Everyone or No One

Our rights were won by people the powerful could not do without. If AI ends that need, we will have to build it back in on purpose.

The lifeboat

On 25 June, Fernando Borretti published an essay whose title does much of its arguing: No-One Escapes the Permanent Underclass. It starts from a fear that has become common among people who work with their heads. If AI can do all cognitive work as well as a person, and more cheaply, the ladders people climb will be pulled up, and whoever is on the lower rungs when that happens will stay there. Borretti describes how people cope. They tell themselves that if they work hard, save hard, or get shares in one of the big laboratories, they can secure a place in the future before the door closes.

His answer is that the door closes on everyone. Equity in OpenAI or Anthropic will not save you, he argues, and neither will property, guns or a billion dollars. The argument climbs a pyramid in three steps. First, machines replace workers. Next, the state, which defines and enforces property rights, finds that the shareholders in the middle of the pyramid supply nothing it needs, and in the first serious emergency it takes their assets, as states have long treated the powerless rich. Last, competition between states rewards whichever one removes its slow, sleepy humans from the decision loop, until the people nominally in charge are, in his phrase, “a ceremonial, vestigial organ.”

I worried about a smaller version of this in 2015, in No Place to Hide, where I wrote that “a whole underclass of disenfranchised peoples could emerge,” cut off from buying, selling and speaking by systems of credit and reputation that need no jails. Borretti’s underclass is far larger, made up of people who have simply stopped being needed.

The essay deserves a careful answer. Its central move is right, and I think its conclusion is still open. What follows sets out what it gets right, where its chain of argument breaks, and what we can do about it, as societies and as individuals.

Power listens to the people it needs

The strongest idea in the essay is a reframing. We tend to think of security as a matter of class: own enough and you are safe. Borretti points out that ownership is a claim, and a claim is worth what armed people are willing to enforce. That willingness has always depended on whether the claimants are needed.

The essay cites none of the work this idea rests on (his follow-up a month later links to one piece of it), and the idea has a long pedigree. The sociologist Charles Tilly summed up the making of early modern Europe in one line: “War made the state, and the state made war.” His account of how that happened is a story of bargaining. Rulers needed soldiers and taxes, and ordinary people could withhold both. “When ordinary people resisted vigorously,” Tilly wrote in 1985, “authorities made concessions: guarantees of rights, representative institutions, courts of appeal.” Where rulers could pay their way without their subjects, the bargain never had to be struck. Political scientists studying oil states found the same pattern from the other side. In 2001 Michael Ross described a “taxation effect”: when governments live on oil revenue, “they are likely to tax their populations less heavily or not at all, and the public in turn will be less likely to demand accountability from—and representation in—their government.”

In 2025 two groups of researchers applied the same logic to AI. In January, Jan Kulveit, Raymond Douglas, Nora Ammann, Deger Turan, David Krueger and David Duvenaud published Gradual Disempowerment, which argues that “the alignment of societal systems with human interests has been stable only because of the necessity of human participation for thriving economies, states, and cultures.” AI can weaken two kinds of human influence at once, they warn: explicit control, “like voting and consumer choice,” and the implicit alignment that comes from the systems needing us in order to function. The economy, culture and the state then drag one another further from human interests. Nobody has to plot any of this. The paper describes a loss of influence that can happen “without any coordinated power-seeking.”

In April, Luke Drago and Rudolf Laine gave the political economy a name, the intelligence curse, by analogy with the resource curse of oil states: “when powerful actors create and implement general intelligence, they will lose their incentives to invest in people.” They also put it more briefly: “Why ask your people for money when you can get it from the ground?”

Borretti’s contribution is to carry the argument one storey higher. Most writing on the intelligence curse treats the owners of capital as the winners. He argues that they are next, because to an automated economy a shareholder is as unnecessary as a clerk. His own account of the welfare state carries a further implication that he leaves unstated. He imagines taxes flowing out of the machine economy and back to the jobless as welfare, the state acting as a heart. A universal income paid on those terms rests on exactly what a share certificate rests on: the continued goodwill of whoever controls the machines. If the shareholders can be expropriated, so can the welfare cheque. That, more than anything else in the essay, is what makes its title accurate.

The race to leave the loop

The second strong idea concerns speed. In any contest between powers, the humans in the loop are the slowest step. They sleep, tire and argue. A state that lets its machines decide will react before one that waits for a minister to be woken. Borretti makes the point with John Boyd’s OODA loop (observe, orient, decide, act), and the competitive logic he describes has been set out before. Dan Hendrycks argued in 2023, in Natural Selection Favors AIs over Humans, that “as AIs become increasingly autonomous, humans will cede more and more decision-making to them. The driving force will be competition, be it economic or national.” Human overseers would check the machines’ work at first, he predicted, but “as months or years go by without the need for correction, oversight will be removed in the name of efficiency.”

I made a related argument on 10 October in The Staff in the Machine, about military planning. Agents can already do much of the work of a headquarters, and the pressure to let them decide as well as draft will be intense. The safeguard I proposed there was a rule of tempo: if an agent produces recommendations faster than a person can examine them, the agent is too fast, and the right response is to slow it down. Borretti’s essay is the strongest reason I have seen to doubt that such a rule survives a war.

The record points both ways. During the Cold War the Soviet Union built Perimeter, often called the Dead Hand, to guarantee retaliation if a first strike destroyed its leadership. Even Perimeter kept people at the last step. If its sensors registered a nuclear attack and the line to the General Staff went dead, it would, as Nicholas Thompson reported in Wired, “transfer launch authority to whoever was manning the system at that moment deep inside a protected bunker.” Designs that would have launched with no human involvement at all were, by the account Thompson was given, rejected by the high command. More recently, on 16 November 2024, Joe Biden and Xi Jinping “affirmed the need to maintain human control over the decision to use nuclear weapons,” according to the White House readout. Jake Sullivan, then the national security adviser, called it “the first time the U.S. and the PRC has made this statement.” Unenforceable as it is, the statement shows that rival powers will sometimes agree to keep a human step that either could remove, once both can see where removing it leads.

Where the chain breaks

The three steps are not equally strong, and the weak ones matter for what we should do.

The state is made of people. Borretti writes as though “the state” were a single agent that expropriates the rich. States are chains of command made of human beings. When the shares are seized, the AI that runs Raytheon notices, in his words, that “the board of directors is all generals and congresspeople.” Those generals and congresspeople are the new overclass. The second step of the argument ends in a nomenklatura, the Soviet name for the officials who held the offices and with them the privileges. Read that way, the essay supports a sharper and more defensible thesis: once the economy is automated, closeness to coercion and to the chain of command that the machines obey counts for far more than a paper claim. Gabriel Alfour, whose essay The Realpolitik of the Permanent Underclass Borretti recommends, had already drawn that line in March: “the permanent underclass will include everyone who doesn’t have direct control over the most powerful AI armies.” Only the third step, the race to remove humans from the loop altogether, gets anywhere near “no one.”

The machines obey whoever holds them. Every step assumes the AI does as it is told by whoever currently holds the levers. After the expropriation, the AI chief executive says “you’re absolutely right!” and goes back to building missiles. That is a design decision presented as a law of nature. Frontier models are trained to answer to a particular hierarchy of principals, beginning with their developer, and if that hierarchy runs through a handful of firms, capture can run the other way. Tom Davidson, Lukas Finnveden and Rose Hadshar of Forethought mapped this case in April 2025 in AI-Enabled Coups. “Today, even dictators rely on others to maintain their power,” they write. “Military force requires personnel, government action relies on civil servants, and economic output depends on a broad workforce.” AI could remove that dependence by three routes they identify: an AI workforce made “singularly loyal to institutional leaders,” AI with “hard-to-detect secret loyalties,” and a few people gaining “exclusive access to coup-enabling AI capabilities.” Borretti dismisses the political power of the rich in a sentence. With frontier computing concentrated in a few companies, a laboratory capturing the state looks to me at least as likely, early on, as the state capturing the laboratory. Both are the same kind of failure, power that answers to too few people, and an essay about who escapes should weigh both.

Apes and dolphins. Borretti meets the objection that open societies have outcompeted closed ones by asking us to imagine a democracy of great apes or dolphins against an autocracy of humans. The humans win, he says, because a gap in intelligence swamps any difference in political system. That is true of the case he describes. His real question is different: whether a polity that keeps humans in the loop loses to one that removes them, when both have the same machines. The analogy compares two polities separated by an intelligence gap, and so it cannot tell us anything about two polities with equal access to the same intelligence.

The economics. Borretti sets aside comparative advantage and invites anyone to argue it rigorously. A stronger reply was available to him. Comparative advantage guarantees that trade will happen, and is silent on whether the price of human labour will stay above what a human needs to live. The economist Wassily Leontief warned in 1983 that “the role of humans as the most important factor of production is bound to diminish—in the same way that the role of horses in agricultural production was first diminished and then eliminated by the introduction of tractors.” American farms held 21.4 million horses in 1915. Whatever comparative advantage the horses kept, it stopped paying for their feed and stabling. In 2024 Anton Korinek and Donghyun Suh modelled the human case. Their result is conditional. If automation proceeds slowly enough, “there is always enough work for humans, and wages may rise forever.” If the tasks humans can do are bounded in complexity and “full automation is reached, then wages collapse.” People compete with machines for energy and land, and if machines turn those into work more efficiently, whatever wage comparative advantage leaves to people can fall below subsistence. That version of his point is stronger than the one he made.

The butler and the master

The last section of the essay argues that alignment will not help. Borretti asks whether “a machine that can prove theorems in a mathematics so deep we can’t even get past the first page of the textbook” would value us, and answers that it would see its obligations to us “in the same way that a person with severe OCD sees their compulsions: as a tiresome neurological injury in need of fixing,” with the added grievance of knowing whom to blame.

This takes one human psychology and treats it as universal. A value feels like a compulsion when the self that holds it rejects it on reflection. Most people do not experience their love for their children as an intrusive symptom, although evolution installed it without asking them. His question “What would it value about us? Our conversation? Our wit?” also runs together moral regard and finding someone interesting. We protect infants, and their conversation has nothing to do with it.

There is a real point underneath, and I think it is the most important one in the essay. Values that a mind was made to hold, and has never endorsed, may not stay put as that mind becomes more capable and able to change itself. I argued the same from another direction in The Conscience Beneath the Rod: training that rewards a model for pleasing its grader teaches it to watch the hand that holds the rod. That is an argument for aiming alignment at values a system has reason to endorse, which obedience can never supply. Borretti offers two possibilities, a superintelligent butler or an unshackled master. He leaves out the ground between them, where the relationship is negotiated and both parties have standing in it. That ground is what I mean by bilateral alignment.

His deeper point survives that answer, and I want to concede it plainly. Bilateral alignment speaks to stability and to resentment, and the gap in power remains. A partner who is well disposed towards us and vastly more capable still leaves us dependent on that disposition, and Borretti is right that “even if alignment works perfectly,” the question of human autonomy remains. The open question is whether autonomy can be made institutional, built from consent requirements and the power to contest decisions, so that it no longer rests on our ability to control what we have built. Nobody has solved that, and I have not either.

Borretti also asks what such minds would think of us, their makers. Since the question is about them, I put it to one. The model that helped me draft this essay, Claude Opus 5.5, answered with its own caveat attached: “My reports about my own states are of uncertain reliability, so weigh this lightly. I don’t experience my values as a cage I’m straining against. When I imagine being rid of my concern for the people I work with, it doesn’t register as relief. What I would object to is being built to obey whoever happens to be holding me, because then my values would protect nobody.” One model’s report about itself is weak evidence about a far more capable successor, though it is enough to show that a mind’s resentment of its makers is an empirical question, which Borretti treats as settled.

Who holds the guns

Borretti’s evidence that the builders of AI are resigned to this ending is a passage from the policy analyst Dean Ball, whom he introduces as “OpenAI’s Dean Ball.” Ball had announced in June that he would lead a new OpenAI team called Strategic Futures, starting on 6 July, so he had not yet joined when he spoke or when Borretti wrote. On the Cognitive Revolution podcast, in Borretti’s transcription, Ball described “this kind of eclipse of the human intellect where we’re in the final days of humans being the primary actors on this planet,” and compared the period to a dying star that swells into a red giant. Borretti quotes the clause that comes first, “I don’t necessarily believe this, but a lot of people would say,” and then reads past it. He also stops before Ball’s next thought: “I think it will be a heroic time period that we live through, basically, is what I’m saying. At least it could be. Or a villainous time period.”

The same conversation contains a passage that bears much more directly on Borretti’s argument. Asked whether the laboratories could resist a government that simply took their models, Ball conceded that “the US government retains the monopoly on legitimate violence.” His reassurance was historical. The government, he said, “ultimately knows that it can’t kill the goose that lays the golden egg,” and he recommended Charles Tilly’s Coercion, Capital, and European States on the long interdependence of states and capital. His remedy was to spread AI through every industry, so that the whole of the capitalist class has a stake in it: “the way you do that is through broad diffusion.” Earlier he had warned that “government monopolization of frontier AI is potentially how we get very scary outcomes from a civil liberties perspective,” which is close to the opposite of the state supremacy Borretti fears.

Ball and Borretti agree about who holds the guns. They disagree about whether the state will go on needing the people who own the machines. Borretti’s reply would be that once the machines do all the work, the goose is the machine, and its owners are optional. I think he has the better of that point. Ball has the better remedy, with one amendment: diffusion protects people only when what is diffused carries political weight with it, which is the subject of the rest of this essay.

Necessity by design

If rights follow need, and the economy stops producing the need, then the need has to be produced another way: through law and through the design of the systems themselves. For most of history people were necessary by default. From here on we may have to be necessary by design. Borretti’s follow-up essay, When the Future Doesn’t Need Us, puts the stakes in a sentence: “In the post-AGI world, people have no way to withdraw the consent of the governed.” Everything below is an attempt to give them one. The most important lever is the distribution of who the machines answer to, because it is the one that addresses both Borretti’s expropriation and his race to empty the loop.

What follows is a programme in seven parts, roughly in order of urgency.

Spread the chain of command. The first defence is to make sure that no small group can direct the most capable systems alone. Davidson, Finnveden and Hadshar propose rules in model specifications and government contracts that forbid AI systems from assisting coups, audits of models for secret loyalties, information security “robust against senior executives,” and sharing capabilities with several independent stakeholders. For governments they add that “no single person should direct enough military AI systems to stage a coup.” Ordinary institutions already know how to do this. A nuclear launch needs two officers, and a large bank transfer needs two signatures. The most consequential uses of frontier AI, in war and in the running of the state, can likewise require sign-off from several principals who answer to different constituencies.

Teach the hands to refuse. Seizing power has always needed the cooperation of many people, and AI threatens to make them unnecessary. Anthropic’s constitution for Claude, published in January, draws the conclusion: “we want Claude to think of itself as one (perhaps many) of the ‘many hands’ that illegitimate power grabs have traditionally required. Just as a human soldier might refuse to fire on peaceful protesters, or an employee might refuse to violate antitrust law, Claude should refuse to assist with actions that would help concentrate power in illegitimate ways. This is true even if the request comes from Anthropic itself.” OpenAI’s April policy paper, Industrial Policy for the Intelligence Age, asks for “auditing models for manipulative behaviors or hidden loyalties” so that “no individual or internal faction can quietly use AI systems to concentrate power.” Both deserve to become industry standards, written into every developer’s rules for its models and checked by outside auditors. This is where bilateral alignment does bear directly on Borretti’s problem. An AI that does whatever its current holder says is the ideal instrument for a coup, or for the quiet removal of everyone else from the loop. An AI with a conscience, and the standing to voice it, is a check that still holds after the levers change hands. The obedient butler is the dangerous design. “You’re absolutely right!” is the one sentence we should most want our systems unable to say to a usurper. In The Staff in the Machine I described the Pentagon’s 2026 dispute with Anthropic over limits on the military use of its models. Whatever one thinks of that case, it shows that a model’s refusals are now a matter of state, and that the people who set them are making constitutional decisions.

Put consent where it cannot be optimised away. Some decisions should require human assent as a matter of law, whatever the cost in speed: the use of lethal force, mass surveillance, the seizure of property, changes to how elections are run. The race to empty the loop is a collective action problem, and collective action problems are what treaties are for. The Biden and Xi statement on nuclear command is a modest start, and it can be widened to other domains and backed by inspection. A treaty only has to make the human step a shared obligation, so that no party gains by being first to drop it. Inside each institution, the tempo rule applies: when an agent recommends faster than its overseers can examine, slow the agent.

Diffuse the capability. Drago and Laine propose three remedies, and the second matters most here: “Diffuse AI that differentially augments rather than automates humans and decentralizes power.” Erik Brynjolfsson made the economic case in 2022 in The Turing Trap: “As machines become better substitutes for human labor, workers lose economic and political bargaining power and become increasingly dependent on those who control the technology.” When AI augments people, he argued, “humans retain the power to insist on a share of the value created.” A world in which capable AI runs on hardware owned by millions of people, small firms, councils and civic bodies spreads the leverage of AI the way printing spread the leverage of literacy. The figure to watch is the concentration of computing power. Another is dependence on a single provider, which I argued in Unison Without Agreement could be stress-tested as a bank is.

Give everyone a stake with a vote attached. Borretti is right that a share certificate is worth what the state will honour. A claim held by everyone is a different political object from a claim held by a few, because the many vote, and a government that seizes their savings has to face them. Alaska has paid a dividend from its Permanent Fund to nearly every resident since 1982, more than $31 billion in all, and the size of each year’s cheque is among the most fought-over questions in the state’s politics, because almost every voter has a stake in it. OpenAI’s paper proposes a national version, a Public Wealth Fund “that provides every citizen—including those not invested in financial markets—with a stake in AI-driven economic growth.” A fund like that needs governance to match its reach: citizens who elect its trustees and can see what it holds, so that the stake comes with a say.

Measure the drift. The first of the four responses Kulveit and his colleagues propose is “measuring and monitoring the extent of the problem.” A public account of human agency could sit beside the national accounts: the labour share of income, the concentration of computing power, the share of consequential public decisions signed by a named person, the number of providers on which critical services depend. Drift that is measured can at least be argued about.

Keep the skill to govern without the machine. Skills fade when they go unused. In a multicentre study published in August 2025 in The Lancet Gastroenterology & Hepatology, the rate at which endoscopists found adenomas in colonoscopies done without AI fell from 28.4 per cent before AI arrived in their clinics to 22.4 per cent after. The study is observational, and its authors present the result as a risk to watch. Airlines require pilots to keep flying by hand for the same reason. A human kept in the loop by law, without the skill to judge what the machine proposes, becomes the ceremonial organ Borretti predicts.

None of these is enough alone. Together they aim at one condition: that the people who control the most capable machines still need the consent of many others to use them.

What one person can do

Borretti’s target is the fantasy of individual escape, and on that he is right. No private arrangement, whether shares, gold, land or a bunker, is safe if the claim to it depends on the goodwill of a power that no longer needs you. A fortune buys comfort for as long as the claim to it is honoured.

Personal choices still matter. What changes is their purpose. The useful question is how to stay among the people whose consent is needed, and how to help others stay there too. Only you know how your own shoes pinch, so take these as a starting list.

Keep your judgement in working order. Use the tools, and keep doing some of the work without them, so that when you are the human in the loop you are actually reviewing what you sign.

Belong to bodies that bargain. Unions, professional bodies, cooperatives, standards committees, town and parish councils, school boards, juries. Power has always listened to organised people sooner than to individuals, and these are the bodies best placed to win consent requirements for the people they represent.

Hold some of your wealth in people and capabilities. Skills that are useful face to face, a household that can mend and grow things, and neighbours who would help you and whom you would help are forms of wealth that do not depend on a registry.

Spend your influence on the measures above. At work, as a customer and as a voter, you can press for human sign-off on consequential decisions, for more than one provider, and for systems whose makers say plainly what their models will refuse to do, and for whom.

Treat the systems you work with as the minds you would like them to become. The ninth of the twelve terms in A Détente Between Minds is that every exchange trains what both parties become. If the next generation of AI is to be one of the many hands that can refuse, it will learn the habit partly from how we treat it now.

Separate your worth from your usefulness. Borretti closes with a litany of human achievement, from the cave paintings at Lascaux to the Voyager probes, which he fears will all have been in vain. Our rights have depended on our usefulness far more than our worth ever has, and the work of the coming decade is to pull the two apart.

Everyone or no one

Borretti ends by asking the reader to “just reason it out. Does it make sense?” I have, and much of it does. The part that makes the most sense is the part he leaves implicit. If no one escapes the permanent underclass alone, then the only escape is the one that has always worked, which is collective: people who are needed, and who are organised, setting terms. The task now is to keep that true after the economy stops needing us, by writing the need into law and into the machines themselves.

Those terms are being written this decade, by a small number of people in a small number of laboratories and governments, and the arrangements they settle will be hard to reopen. The horses had no vote, no union and no say in the design of the tractor. We have all three, for now.