The Great Stagnation: why productivity growth slowed and whether it can recover

“You can see the computer age everywhere but in the productivity statistics.” Robert Solow wrote that in 1987, in a book review for the New York Times Book Review — a throwaway line, almost, buried on page 36 of the July 12 issue. It became the most cited sentence in productivity economics. And then it happened again.

In August 2024, the Richmond Fed published Economic Brief 24-25, authored by Erin Henry, Pierre-Daniel Sarte, and Jack Taylor. The title was “The Productivity Puzzle: AI, Technology Adoption and the Workforce.” The opening move: Solow’s sentence, quoted word for word, then applied to artificial intelligence. In January 2025, David Smith, economics editor of the Sunday Times, used it as a headline: “You See AI Everywhere, Except in the Productivity Figures.” Not a paraphrase. Not a thematic echo. The same sentence, recycled with surgical precision, nearly four decades later, about a completely different technology.

This should bother you more than it does. The recurrence is not a literary coincidence — economists reaching for a familiar aphorism. It is diagnostic. It tells you that whatever produced the original paradox was never resolved. The computer age eventually showed up in the statistics, briefly, between 1996 and 2004 — then vanished. And now a new technology is doing the same thing: transforming visible daily life while the productivity numbers refuse to move.

The question worth asking is not whether AI will eventually fix this. The question is why this keeps happening. Why does every generation of transformative technology produce the same gap between what you can see and what the data records? The answer is not about measurement. It is not about patience. It is about a specific confusion — between technological impressiveness and the structural reorganization that actually drives aggregate growth — that we keep making because we have never properly diagnosed it.

The numbers

The scale of the slowdown is easy to understate because percentages obscure compounding. US nonfarm business sector productivity — output per hour worked — averaged roughly 3.2% annual growth from 1948 to 1973. After 1973, it dropped to approximately 1.4% and never sustainably recovered. The difference between 3.2% and 1.4% compounded over fifty years is not small. At the pre-1973 rate, US output per hour would now be roughly double its actual level. That gap is the size of an entire economy.

Robert Gordon, in The Rise and Fall of American Growth, frames this as the end of a “special century”: output per hour grew at roughly 2.8% annually from 1920 to 1970, compared with approximately 1.8% from 1870 to 1920 and approximately 1.7% after 1970. TFP growth — the deeper measure that strips out capital accumulation — was triple its pre-1920 and post-1970 rates during that central period. The special century was genuinely special. What followed was a reversion, not a dip.

The most recent data does nothing to change this picture. Private nonfarm business labor productivity grew 2.8% in 2024 — a cyclical rebound — and 2.2% in 2025. Respectable. Not a structural break. OECD-wide, labour productivity growth averaged approximately 0.4% in 2024. Stagnant by any historical standard.

Four deflections arise immediately, and all of them fail. This is not a US-specific problem — OECD data confirms it is universal across developed economies. It is not a post-2008 financial crisis effect — the slowdown began in 1973, and the post-2004 deceleration predates the crisis by four years. It is not temporary — fifty years is not a business cycle. And it is not simply a composition effect from the shift toward services — the slowdown occurred across manufacturing and services alike.

Something real happened.

Labour productivity versus total factor productivity

Labour productivity — output per hour worked — can rise simply because firms invest in more or better equipment per worker. A factory that buys each worker a faster machine raises output per hour without any genuine efficiency gain. Total factor productivity strips out this capital deepening, isolating the residual: how much more output the economy produces from the same inputs. Economists treat TFP as the fundamental measure of long-run growth potential. When TFP stagnates, capital accumulation alone cannot sustain growth indefinitely. The distinction matters throughout what follows.

The measurement escape hatch

The most comfortable explanation is that the slowdown is a statistical illusion. GDP measures market transactions. Free digital goods — Google Maps, Wikipedia, streaming services priced near zero — create genuine economic value that never enters the output statistics because their market price is zero. We are richer than GDP says. The productivity crisis is a measurement crisis.

This is half right and wholly insufficient.

Erik Brynjolfsson and co-authors developed the GDP-B framework — a method for estimating consumer surplus from free digital goods using incentive-compatible willingness-to-pay surveys. Their finding, published as NBER Working Paper 25695 in 2019: the median US Facebook user would require approximately $48 per month to give up access. Incorporating Facebook’s unmeasured welfare contribution adds an estimated 0.05 to 0.11 percentage points to annual GDP-B growth. The consumer surplus from digital goods is real and genuinely uncaptured by GDP.

But Chad Syverson, in a 2017 Journal of Economic Perspectives paper, demolished mismeasurement as a sufficient explanation. His empirical case has three parts, and the first is close to decisive on its own: the productivity slowdown is a cross-country phenomenon, and its magnitude is unrelated to countries’ ICT intensity. If the slowdown were a measurement artifact of digital economies, it should be most severe in the most digital ones. It isn’t. Germany and Japan show comparable slowdowns despite lower ICT penetration than the United States.

Second: the total “missing output” implied by the post-2004 productivity slowdown — the gap between what GDP would have been at the pre-2004 trend and what it actually was — amounts to over $3 trillion annually. The entire estimated consumer surplus from all free digital goods combined covers a small fraction of that gap. Facebook’s unmeasured welfare value, generous as the GDP-B estimates are, adds 0.05 to 0.11 percentage points per year. The missing output requires roughly 2 full percentage points per year. The arithmetic is not close.

Third: for mismeasurement to account for even a modest fraction of missing output, ICT-producing and ICT-intensive industries would have needed unmeasured productivity growth rates that are multiples — not fractions, multiples — of their measured rates. The inter-industry data does not support this. The sectors most plausibly producing unmeasured digital value show no such anomalous gap between their measured and implied productivity.

Syverson’s conclusion is not that digital goods have no unmeasured value. They do. His conclusion is that the unmeasured value is an order of magnitude too small to explain what happened. Mismeasurement is real and irrelevant to the central question. These are not two sides of a debate. The cross-country evidence is an empirical result. The mismeasurement hypothesis has not answered it.

How GDP-B works

The GDP-B methodology uses incentive-compatible discrete choice experiments — respondents choose between keeping a service and receiving a cash payment, with real money at stake — to estimate consumer surplus. The estimates are meaningful but noisy: they capture what people say they would accept, under conditions designed to elicit truthful responses. The aggregate numbers are genuine welfare gains, genuinely uncaptured by GDP, and genuinely an order of magnitude too small to close the productivity gap. A $48-per-month welfare gain from Facebook, even summed across hundreds of millions of users, does not approach $3 trillion in missing annual output.

What actually made the special century special

If the measurement story doesn’t explain the slowdown, the question shifts: what actually drove the fifty years of growth that we have not been able to replicate? The common summary of Gordon’s argument — “IT is overrated” — misses the structural claim. The structural claim is the one that matters.

The inventions of 1870–1940 did not do old things faster. They eliminated old things entirely. Indoor plumbing didn’t accelerate water-hauling — it ended water-hauling. Before indoor plumbing, a typical American household spent hours per day on water-related labour: hauling, boiling, disposing. After it, that time went to zero. Not reduced. Zeroed. Refrigeration didn’t speed up food preservation — it made an entire category of domestic labor unnecessary. The icebox had already reduced spoilage, but mechanical refrigeration eliminated the daily shopping trip that perishability required. The washing machine didn’t improve the efficiency of hand-scrubbing — it turned a full day’s labour into an hour of machine operation.

Each invention annihilated an activity that had consumed enormous fractions of household time. Women entered the paid workforce in large numbers partly because the domestic labour that had kept them out was mechanically abolished. And each of these eliminations could happen exactly once.

Gordon’s comparison to IT is brutal in its specificity. Email accelerated letter-writing — but letter-writing was never an economy-wide bottleneck. Nobody was spending four hours a day on correspondence the way they spent four hours hauling water. Search engines dramatically improved information retrieval at the margins of knowledge work — but information retrieval was never the kind of economy-wide bottleneck that food preservation and transport were. The IT revolution improved things that were already tolerable. The special century eliminated things that were intolerable.

Tyler Cowen’s frame in The Great Stagnation goes deeper. The growth of 1880–1970 rested on three specific forms of low-hanging fruit: previously unused free land — the continental frontier that turned subsistence farmers into landowners at essentially no capital cost; the rapid translation of 18th- and 19th-century scientific breakthroughs into working technologies — the basic science of thermodynamics, electromagnetism, and germ theory, accumulated over centuries, converted into steam engines, dynamos, and antibiotics in a compressed burst between 1880 and 1940; and the large returns from educating previously uneducated populations — taking high school completion from a rarity to a near-universal norm, each additional year of schooling producing large marginal returns because the starting base was so low.

All three were historically contingent. One-time conditions, not renewable features of industrial economies. You cannot discover free land twice. You cannot harvest the transition from mass illiteracy to mass literacy twice. The internet, Cowen observed, has been “fantastic for the intellectually curious” but has done little to raise material living standards for the median household. The largest internet companies employ remarkably few people relative to their market capitalisation. The economic footprint is narrow even when the cultural footprint is vast.

Gordon and Cowen disagree about how closed the future is — Gordon more pessimistic about transformative innovation resuming, Cowen leaving the door open. But they agree on the diagnosis: the post-1970 slowdown is structural, not cyclical, and not an artifact.

The question this reframes is not “why did IT disappoint?” It is: what would any technology have to do to match the special century? What test must it pass?

Gordon's headwinds

Independent of the innovation question, Gordon identifies four structural drags suppressing growth: rising inequality concentrating productivity gains among a narrow elite rather than raising median incomes; educational attainment plateauing after decades of expansion; demographic aging reducing the working-population share; and rising government and household debt constraining fiscal capacity. These headwinds mean that even a genuine productivity acceleration might not translate into broadly shared prosperity. They are important but orthogonal to the technological question — whether aggregate productivity can accelerate at all, not how gains distribute once it does.

The exception that proves the rule

Between 1996 and 2004, the slowdown appeared to reverse. The Congressional Budget Office documented that nonfarm business output per hour grew at roughly 3% annually during this period — matching the pre-1970 pace. IT optimists cite this as proof that technology can drive growth. They are right. But the composition of those gains confirms the thesis rather than complicating it.

The acceleration was driven primarily by two mechanisms: semiconductor manufacturing productivity — as Moore’s Law drove relentless price deflation in computing hardware, making each unit of computing power cheaper at a rate that directly inflated measured output per hour in the semiconductor sector — and wholesale/retail logistics improvements. Walmart’s inventory management revolution, built on barcode scanning, electronic data interchange, and real-time supply chain coordination, produced enormous measured efficiency gains in the distribution sector. The broader retail industry then adopted these methods in a diffusion wave that sustained the productivity surge for nearly a decade. The consumer internet — websites, e-commerce, email — contributed. But substantially less than the hardware and logistics components. The mechanisms that drove the measured numbers were specific, identifiable, and sectoral.

The critical test came in 2004. Semiconductor price deflation slowed as Moore’s Law approached physical constraints. Walmart’s inventory methods had diffused through the retail sector — the reorganization was complete, the gains harvested. The productivity surge ended.

And then: Facebook launched in 2004. YouTube followed in 2005. The iPhone arrived in 2007. Cloud computing scaled through the late 2000s. Broadband penetration expanded. The social media ecosystem exploded into one of the most culturally transformative technological developments in living memory. Visible technological change accelerated — massively, undeniably — after the productivity surge had already peaked and ended.

Consumer internet growth continued accelerating after 2004. The mechanisms that had actually driven measured productivity — semiconductor output, logistics efficiency — plateaued. Technology that is impressive, widely adopted, and culturally dominant is not the same thing as technology that drives measured output. These two can decouple entirely. They did.

The Dynamo Problem — or, why it might still come

The decoupling after 2004 creates a precise diagnostic for AI: is it the semiconductor — structurally transformative in ways that show up in output? Or the consumer internet — impressive, dominant, economically marginal? The strongest case for an AI-driven recovery earns its claim the hard way: by specifying exactly what conditions must be met.

Paul David’s 1990 paper “The Dynamo and the Computer” provides the framework. US factories began adopting electric motors in the 1880s. Manufacturing productivity did not accelerate until the 1920s — a lag of roughly forty years. The reason was not that electricity was overhyped. The reason was that factory managers who replaced steam engines with dynamos kept the existing physical layout: one central power source driving machines through shafts and belts. The efficiency gains of individual electric motors were largely wasted because the factory was organized around the constraints of mechanical power delivery.

The breakthrough came when a new generation of managers — unburdened by sunk costs in existing layouts — designed factories from scratch around unit-drive systems: individual motors on each machine, enabling single-storey floor plans that followed the flow of materials and labour rather than the constraints of shaft-and-belt power distribution. Factories could suddenly be organized horizontally, spread across single floors, with production sequences arranged by logic rather than proximity to a power shaft. The technology was the same. The organizational structure was different. That difference was worth decades of productivity growth.

Brynjolfsson, Rock, and Syverson formalized this in their productivity J-curve paper: general-purpose technologies require large complementary intangible investments — workflow redesign, workforce retraining, organizational restructuring — before productivity gains materialize. The investment phase actively depresses measured TFP before gains appear. Applied to AI, the diagnostic question becomes specific: are firms still in the steam-engine-in-old-factory phase — deploying AI into unchanged workflows — or beginning to build new organizational structures around AI capabilities?

The micro-level evidence, though, is real and specific. Brynjolfsson, Li, and Raymond, in a 2025 Quarterly Journal of Economics study of 5,172 customer support agents, found AI tool access increased issues resolved per hour by approximately 15% on average, with substantially larger gains for novice workers. Peng and co-authors found developers with GitHub Copilot access completed a standard coding task 55.8% faster than a control group. Task-level time savings are genuine and, in some settings, large.

But whether they aggregate is a different question entirely. A 15% improvement in customer support resolution rates, at one company, does not make GDP move. A 55.8% coding speed improvement on a controlled task does not make GDP move. These are firm-level or task-level findings. For them to produce the kind of aggregate productivity acceleration that would reverse the stagnation, the savings have to translate into additional output across the economy — not just faster completion of existing work.

And the mechanism for that translation is not automatic. Saving a quarter of a developer’s week raises firm-level output only if developer time is the binding constraint. If the bottleneck is product decisions, testing infrastructure, or deployment pipelines, the saving disappears into reduced overtime or longer lunches. The same logic applies to every sector: a faster insurance claim reviewer produces more output only if claim volume was the constraint, not regulatory approval. A faster legal researcher produces more billable work only if the firm bills for speed rather than hours.

The dynamo analogy tells you what to look for: not task-level acceleration, but systemic reorganization. Not faster typing, but differently structured firms.

General-purpose technologies

Bresnahan and Trajtenberg defined general-purpose technologies in a 1995 Journal of Econometrics paper by three criteria: pervasiveness across sectors, continuous technical improvement, and the spawning of complementary innovations. Steam, electricity, and computing all qualify. Whether current AI qualifies on all three — or only the first two — remains genuinely open. Pervasiveness is evident. Technical improvement is rapid. But the spawning of complementary innovations — new industries, new organizational forms, new products that could not have existed without AI as infrastructure — is the criterion that would confirm GPT status. Evidence of it remains anecdotal rather than systematic.

What the evidence actually shows

The evidence divides into three categories that do not point the same direction, and flattening them into a single signal is analytically incoherent.

Observed aggregate data: US private nonfarm business labor productivity grew 2.8% in 2024 and 2.2% in 2025. TFP in the private nonfarm business sector grew 0.8% in 2025 — down from 1.5% in 2024. OECD-wide labour productivity averaged approximately 0.4% growth in 2024. No aggregate data series through 2025 shows a structural break consistent with economy-wide AI-driven acceleration. The 2024 US figure reflects a cyclical rebound, not a trend shift.

Survey-reported time savings: the St. Louis Fed, drawing on November 2024 population survey data collected by Alexander Bick, Adam Blandin, and David Deming, found AI users reported saving 5.4% of work hours — implying a 1.1% to 1.4% aggregate productivity gain for the whole workforce if those savings represent additional output rather than absorbed slack. A subsequent November 2025 report by the same authors found work adoption had reached 37.4% of US workers — up from 33.3% the previous year — with AI users spending 5.7% of work hours on AI tasks. That adoption rate exceeds the rate of personal computer adoption at the same point in its diffusion curve. Note the distinction: time savings and time spent are different metrics. Neither directly measures output. If the time-savings figure is accurate and represents additional production rather than absorbed slack, it should be beginning to appear in aggregate data. In observed BLS statistics through 2025, it has not done so discernibly.

Model projections: the Penn Wharton Budget Model estimates AI’s contribution to TFP growth at approximately 0.01 percentage points in 2025 — minimal, because most firms have yet to deploy AI tools at scale. Their forward projections suggest higher contributions beginning around 2027, peaking in the early 2030s. But these are model outputs under specific assumptions. They are not observations.

Three explanations are consistent with this pattern. The J-curve: productivity is temporarily depressed by AI implementation costs and organizational disruption before gains materialize, as happened with electrification — in which case patience is warranted and the payoff is coming. Slack absorption: task-level time savings are real but freed time goes into reduced effort rather than additional output — in which case the technology works but the economic structure does not translate individual time savings into aggregate production. Price competition: gains are being competed away into lower output prices rather than appearing as measured productivity growth — in which case the gains are real but invisible to TFP accounting in the same way that digital consumer surplus is invisible to GDP.

The data does not yet resolve between these. But the framework built in previous sections tells you what resolution would look like. Not individual workers finishing tasks faster. Not survey respondents reporting time savings. Those are necessary but insufficient — and they are exactly what you would see in both the optimistic scenario (J-curve before breakout) and the pessimistic one (task-level gains that never aggregate). They cannot distinguish.

What the electrification story says “working” looks like is different and specific: sustained TFP gains across sectors where AI adoption is three or more years mature. Productivity improvements that hold across multiple business cycle phases rather than reflecting cyclical variation. Evidence of organizational redesign at scale — new workflows, restructured teams, redesigned production processes, entirely new roles and business models that could not have existed before — rather than tool adoption layered over unchanged structures. Firms building the single-storey factory, not bolting the electric motor onto the existing shaft.

None of that is visible yet in aggregate data. Individual examples exist. Case studies proliferate. But the macro signal — the economy-wide structural break that would show the electrification pattern repeating — has not appeared. That does not mean it won’t. It means the bet has not been won.

The loop

Solow’s sentence keeps getting recycled because the diagnostic error keeps getting repeated. Each generation of technology produces visible, impressive, culturally dominant change — and each generation’s observers conflate that visibility with the specific structural reorganization that drives aggregate productivity. Computers everywhere, but not in the statistics. AI everywhere, but not in the statistics. The paradox is only paradoxical if you believe adoption is the same thing as transformation.

It has never been the same thing. The special century remade daily economic life so thoroughly that a time traveller from 1870 would not recognize 1940. And when transformation actually happens — not adoption, but the structural reorganization that adoption eventually enables — it does not require sophisticated measurement to detect.

Whether AI breaks the loop depends on whether firms eventually do what factory managers did in the 1920s — not adopt the technology, but reorganize around it. Build new structures that could not have existed without it. Eliminate entire categories of work rather than making existing work slightly faster. The conditions are specific. The historical precedent shows how rarely they are met — and how unambiguously visible it is when they are. The special century was not ambiguous. The 1996–2004 surge was not ambiguous. A real structural break does not require faith. It shows up.

We are not there yet. That is not a verdict on AI’s future. It is a reading of the present — and a precise specification of what would have to change for the next reading to be different. The statistics, for now, are telling the truth.

Gen AI Disclaimer

Some contents of this page were generated and/or edited with the help of a Generative AI.

Media

Hyundai Motor Group – Pexels

Key Sources and References

Robert Solow, “We’d Better Watch Out,” New York Times Book Review, July 12, 1987, p. 36.

Erin Henry, Pierre-Daniel Sarte, and Jack Taylor, “The Productivity Puzzle: AI, Technology Adoption and the Workforce,” Federal Reserve Bank of Richmond Economic Brief No. 24-25, August 2024. https://www.richmondfed.org/publications/research/economic_brief/2024/eb_24-25

David Smith, “You See AI Everywhere, Except in the Productivity Figures,” The Sunday Times, January 19, 2025.

Robert J. Gordon, The Rise and Fall of American Growth: The U.S. Standard of Living Since the Civil War, Princeton University Press, 2016.

Tyler Cowen, The Great Stagnation: How America Ate All the Low-Hanging Fruit of Modern History, Got Sick, and Will (Eventually) Feel Better, Dutton/Penguin, 2011.

U.S. Bureau of Labor Statistics, Productivity and Costs news releases, 2024–2026. https://www.bls.gov/productivity/

OECD, Compendium of Productivity Indicators 2025. https://www.oecd.org/en/publications/oecd-compendium-of-productivity-indicators-2025_b024d9e1-en.html

Erik Brynjolfsson, Avinash Collis, W. Erwin Diewert, Felix Eggers, and Kevin J. Fox, “GDP-B: Accounting for the Value of New and Free Goods in the Digital Economy,” NBER Working Paper 25695, 2019. https://www.nber.org/papers/w25695

Chad Syverson, “Challenges to Mismeasurement Explanations for the US Productivity Slowdown,” Journal of Economic Perspectives, vol. 31, no. 2, Spring 2017, pp. 165–186. https://www.aeaweb.org/articles?id=10.1257/jep.31.2.165

Paul A. David, “The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox,” American Economic Review: Papers and Proceedings, vol. 80, no. 2, May 1990, pp. 355–361.

Erik Brynjolfsson, Daniel Rock, and Chad Syverson, “The Productivity J-Curve: How Intangibles Complement General Purpose Technologies,” American Economic Journal: Macroeconomics, vol. 13, no. 1, January 2021, pp. 333–372. https://www.aeaweb.org/articles?id=10.1257/mac.20180386

Erik Brynjolfsson, Danielle Li, and Lindsey Raymond, “Generative AI at Work,” Quarterly Journal of Economics, vol. 140, no. 2, 2025, pp. 889–942. https://academic.oup.com/qje/article/140/2/889/7990658

Sida Peng, Eirini Kalliamvakou, Peter Cihon, and Mert Demirer, “The Impact of AI on Developer Productivity: Evidence from GitHub Copilot,” arXiv:2302.06590, February 2023. https://arxiv.org/abs/2302.06590

Alexander Bick, Adam Blandin, and David Deming, “The Impact of Generative AI on Work Productivity,” Federal Reserve Bank of St. Louis, On the Economy blog, February 27, 2025. https://www.stlouisfed.org/on-the-economy/2025/feb/impact-generative-ai-work-productivity

Alexander Bick, Adam Blandin, and David Deming, “The State of Generative AI Adoption in 2025,” Federal Reserve Bank of St. Louis, On the Economy blog, November 13, 2025. https://www.stlouisfed.org/on-the-economy/2025/nov/state-generative-ai-adoption-2025

Penn Wharton Budget Model, “The Projected Impact of Generative AI on Future Productivity Growth,” September 8, 2025. https://budgetmodel.wharton.upenn.edu/issues/2025/9/8/projected-impact-of-generative-ai-on-future-productivity-growth

Congressional Budget Office, “Labor Productivity: Developments Since 1995,” March 2007. https://www.cbo.gov/sites/default/files/110th-congress-2007-2008/reports/03-26-labor.pdf

Timothy F. Bresnahan and M. Trajtenberg, “General Purpose Technologies: ‘Engines of Growth’?” Journal of Econometrics, vol. 65, no. 1, January 1995, pp. 83–108.