The Productivity Paradox: A History from the Dynamo to AI

By Victor Hoang · Updated Oct 24, 2026

Picture the Paris exposition of 1900. Visitors saw enormous dynamos massed on the Champ de Mars. The city's outdoor festivals ran on electric light into the night. The technology looked like a finished promise.

Yet inside most factories, power still moved the old way, and the economic statistics hadn't noticed the dynamo. That gap is where this story starts. The productivity paradox is the finding that a powerful new technology can spread everywhere before it shows up in measured output. The name arrived in the late 1980s, when economist Robert Solow wrote that the computer age was visible everywhere but in the productivity statistics. The puzzle itself is older. Economist Paul David showed in 1989 that electric dynamos had the same problem around 1900. The payoff came about forty years after the first power stations opened.

That's the throughline of this history: each time, the machine arrived first and the redesign of work arrived later. The statistics were not lying, and neither were the people who saw the machines. They were looking at two different clocks. If you want the definition first, start with what productivity is. For the wider story, see the history of productivity.

Timeline at a Glance

The Dynamo Lag (1880s-1920s)

Electric power arrives early and pays off late

This is the precursor, not the paradox itself. Nobody in 1900 called it a productivity paradox, and the word wasn't used for another eighty-odd years. But the pattern was already there, and David found it.

Walk into an American factory in the 1890s and you would see one power source and one way of moving it. A steam engine or water wheel turned a main shaft, and belts carried the motion to every machine. When electricity arrived, most owners didn't rebuild that system. David, quoting the historian Harry DuBoff, describes how innovating plants kept their belts and shafting and installed motors to drive separate shafting sections. This "group drive" was cheaper at first because the steam engine and shafting could be kept. David dates that phase from the mid-1890s to the eve of the 1920s.

That's why electricity could be everywhere and still not show. David reports that during the 1890s most U.S. engineers advised against running any but the largest machines with their own motors, and that by the middle of the next decade informed opinion had shifted toward individual, or "unit," drive, with a motor on each machine. The real gain came when factories could be laid out around the flow of work instead of the line of the shaft. David notes that the industrial cost of electric power fell dramatically between 1907 and 1917, and that the wave of unit drive and redesigned plants came in the 1920s. In his account, the full transformation did not come until more than half of factory mechanical drive capacity had been electrified, some forty years after the first central power stations began operating.

The result was visible in the numbers. David reports that total factor productivity growth in U.S. manufacturing sped up by about 5 percentage points between 1909-19 and 1919-29. He also cautions that electrification explains only part of it, because continuous-process industries such as oil refining and chemicals were improving at the same time. The lesson holds without overclaiming: the technology was ready decades before the work was.

David first set this out in a 1989 working paper at Stanford, a revised version of a paper presented at an OECD seminar in Paris. His abstract made the comparison directly: in 1900, contemporaries might well have said the dynamos were to be seen "everywhere but in the economic statistics". The shorter, published version came a year later in the American Economic Review's May 1990 Papers and Proceedings, as cited in Bakos's 1998 review in Science. The analogy is imperfect, and David said so, but it gave economists a pattern to look for. For a longer look at how work itself changed in this period, see the history of knowledge work.

What this means for teams today: a tool bolted onto an old process earns an old result. The gain is in the new layout of the work.

The Missing Statistic (1970s-1987)

Computers multiply while growth slows

By the 1970s the economy had two facts that didn't fit. Computing was getting cheaper and more common every year. And measured productivity growth was falling. Robert Gordon's 2012 NBER paper describes about 80 years of rapid productivity growth from 1890 to 1972, followed by a slowdown during 1972-96. The slowdown arrived just as the mainframe gave way to the minicomputer and then the desktop machine.

Computers were moving from the data center to the desk, and the national statistics showed no matching payoff. Erik Brynjolfsson later put numbers on the mismatch: delivered computing power in the U.S. economy rose by more than two orders of magnitude from 1970, yet productivity, especially in services, appeared to have stagnated. For the software side of that decade, see the history of office software, and for the rooms those desks sat in, the history of the office.

Then the growth economist Robert Solow wrote the sentence that stuck. In his review "We'd better watch out," published in the New York Times Book Review on July 12, 1987, he wrote: "You can see the computer age everywhere but in the productivity statistics." Solow's earlier growth work is part of why the remark carried weight, and we cover it in how productivity is measured.

What this means for teams today: spending is not output. If you can see the tool everywhere and the result nowhere, ask what had to change in the work and hasn't.

The Paradox Named (1987-1995)

Solow's line becomes a research question

The remark turned into a field of study within a few years. By 1990, a magazine could write that computers did not boost productivity "at least not most of the time", a line the Economist ran that year and Brynjolfsson quoted. CEOs and line managers were asking why their computer budgets weren't paying back. Economists were asking whether the question was even framed correctly.

Two scholars gave the debate its two main answers. David's answer was historical: wait, and look at how long the dynamo took. Brynjolfsson's answer was a sorting exercise. His paper circulated in December 1991 and September 1992 drafts and was published in Communications of the ACM in December 1993. It listed four explanations: measurement errors, lags, redistribution of gains between firms, and mismanagement. It also warned against over-reading the evidence. His line: "a shortfall of evidence is not necessarily evidence of a shortfall." That distinction is still useful. A missing number tells you the measurement failed or the effect is absent. It doesn't tell you which.

Jack Triplett's March 1999 Brookings review took the sorting further. He worked through seven explanations, including that computers were a small share of the capital stock, that output in finance and insurance was poorly measured, that some benefits went uncounted, and that productivity gains would lag the way electricity's had. The last one is David's argument. By the end of the 1990s the paradox had a literature, and the literature agreed on one thing: the answer would not be a single cause.

What this means for teams today: before declaring a tool a failure, separate four questions. Is it measured correctly, is it too early, did the gain go elsewhere, or was it run badly?

The Payoff Years (1995-2004)

The statistics finally move

Then the numbers turned. In May 2000, Stephen Oliner and Daniel Sichel of the Federal Reserve Board reported a rebound in labor productivity growth after nearly a quarter century of sluggish gains. Their estimate: IT use and computer production together accounted for about two-thirds of the one-percentage-point step-up between the first and second halves of the 1990s. Their conclusion was blunt: IT largely was the story.

Mun Ho, Dale Jorgenson and Kevin Stiroh reached a compatible view in a 2002 Federal Reserve Bank of Atlanta paper. They found that IT played a substantial role in the U.S. revival, with productivity growth averaging 2.36 percent a year from 1995 to 2000, and projected 2.21 percent for the following decade. The computers had been spreading since the 1970s, and the measurable payoff came in the second half of the 1990s. That sequence matched David's pattern of a long lag between the machines and the result. But the revival was brief. Gordon's 2012 paper describes only a short-lived revival from 1996 to 2004. The paradox had a resolution, and the resolution didn't last. For the matching story of how that decade changed office work, see the productivity timeline.

What this means for teams today: gains tend to arrive after the process changes, not after the purchase. Give a rollout enough time to show the effect, and measure the process, not the license count.

The Slowdown (2005-2019)

The revival fades and the doubt returns

By the early 2010s, the fast years looked like the exception. Gordon's August 2012 paper argued that the third industrial revolution, computers and the web, had produced a short revival and that headwinds such as demography, education and debt could push long-term growth to half or less of the 1.9 percent a year seen from 1860 to 2007. It was a forecast, and mismeasurement remained one of the rival explanations.

The government data showed the slowdown, whatever its cause. In the BLS release that covers the whole period, nonfarm business labor productivity grew at 1.5 percent a year from Q4 2007 to Q4 2019, compared with a long-term rate of 2.1 percent since the first quarter of 1947. Smartphones, cloud software and collaboration tools spread through offices in those years while productivity grew more slowly than its long-run pace. It was Solow's gap again, with a different set of machines.

The 2017 paper by Brynjolfsson, Rock and Syverson put the new gap in plain terms. Measured productivity growth had roughly halved over the past decade, while machine learning systems were beginning to match human performance on narrow tasks. They offered four explanations, false hopes, mismeasurement, redistribution and implementation lags, and judged lags the most important. In October 2018 the same authors published a model of why. Their Productivity J-Curve says that general-purpose technologies need intangible investments, such as new processes and skills, that national accounts capture poorly. Measured productivity growth is understated early and overstated later. When they adjusted for software and hardware intangibles, they found total factor productivity about 11.3 percent higher than official figures at the end of 2004, and 15.9 percent higher at the end of 2017.

What this means for teams today: some of the work of change is invisible in the numbers. Retraining, redesign and new roles look like cost before they look like output, so plan to defend them.

The AI Return (2017-present)

The old argument gets a new technology

This era overlaps with the slowdown, because the same economists carried the argument forward. The 2017 NBER paper was written when generative AI was still a research topic, and it already argued that "Artificial Intelligence and the Modern Productivity Paradox" was a clash of expectations and statistics. Then the tools arrived at scale.

The early evidence was mixed, and it was measured at the task level. In an April 2023 working paper, Brynjolfsson, Danielle Li and Lindsey Raymond studied the staggered rollout of a generative AI assistant to 5,179 customer support agents. Issues resolved per hour rose 14 percent on average and 34 percent for novice and low-skilled workers. Two years later, METR ran a randomized trial with 16 experienced open-source developers working on 246 real issues. They took 19 percent longer with AI tools, though they had expected a 24 percent speed-up and afterward believed AI had made them 20 percent faster. METR has since published updated data on later tools, so treat that result as a snapshot of early 2025.

Both findings fit the old pattern. The support agents gained most when they had the least experience. The developers worked in large, mature codebases they had contributed to for years. Neither study says anything about the economy as a whole. For that, the latest BLS release shows nonfarm business productivity growing at 2.1 percent annualized from Q4 2019 through Q2 2026, above the 1.5 percent of the previous cycle and equal to the long-term rate since 1947. That is a better number than the slowdown years produced. The release doesn't attribute that rate to AI, and the lesson of this history is to be careful about credit in either direction. For the enterprise version of the question, with adoption and financial-impact figures, see the AI productivity paradox.

What this means for teams today: measure a specific workflow before and after the change, and keep perception separate from results. People can feel faster and be slower.

Key People

Key Facts

Key Facts: The Productivity Paradox

• Electric drive reached more than half of factory power capacity about forty years after the first central power stations, according to David's 1989 working paper.
• Solow's line ran in the New York Times Book Review on July 12, 1987, under the title "We'd better watch out."
• Oliner and Sichel estimated that IT use and computer production accounted for about two-thirds of the 1 percentage point step-up in productivity growth in the second half of the 1990s.
• BLS data show nonfarm business productivity growth of 1.5 percent a year from Q4 2007 to Q4 2019, against 2.1 percent since 1947.
• The Productivity J-Curve authors estimate that adjusted total factor productivity was 15.9 percent higher than official figures at the end of 2017.

The Pattern Behind the History

Across 140 years, the same four steps repeat. We call this the Lag and Rebuild pattern, our own synthesis and not a published model.

1. The machine arrives first. Power stations in the 1880s, mainframes in the 1960s, smartphones in the 2000s, generative AI in the 2020s.
2. The old process absorbs it. Motors on old shafts, terminals on old paper workflows, chatbots on old approval chains.
3. The statistics wait. Output is flat, costs are real, and skeptics sound right.
4. The rebuild pays. Unit drive in the 1920s, integrated data in the late 1990s. Then the numbers move, usually credited to something else.

Each cycle also adds a measurement argument. Every era has argued about whether the gains were missing or just unmeasured, and the honest answer has usually been some of both.

What It Teaches Teams Today

The history gives a team four practical rules.

• Plan for the lag. Budget for years of change, not one quarter of license savings.
• Rebuild the workflow, not just the tool. The dynamo paid off when factories were laid out around motors. Ask what the process would look like if the tool had been there from the start.
• Measure the process. Track cycle time, error rate or cost per ticket against a baseline, not logins and seats.
• Separate feeling from result. The METR developers believed they were faster. A controlled before-and-after comparison protects you from that error in either direction.

The paradox isn't an argument for patience alone. It's an argument for doing the unglamorous rebuild while others wait for the statistics to move. The same discipline appears in the productivity traps that individuals fall into, and the long view is in the history of computers.

The productivity paradox is less a puzzle than a schedule. Technology arrives on one clock and work changes on another, and the teams that close the gap soonest are the ones that start redesigning before the numbers say they should.