There is a contradiction sitting inside the praise for Jill Lepore's new book, and it is worth naming before anyone gets swept up in the elegy.
Jaron Lanier calls her "our best scholar of the culture of governance." The Times Literary Supplement's Carl Benedikt Frey notes that she "is undoubtedly more interested in tracing the intellectual history of machine rule than in explaining its institutional underpinnings."
Both are correct. And the gap between them is the whole story.
Lepore's phrase for what she fears is six words long: the Artificial State is "government without consent, even government without humans." Read it slowly, because the important word is not "humans." It is "consent."
For two years the argument over artificial intelligence in government has been fought on the terrain of performance—will the model be accurate, will it be cheaper, will it be less biased than the tired official it replaces. Lepore, by reaching back to Hannah Arendt, quietly changes the subject to legitimacy.
That is her real contribution, and it arrives before any reviewer has praised or faulted a sentence. The question is not whether the machine decides well. It is whether a decision the machine makes is the kind of thing a free people can be said to have authorized at all.
She takes this move from Arendt's Origins of Totalitarianism, the 1951 book that argued the machinery of modern life was reshaping the fundamentals of human existence. Lepore reads the AI state as a descendant of that machinery—not a tool of self-rule but a solvent of it.
Notice how thoroughly this reframes the stakes. Efficiency is a virtue you can concede without losing your standing as a citizen. Consent is not. A government you did not authorize is illegitimate no matter how well it performs, and a government run by systems no one authored is one you cannot withdraw consent from, because there is no one there to withdraw it from.
Strikingly, an older authority arrives at the same threshold from an entirely different tradition. Pope Leo XIV, in his recent document on AI, warned of the risk of dehumanization—the Vatican positioning itself, as the BBC put it, as one of the most prominent global voices questioning the path the tech billionaires have set.
Two ancient authorities—the secular republican tradition and the Catholic Church—converging on the same worry: not that the machines will get the answer wrong, but that they will make the human act of answering disappear. That convergence is a signal. The legitimacy frame is not a quirk of one Harvard historian. It is where the serious objections keep landing.
Two ancient authorities converging on the same worry: not that the machines will get the answer wrong, but that they will make the human act of answering disappear.
So Lepore's diagnosis is right, and it is the diagnosis the policy debate needed. Then comes the turn.
Frey's review in the TLS is admiring and unsparing at once. The lessons Lepore draws, he writes, "come down to: do not surrender judgement to machines. As counsel, this has appeal. As a guide to explaining or reversing institutional decay, environmental damage and democratic decline, it is less satisfying."
The Guardian's Farrah Jarral lands the same blow more gently. Lepore ends the book with an exhortation to "unplug" and reconnect with nature and animals, and "prescribes nothing further." Good advice, Jarral allows—but "logging off to touch grass can only be the first step."
This is a fair charge, and it exposes precisely what the elegiac mode cannot supply. An elegy mourns a vanishing world. It does not tell you how to build the load-bearing beam that keeps the roof up.
The reframe from cost to consent is necessary. But "consent" cannot stay a mood. If it is to do any work, it has to be cashed out as institutional design—as an account of the conditions under which a specific human being can be made to answer for a specific machine's decision.
And that account exists. It just lives in the technical scholarship Lepore's narrative history skips.
Consider a Springer paper published this June with a title that reframes everything: "When responsibility fails to arise." Its argument is sharper than the usual accountability complaint. The standard worry is that in complex algorithmic systems, responsibility becomes hidden—hard to trace, distributed, opaque. Fix the transparency and you recover it.
The paper says no. Under optimization-driven governance, responsibility does not go into hiding. It fails to arise in the first place, because outcomes are produced without ever being constituted as the kind of act to which responsibility could attach.
Drawing on an externalist reading of Arendt—the same Arendt Lepore invokes—the author names three conditions any decision must meet before anyone can be held to answer for it: discreteness, authorship, and publicness. An outcome grounds responsibility only if it is a bounded act, attributable to an identifiable officeholder, and situated in a forum where it can be contested.
The case study is Australia's Robodebt scheme, in which automated debt notices "were issued at scale without identifiable decision events, attributable authors, or meaningful forums of contestation."
Sit with that description, because it is Lepore's "government without humans" rendered in operational terms. No moment you can point to and call the decision. No name attached to it. No room in which to argue back. The machinery did not decide badly. It arranged things so that no one decided at all—and therefore no one could be blamed.
The choice of Robodebt is deliberate and damning. It was deterministic and rule-based, not some inscrutable neural network. If responsibility can fail to arise in a system that simple, the problem is not exotic AI. It is the institutional form the decision was allowed to take. The paper notes the erosion only intensifies with generative systems.
Here, at last, is what the elegy could not give us. The paper closes with three concrete design principles: event-marking, authorship preservation, and deliberative apertures. Mark the moment a decision occurs so it becomes a discrete act. Attach a human author to it. Build an aperture through which it can be contested.
That is what consent looks like when you stop mourning it and start engineering it.
Melis Bas's work on technopolitical mediation, also drawing on Arendt, pushes the same conclusion from the other side. Technologies, Bas argues, do far more than exert power over individuals—they "condition, frame, create, and organize politics." Which means the political question is never merely whether to permit a system, but how its design shapes the forum in which people appear to one another as citizens.
Put Lepore, the Springer paper, and Bas in a room and the division of labor becomes obvious. Lepore establishes why the AI-governance fight is about consent rather than accuracy. The technical scholarship establishes what consent requires as a matter of institutional plumbing.
The elegy diagnoses. It cannot repair, because repair lives in exactly the register—design principles, failure modes, event-marking—that the mournful mode passes over on its way to the wildlife and the touch of grass.
This matters beyond one book review, because the legislative moment is arriving. The open question is whether the laws now being drafted will contain binding structural requirements or merely gestures.
The Springer paper draws the line cleanly. Most of the existing accountability toolkit—transparency, explainability, auditability, oversight—assumes there is an underlying act to be explained or reviewed. Where responsibility never arose, that toolkit "operates at the wrong level." You cannot audit a decision that was never constituted as one.
So the binding clause worth fighting for is not a right to an explanation, or a prize for verifying a model after the fact. It is a requirement that consequential governmental outputs take the form of discrete, authored, contestable acts. No event-marking, no authorship, no aperture—no lawful decision.
Lepore is right that nothing about the Artificial State was inevitable, and right that dismantling it "requires identifying the parts, tracing the sources." That is the historian's gift, and it is real.
But the parts that matter most are not in the past. They are in the architecture of the next system a government procures. The debate over consent is won only when it becomes a debate over institutional design—and that is the one debate the elegy, for all its beauty, was never built to have.