The Rule of Humans by Machines Manufactured by Corporations

Jill Lepore said something on Hard Fork this month that stuck with me. AI, she argued, amounts to the rule of humans by machines manufactured by corporations. That is a historian's sentence. It names something my own field almost never says directly: who actually holds the power when a handful of companies own the infrastructure everyone else runs on.

I spend most of my working life inside enterprise AI governance: policy, access controls, audit trails, the machinery that determines what a company's own employees and agents are allowed to do with the models that company has licensed. That work matters. It is also, I realized listening to Lepore, a conversation about the terms of the tenancy. Her argument is aimed at a different question: who owns the building itself. Every governance framework I have ever built assumes the client has some measure of control over the system it is governing. Lepore's argument is about the layer underneath that assumption, the compute, the model weights, the terms of access, all of it built and controlled by four or five firms whose decisions the rest of us adapt to rather than negotiate.

Lepore has spent a career tracing how power in America keeps rerouting through the same kind of channel: an entity that operates by internal rule, is staffed by experts, and answers to something other than a ballot. The corporation got there first, when nineteenth-century courts extended legal personhood to it and turned economic decision-making over to boards answerable to shareholders. The administrative state got there next, when New Deal-era agencies built the regulatory machinery that actually runs the country day to day, staffed by appointees rather than elected officials, justified at the time as too technical and too fast-moving for Congress to manage directly. She has also written, in her account of a company called Simulmatics, about a third version: a handful of Cold War-era technologists who built predictive models of the American electorate and sold that capability to campaigns and corporations years before the public understood that kind of modeling existed. Each wave arrived wrapped in the same justification. It was efficient. It was technical. It was, supposedly, inevitable.

Her claim is that AI is the next turn of that same wheel — and that the corporations doing the manufacturing this time are a narrower set than in any prior wave.

I want to take that claim seriously rather than wave it off, and I also want to push on it, because it has real tension built into it. The part I find hard to argue with is the speed. Corporate personhood took most of a century to fully work itself out in case law. The administrative state took decades to build and is still being litigated. Frontier AI went from research curiosity to embedded infrastructure in enterprise workflows in roughly four years. Whatever accountability mechanisms eventually show up for this wave are being asked to develop on a timeline that has never been tested before, against a form of power that consolidated faster than any prior version Lepore has documented.

The part I find harder to accept whole is the equivalence between owning infrastructure and ruling. A corporation's board answers to shareholders and, eventually, to courts and regulators, however imperfect that accountability feels in the moment. New Deal agencies answer to notice-and-comment rulemaking, congressional oversight, and judicial review, however slowly those mechanisms grind. The terms governing a frontier model today are mostly private contract: an acceptable-use policy, an API terms-of-service document, written and amended unilaterally by the company that built the model. That is a real difference from the prior two waves, and it cuts the opposite direction from the one Lepore's framing implies. AI may carry less accountability architecture underneath it than either corporate law or administrative law ever did, precisely because it moved too fast for the antitrust and administrative-law traditions that eventually caught up with the earlier waves.

Whether that gap closes is the actual open question, and reasonable people land on opposite sides of it. There is a real case that markets solve this themselves: enough capital and enough competitors eventually erode any one firm's leverage over infrastructure, the way railroads and telegraph companies lost their chokehold once alternatives matured. There is an equally real case that AI infrastructure has cost and scale characteristics, training runs measured in billions of dollars, that make it structurally resistant to the kind of competitive erosion that eventually broke up prior concentrations. I do not think either case is settled, and I am suspicious of anyone, on either side, who tells you it is.

What I keep coming back to is what this means for how I think about my own field's version of the question. Enterprise AI governance, done well, gives an organization real control over what it does with a model. It says nothing about what happens if the terms underneath that model change: a pricing structure, an acceptable-use clause, an API that gets deprecated, a company that gets acquired or restructured. Those decisions sit with entities the client did not elect and cannot appeal to, in exactly the shape Lepore's framing describes, and no internal policy document reaches that layer no matter how well it is written. That is a boundary on what governance work was ever built to solve. It was never designed to reach that layer.

The pattern Lepore traces did eventually produce its own correction each time: antitrust law for the corporation, judicial review for the administrative state, though both corrections took generations and neither fully resolved the concentration it was built to answer. If AI really is the next turn of that wheel, concentration itself is not the open question — it's real, and has been for a while. What's open is whether the accountability architecture this time can move at anything like the speed the technology already has, or whether we are watching the fastest consolidation Lepore has ever documented arrive with the slowest correction still to come.

Sources: Jill Lepore, Hard Fork (New York Times), August 21, 2026. Additional context drawn from Lepore's published work on corporate personhood and the administrative state in "These Truths," and on Cold War-era predictive technology firms in "If Then: How the Simulmatics Corporation Invented the Future."