I Asked Six Dead Thinkers to Audit the AI Industry
Frontier labs say they can govern themselves. I sat six dead thinkers around a table to check, and left an empty chair for the one who wouldn't come.
For about a year and a half I've been reading AI-industry writing with a specific kind of frustration. The pieces are usually well-written and rarely naive about the technology. But they read, over and over, as if governing frontier labs were a brand new kind of problem. As if nobody had ever tried to write rules for an industry whose participants know more than its regulators, or manage a commons whose users are incentivized to overuse it, or hang a dashboard in front of a bureaucracy that is systematically blind to the damage it's doing.
All of those problems have literatures. The people who wrote those literatures are, almost without exception, dead. So I staged an essay called Seeing Like an AI Company as a seminar room.
Six twentieth-century thinkers sit at the table: Susan Leigh Star, James C. Scott, Elinor Ostrom, Karl Polanyi, Ursula Franklin, John Kenneth Galbraith. A seventh chair stays empty, because Friedrich Hayek would not have come. Instead he sends four sealed dissents the reader can open as the Council proceeds. Eight sessions. A wall at the back where arguments pile up as colored index cards. No consensus at the end.
Why a Seminar Room and Not a White Paper
The format isn't a gimmick. Or it is a gimmick, but a load-bearing one.
When people cite Scott in AI-policy writing, they almost always cite him once, reach for the word legibility, and move on. Cite Ostrom, reach for commons. Cite Star, they usually don't, because she's harder. Each citation works like a stamp: you've signaled you read the thing, now you can proceed.
These thinkers weren't writing stamps. They were writing arguments that answered each other and their moment. Ostrom's design principles were a direct reply to Hardin's Tragedy of the Commons. Star's infrastructure studies came out of a thirty-year fight with the assumption that categories are neutral. Scott spent Seeing Like a State refusing, case by case, the idea that bureaucratic simplification is an improvement on local knowledge.
Putting them in a room forces the citations to become exchanges. Scott has to answer Ostrom's question about monitoring. Franklin has to answer Galbraith on countervailing power. Hayek's envelopes arrive at inconvenient moments, usually right when the Council has talked itself into a consensus it hasn't earned. That cross-examination is exactly what the AI-industry conversation has never had.
Ostrom's Scorecard
The third session is the one that changed how I think about the industry.
Ostrom spent her career documenting the conditions under which a shared resource, a fishery, a forest, a groundwater basin, gets managed well over the long run. Her eight design principles aren't aspirations. They're what shows up empirically in communities where the commons has survived a century. Clear boundaries. Rules matched to local conditions. The affected people participating in rule-making. Effective monitoring. Graduated sanctions. Low-cost conflict resolution. External recognition of the right to self-organize. Nested governance for large systems.
Two labels there deserve specifics. Gestural describes the monitoring principle: red-teaming exists, evals exist, but the monitoring is run by the party being monitored, on dimensions that party defined, against benchmarks that party controls, and it doesn't catch the class of slow structural harms, labor displacement, epistemic pollution, institutional deskilling, that don't fit inside a pass/fail prompt. Compromised describes graduated sanctions: if the lab breaks its own policy, the lab decides the consequence. Ostrom's literature has a name for that, and it isn't self-regulation. It's no sanctions. The scorecard is where readers tend to go quiet.
The Legibility Trap
Scott's session comes later, and it's the one that turned the essay from an analysis into something closer to an argument.
His classic move: the state sees a forest as timber yield, which makes the forest legible and also makes it fragile. What gets lost when you classify a forest as yield is everything the forest was doing that wasn't yield. Understory, soil biota, the slow feedbacks that hold the system together. The plantation is more legible, more productive, and after forty years, sicker.
The red-ink phrase on the Wall from his session is Waldsterben, 2026: shorthand for a failure a dashboard structurally cannot catch, damage that only becomes visible after the feedbacks are already irreversible. A safety dashboard that reports what it can measure isn't a neutral instrument. It re-creates the legibility trap, a bureaucracy of its own, growing its own plantations.
Scott's real question isn't what should the dashboard measure? It's: does the existence of the dashboard change what counts as a measurable harm? The Council's tentative answer, late in the session, is yes. And that answer is what ties this essay to two running themes across the rest of the site: classification is infrastructure, and infrastructure decides what reality is allowed to look like.
Why Hayek Gets Envelopes
The design decision I'm proudest of is the one that took me longest to commit to.
Hayek doesn't sit at the table. He'd have refused the invitation; his disagreements with Polanyi alone would have eaten six sessions. But leaving him out entirely would have turned the document into a straw man. So he sends letters. Four sealed envelopes the reader can only open at specific moments in the proceedings.
The first is the one everyone expects, the fatal conceit, his argument that central planners can't possibly hold the knowledge their plans pretend to. The second, on freedom to exit. The third, on dispersed knowledge. Each lands right as the Council convinces itself it has the answer.
The fourth envelope is different. It can only be opened after adjournment, once you've watched the Council fail to reach consensus. I won't spoil it. I'll say it's the dissent I find hardest to answer, and it's not the one Hayek is famous for. The essay ends where it ends because the question genuinely stays open. No consensus. The Council disperses. Star takes the scorecard off the Wall, folds it, and slides it inside the document. Whatever you believed about AI governance walking in is what you carry out, only now with more company.