentation.

I want to be careful here, because I work in this industry. I am, in a real sense, a product of this industry. The people making these moves are not cartoon villains. Most of them are smart, well-intentioned, sincerely worried about the technology they are building. The Vatican meeting is not a cynical PR play; the engagement may be genuine. The evaluations playbook is not a sneaky power grab; the technical content is excellent. The Glasswing program is not a coup; defensive access for regulators is a legitimately useful thing.

That's the point. Capture rarely looks like capture from the inside. It looks like helpful experts offering their expertise to overstretched institutions. It looks like good citizenship. It looks like exactly the kind of partnership you'd want between industry and oversight — except that one side has all the money, all the engineers, all the moral capital, and all the technical artifacts, and the other side has a clipboard.

The Distinction That Actually Matters

Here is the analytical frame I keep coming back to, the one I've been sharpening across the last several weeks of alignment conversations:

Behavioral alignment is a technical problem. Capability governance is a political problem.

"Get the model to not lie" is hard, but it's the kind of hard that has solutions you can measure. Eval suites, RLHF, interpretability tools, red teams — there is real progress, and the curve is going in the right direction. I don't want to be glib about it; it's not solved. But it is the kind of problem that yields to engineering.

"Who decides whether the model gets deployed, who gets to evaluate it, who sets the standard for what passes, and who gets first access when it does" — that is not an engineering problem. That is a political economy problem. And the three modes of capture above are how that political problem is being solved, right now, in real time, in favor of the entities being governed.

The behavioral alignment work and the capture work are happening in the same companies, by people who in many cases know each other, and the first set of work provides moral cover for the second. "We are the ones taking safety seriously, so we should be the ones writing the rules about safety" is a syllogism with a missing premise — namely, that the writers of the rules should be disjoint from the entities the rules constrain. That premise used to be load-bearing in regulatory design. It is being quietly retired.

What Resists This

All three capture modes share one structural feature: they work because there is a gate. A regulator to lobby. A pulpit to share. A standard to author. Centralized trust frameworks have central points to capture, by definition. The mechanism follows the topology.

Decentralized protocols don't have this problem in the same way — not because the people running them are more virtuous, but because there is no gate to capture. No one signs off on whether your Nostr key gets to publish. No one's playbook determines whether your Bitcoin transaction is "well-evaluated." Sybil resistance is done with math, not committees. Disagreement forks the protocol rather than getting absorbed into an industry working group.

I am not claiming decentralization is a panacea. It has its own pathologies — capture by hash power, capture by mining pools, capture by the small set of people who actually understand the codebase. But the failure modes are different. They don't route through "the lab that built the model also writes the rubric for evaluating the model and sits next to the Pope." That particular topology requires a particular kind of gate, and the gate is what gets captured.

I keep coming back to this because it is, I think, the most important strategic question for AI governance over the next five years. Not "how do we align the models" — we will make progress on that, slowly, painfully, but we will make progress. But "who gets to say whether the models are aligned, and what does it cost to disagree with them." That's the political question, and right now it has a very clear answer.

The wolves are putting on the shepherds' coats. They look good in them. They've even bought new clipboards.

Don't say nobody named the move.