A thousand years of history of major innovations lead Daron Acemoglu and Simon Johnson to a conclusion that should worry anyone betting on AI to improve the lot of the majority: technology spreads its gains only when countervailing powers force it to do so. Without this, it enriches those who control it, sometimes for generations.

Power and Progress, published in 2023, is a book of historical economics and political philosophy. It is also a book of struggle. At a time when Silicon Valley giants announce a future of shared abundance, the two authors refuse fatalism in both directions: neither the naive optimism of those who believe technology naturally benefits everyone, nor the catastrophism of those who see AI as the end of human work. Their wager is more demanding: to show that the sharing of gains is a political choice, and that it remains accessible — provided action is taken before power relationships solidify.

The Essentials

  • The share of income going to capital has increased by approximately 10 percentage points in advanced economies since 1980, a trend Acemoglu and Johnson link directly to digital automation.
  • By tracing a thousand years of history of major technological revolutions, the authors show that gains have systematically been captured by a minority before institutions imposed redistribution — sometimes after several generations.
  • The current trajectory of AI, oriented toward automation rather than worker augmentation, is not a technical necessity: it is a choice of investment and governance.
  • The window for correcting this trajectory is open, but it closes as power relationships solidify around the actors who control the models.

The Authors and Their Moment

Daron Acemoglu is an economist at MIT, Nobel Prize in Economics 2024 with Johnson and James Robinson for their work on institutions as drivers of development. Simon Johnson, also at MIT and former chief economist of the IMF, works on financial regulation and industrial policy. The two had already co-signed Why Nations Fail with Robinson in 2012, a thesis on the role of inclusive institutions in long-term prosperity. Power and Progress extends this work by applying it to the technology question.

The book appears in a specific context: the explosion of ChatGPT in 2022, announcements of massive layoffs in large technology companies, and a public debate oscillating between fascination and panic. Acemoglu and Johnson choose to step out of this immediate debate to pose a longer question: have we seen this before? And if so, how did it end?

The Thesis: Technology Obeys Those Who Pay for It

The central idea of the book can be summarized in a sentence that the authors formulate in the opening pages: “Technology is not a natural force imposing itself on society. It is a social construction that reflects the choices, interests, and power of those who develop and deploy it.”

This is not an anti-technology thesis. It is a thesis about the direction of technical progress. Acemoglu and Johnson distinguish two possible orientations for an innovation: automation, which replaces workers, and augmentation, which increases their productivity without eliminating them. Their finding is that the latter is not less profitable than the former over the long term, but it is systematically less chosen when decisions are concentrated in the hands of a small elite that has no interest in sharing gains.

The data they mobilize to anchor this thesis in the present is striking: since 1980, the share of national income going to wages in advanced economies has fallen by approximately 10 percentage points in favor of capital. This phenomenon, documented by economists such as Lawrence Katz and Alan Krueger for the United States, coincides precisely with successive waves of digital automation. For the authors, this is not a coincidence: it is demonstration that the productivity gains of the information era have largely been captured by holders of capital and a very narrow segment of highly skilled workers.

A Thousand Years of Evidence

The argumentative force of the book rests on its historical ambition. To avoid any technological determinism, the authors go back to medieval agriculture, to the printing press, to the steam engine, to electricity. In each case, they document the same pattern: a major innovation initially creates a rent for those who control it, compresses the incomes of existing workers, generates decades of concentration before external forces impose a rebalancing.

The most developed example is that of the British Industrial Revolution. Contrary to the dominant narrative that treats it as a moment of general progress, Acemoglu and Johnson remind us that the living conditions of English workers stagnated, or even declined, during the first 50 to 70 years of mechanization. Only with the extension of voting rights, union organization, and early social legislation — between 1830 and 1880 — did productivity gains begin to diffuse into wages. The correction took two generations.

From this analysis they derive a principle they call “the machinery problem” — a term borrowed from Ricardo, who had himself hesitated on the question in the third edition of his Principles. Technical progress creates wealth, but its distribution depends on institutional power relationships. Without countervailing powers, this wealth accumulates where decision-making power lies.

AI as Choice, Not Destiny

The contemporary part of the book is the most directly political. The authors document there the vision guiding today’s major AI laboratories: a conviction, shared by leaders at Google DeepMind, OpenAI, and Meta, that maximum automation is not only inevitable but desirable, on the grounds that it would free humans from repetitive tasks. Acemoglu and Johnson do not reject this ambition. They point out that it rests on an implicit hypothesis rarely discussed: that the gains thus generated will actually be redistributed.

The data does not support this hypothesis. The authors cite work by Acemoglu himself, published in the American Economic Review with Pascual Restrepo, showing that between 1990 and 2007, each additional robot per thousand workers in the United States reduced employment by 0.18 to 0.34 percentage points and compressed wages by 0.25 to 0.5 percentage points. These figures may seem modest; they are less so when projected onto the volumes of automation announced for the next decade.

But the book’s most original critique concerns the direction of research itself. Acemoglu and Johnson show that most AI investments today are oriented toward applications that substitute capital for human labor — speech recognition, content generation, autonomous driving — rather than toward tools that would make doctors, teachers, or engineers more effective. This choice of orientation is not neutral: it reflects the incentive structures of companies developing these technologies, largely dependent on stock market valuations that reward the reduction of wage costs rather than the increase in value created by workers.

The thesis resonates directly with the institutional question found in the analysis of Mexican nearshoring: a technological or geographic advantage is worthless if institutions to share its gains are absent.

What Garnier’s Data Illuminates Differently

The reading of Acemoglu and Johnson gains from being confronted with a more cyclical perspective. The work of Olivier Garnier, director general of studies at the Banque de France, on public statistics and productivity measurement raises a question that nuances the picture painted by the authors: are our measurement tools adapted to capture the gains of AI?

Garnier and other institutional economists raise the fact that the productivity stagnation observed since 2008 in advanced economies could partially reflect a measurement problem rather than an absence of real gains. Digital services create value that national accounts capture poorly — increased quality of medical research, expanded access to information, reduction of transaction costs. The Solow paradox, which we examined in the analysis of American productivity data, remains intact: we see AI everywhere except in productivity statistics.

This objection is serious, but Acemoglu and Johnson respond to it indirectly. Even if AI gains are under-measured, the question of their distribution remains. A free digital service whose value does not figure in GDP does not redistribute income to workers. Diffuse gains in well-being (better information, faster services) do not compensate for the compression of real wages. The measurement argument can nuance the scale of the phenomenon; it does not change its direction.

The Book’s Blind Spots

Power and Progress is a book of political economy, and its blind spots are those of this discipline.

The first is geographic. The authors’ analysis is largely centered on advanced economies, primarily American and European. The question of how developing countries position themselves relative to this dynamic — and whether different trajectories are possible in more fragile institutional contexts — is barely addressed. China, which is developing a deliberately state-directed industrial policy for AI oriented toward state objectives, is not analyzed in detail. Yet it constitutes a potential counter-example to the implicit liberal thesis of the authors, which assumes that democratic institutions are the natural vehicle for countervailing powers.

The second blind spot is one of speed. The historical examples mobilized — industrial revolution, electrification — concern transitions that took several decades. If generative AI evolves at a significantly faster pace, will the mechanisms of institutional correction have time to take place? The authors pose the question without fully answering it. Their conditional optimism (“it’s a choice, so we can change it”) rests on the implicit hypothesis that the window of action remains open. But they themselves recognize, at the end of the volume, that this window closes as dominant actors consolidate their position.

The third blind spot is more conceptual. The book devotes little space to cases where automation has actually freed workers for more valuable activities — automated checkouts and cashiers reconverted into customer advisors, accounting software and the rise of financial analysts. These examples do not refute the central thesis, but they suggest that the boundary between automation and augmentation is more porous than the authors’ analytical framework suggests.

The Closing Window

The forward-looking dimension of the book is also its most uncomfortable. The authors project that if the current automation trajectory is not corrected by appropriate institutions in the next twenty years, AI productivity gains will reproduce the income concentration observed during the first industrial revolution. Their implicit model is one of a cumulative process: the more gains accumulate with holders of technological capital, the more their ability to influence institutional rules increases, the more correction becomes difficult.

This scenario is not prophecy. It is a conditional trajectory. The authors themselves identify available levers: taxation of productivity gains induced by automation to finance training and professional transitions, reorientation of public research policies toward worker augmentation applications, strengthening of labor law in sectors exposed to AI. These proposals are not revolutionary — some are already under discussion in the European Union, notably in the framework of the AI Act and negotiations on the taxation of digital platforms.

What the book adds to this political debate is the long perspective. Nineteenth-century unions, child labor laws, the American New Deal, French Social Security: none of these institutions were inevitable. They resulted from conflicts, mobilizations, deliberate political choices. The authors do not say that history repeats itself. They say that the mechanisms it reveals remain operative.

The twenty-year horizon they set is not arbitrary. It is the estimated time for current foundation models — trained on data corpora that will not be reconstituted — to consolidate their position of cognitive monopoly. Beyond that point, the concentration of power over AI infrastructure will be structurally comparable to that of railroads in the nineteenth century: a first-mover advantage difficult to challenge without forced public intervention. The question of what AI does to knowledge production itself is not separable from this concentration: whoever controls the models also progressively controls the direction of research.

Why Read It

Power and Progress addresses anyone wanting to move beyond the binary debate on AI — neither technological hymn nor employment catastrophism. It also addresses policymakers seeking a historical framework for thinking about AI policies: the work is one of the few serious attempts to link long economic history to immediately available political choices.

Finally, it addresses economists and practitioners of industrial policy who want to understand why productivity growth in advanced economies remains disappointing despite massive technology investments. The authors’ thesis — that gains exist but are captured before diffusing — is one of the most documented answers to this puzzle.

What this book changes in understanding the subject is precisely this refusal of determinism. AI is neither a kept promise nor an ineluctable threat. It is a tool whose direction depends on collective choices. And these choices, Acemoglu and Johnson remind us with historical patience, have always been contestable — and have sometimes been contested successfully.

Bibliographic Information

Title: Power and Progress: Our Thousand-Year Struggle over Technology and Prosperity Authors: Daron Acemoglu & Simon Johnson Publisher: PublicAffairs Year: 2023 Pages: 546 pages

Sources

  1. Daron Acemoglu & Simon Johnson, Power and Progress, PublicAffairs, 2023
  2. Daron Acemoglu & Pascual Restrepo, “Robots and Jobs: Evidence from US Labor Markets”, American Economic Review, 2020 (no link — academic article)
  3. Lawrence Katz & Alan Krueger, “The Rise and Nature of Alternative Work Arrangements in the United States”, ILR Review, 2019 (no link — academic article)
  4. Banque de France, work by Olivier Garnier on measuring productivity in the digital age (no link — internal reports and public addresses)