Every general-purpose technology arrives twice: once as a set of capabilities, and once as a set of questions about who gets to use them. The printing press arrived as movable type and left as the Reformation and two centuries of religious war. The factory arrived as machinery and left as the labour movement, the regulatory state, and a redrawn map of who owned what. Artificial intelligence is now making its first arrival as a capability, cheap enough and fast enough that a single person can produce what once required a newsroom, a surveillance agency, or an army of clerks. Its second arrival, the one that decides what it becomes, has not yet happened.
AI is not a neutral tool whose consequences can be understood separately from the arrangements through which it is owned and governed. Under the ownership and governance arrangements that currently exist, it is a machine for redistribution of wealth, of coercive power, and of authority over what counts as true and each of these is moving toward those who are already concentrated. Whether that outcome is fixed or still contested is the question this essay sets out to answer.
Start with the oldest and best-documented trend.
Labor’s share of national income has been falling since the 1980s; Loukas Karabarbounis and Brent Neiman in their 2014 QJE publication tracked the decline across most advanced economies , and Thomas Piketty’s 2014 publication offers one reason why: when the return on capital persistently exceeds the growth rate of the economy, wealth can compound faster than incomes, reinforcing the advantages of those who already own it.
AI does not initiate this trend, but it accelerates it in a way earlier automation could not, because the capital in question is no longer a machine that replaces a worker’s hands. It is software that replaces judgment, the one input automation was never supposed to reach. In his 2003 and 2015 publications, David Autor’s task-based model of automation sharpens the question from “whether” AI takes jobs to “whose”. His framework distinguishes routine from non-routine work, and cognitive from manual. Earlier automation hollowed out routine work at every skill level — the assembly-line worker and the bookkeeper alike — while sparing non-routine work at both ends of the ladder, since neither surgery nor caregiving could be codified into rules. Generative AI is the first widely deployed general-purpose technology to breach that boundary at scale: drafting, diagnosing, coding, and translating are all non-routine cognitive tasks, and it now performs versions of each. The evidence so far is mixed, some studies find AI raising the productivity of junior workers more than it threatens senior ones, but the tier of work that survived four decades of automation intact, white-collar professional labour, is for the first time genuinely exposed.
A second axis of the same redistribution runs not from labour to capital within an economy but from periphery to centre between them. The raw material of the AI economy is data generated globally, disproportionately by people with no claim on the value it produces while the compute, the models, and the revenue sit concentrated in a handful of firms in a handful of countries. It is an old extractive relationship in new technical clothing. But the same body of economics supplies a genuine counter-case, and it deserves more than a passing nod. M-Pesa allowed millions of Kenyans to access financial services without the decades of physical banking infrastructure that wealthier economies needed first; a real instance of leapfrogging, not a hypothetical one. It worked because several conditions held at once: a mobile network that already reached rural areas; a regulator, the Central Bank of Kenya, that chose to permit innovation rather than protect incumbent banks from it; local entrepreneurship layered on top of the platform; and, critically, local ownership Safaricom, Kenyan-domiciled, built and ran it, so a substantial share of the value it generated accrued within the economy that produced it.
India’s experience during the 2020 lockdowns offers a second case, produced by necessity rather than deliberate design. Cash-dependent populations were pushed onto digital rails already built; Aadhaar’s biometric identity system, the UPI payments network, Jan Dhan bank accounts, and adoption that might otherwise have taken a decade normalized within months. But India complicates the ownership condition that made Kenya’s case work: UPI is public infrastructure, not a private platform, which sharpens rather than resolves the question of who should own the rails when the technology in question is AI, where the analogous infrastructure — models, compute, the chips underneath both — is privately held and concentrated almost entirely outside the countries now adopting it. Together, the two cases suggest what leapfrogging actually requires: infrastructure already in place, a regulator willing to permit rather than block, and either local ownership or public control substantial enough to keep the value generated from leaking out.
If wealth describes what AI redistributes toward, power describes what it redistributes through, and the mechanism Shoshana Zuboff identified predates generative AI by a decade. Surveillance capitalism begins innocently enough: behavioural data collected to predict and shape consumer choices, refined for advertising. But the infrastructure built for that purpose — comprehensive, continuous data capture on ordinary life — does not stay confined to commerce. It is directly repurposable, at essentially no additional cost, for political control. The same architecture that predicts what a person will buy predicts what they believe, who they associate with, and when they might dissent. Yet the relationship between surveillance capacity and actual coercive power is not mechanical.
Sheena Greitens’s empirical work on surveillance states finds real variation in how effectively regimes convert monitoring into control; some governments over-invest in visible surveillance technology relative to its practical payoff, because the technology also signals resolve and capacity to internal elites and rival factions, not just the public. What is less contested is the institutional problem underneath both debates: the bodies built to govern AI — the EU’s AI Act, executive orders and frameworks in the United States, China’s own algorithm regulations — are each attempts to constrain a technology whose builders are also the primary source of the technical expertise, funding, and lobbying pressure shaping how it gets governed. Few regulators anywhere can independently evaluate what a frontier model can do; fewer still can do so faster than the firms building it can lobby against the finding.
Hannah Arendt’s account of totalitarianism identifies a target more fundamental than belief in any particular lie. The ideal subject of total domination, she argued, is not someone convinced that false things are true, but someone for whom the distinction between fact and fiction has ceased to matter at all. What totalitarian power requires is not agreement — it is the erosion of the shared ground on which agreement or disagreement could even be tested. That is the precise mechanism behind the sense that AI destabilizes shared reality: the danger is not that people will believe false things it generates, but that the sheer volume and plausibility of synthetic content makes verification exhausting enough that people stop attempting it.
It is worth separating, though, two things AI is often blamed for at once: generating disinformation and spreading it. These have different causal histories. Zeynep Tufekci and Danah Boyd’s work on platform dynamics locates the spread of disinformation in recommendation architecture and algorithmic amplification, systems that predate generative AI by years and would continue rewarding engagement-maximizing content even if no synthetic media existed at all. What AI changes is the cost of producing a convincing fake image, video, or article drops from what once required a studio or a state propaganda ministry to what one person can produce in seconds. Nor does this amount to something unprecedented in human history, a claim that does not survive contact with economic history. The printing press did not just spread literacy and the Reformation; it also spread a century of religious propaganda and witch-hunt pamphlets faster than any prior technology, and it took decades for courts, editorial norms, and eventually libel law to catch up. Radio did the same for both Roosevelt’s fireside chats and fascist mass rallies within the same decade.
The pattern is not that propaganda, forgery, and rumor are new; they predate AI by centuries. The pattern is a lag between the arrival of a capability and the arrival of institutions built to govern it, and what determines the outcome is less the technology itself than who holds power during that gap. AI’s genuine novelty is the marginal cost of production, collapsed close to zero, at a scale no prior lag period had to contend with — and, in a register philosophers of technology like Albert Borgmann and Langdon Winner have explored since long before generative AI existed, its effect may run deeper than false content circulating faster. A recommendation system that curates what a person sees, or a synthetic-media tool that makes any recording potentially fabricated, does not just add more lies into circulation. It changes the default posture toward everything, turning verification from an occasional task into a constant, exhausting condition of consuming any information at all.
Three threads of evidence, then, point toward the same conclusion: wealth, coercive power, and epistemic authority are all currently moving toward those who already hold the most of each. But accumulating evidence for a claim is not the same as establishing that the claim is necessary, and the harder question is whether this outcome is an inherent property of the technology or a contingent one, produced by who happens to own it right now. Winner’s 1980 essay “Do Artifacts Have Politics?” was built for exactly this question, and it refuses the easy answer in either direction. His famous, though contested, example is Robert Moses’s low-hanging overpasses on Long Island parkways, built, by some accounts, specifically low enough to keep buses — and the poorer, disproportionately Black New Yorkers who rode them — away from Jones Beach: a technology whose politics are baked into its physical form, unfixable by any change of ownership. But Winner also offers a counter-example, the mechanical tomato harvester, which favored large agribusiness over small farms not because of anything inherent in the machine but because of who had the capital to buy one when it arrived. Change who owns the harvester, and the politics change with it. Nothing in the evidence gathered here supports treating AI as the overpass, fixed by design regardless of who ends up holding it. Its redistribution of wealth runs through decisions about who owns training data, compute, and the resulting models — decisions made by firms and governments, not dictated by the technology’s architecture. Its redistribution of power runs through how regimes actually deploy surveillance capacity, and through a weapons debate that remains genuinely open. Its redistribution of epistemic authority runs through a lag between capability and governing institutions that has closed before, however imperfectly and however long it took. AI looks, so far, considerably more like the tomato harvester than the overpass.
That distinction matters because it is precisely the condition Karl Polanyi described in his account of nineteenth-century industrialization. His double movement holds that unregulated market expansion is never stable on its own; it provokes, sooner or later, a social and political countermovement that re-embeds the market in rules the market cannot generate for itself; labor law, financial regulation, the welfare state. The double movement is not automatic and it is not fast; it took the worst depression in a century and two world wars for the version Polanyi described to fully arrive. But it did arrive, organized by people who had no reason to assume in advance that it would work.
What the leapfrogging cases and the capture problem show, taken together, is the shape of the obstacle now. A countermovement requires infrastructure already in place, regulators willing to permit rather than protect incumbents, and ownership arranged so value does not leak out of the community that generated it, conditions Kenya and India met only partially, and only through very different routes. It also requires institutions capable of checking power in something close to real time, and the institutions built to govern AI are shaped disproportionately by the actors currently accumulating that power. Above all, it requires a shared enough sense of reality to organize around the exact terrain being eroded fastest and first.
None of this is a prediction. Every previous double movement was organized by people with reasonably reliable information about what was happening to them and reasonably functional institutions through which to respond; AI is the first technology whose redistribution of wealth and power depends on first winning an argument about whether the redistribution is even happening, conducted through channels the same technology is reshaping as the argument unfolds. That is a harder starting condition than the one Polanyi described, not an impossible one. The printing press’s propaganda century still ended in something recognizable as a free press, and the factory’s upheavals still produced something recognizable as a welfare state. What both took was not the passage of time on its own but people who treated the outcome as unsettled and acted accordingly, before the institutions that would later prove it was unsettled had caught up enough to help them. That is the position AI now puts everyone in: not at the end of the argument about what this technology becomes, but at the only point in its history when the argument can still be won by someone other than whoever currently holds it.