Future Horizons

The Productivity Mirage: Why AI Has Not Yet Moved the Economic Needle — and When It Will

US productivity grew 2.2% in the year to Q2 2026, exactly the long-run rate. Meanwhile Epoch AI puts computing infrastructure at about 1.5% of GDP, roughly double its pre-boom share — either the J-curve working, or the growth itself.

Gabriele Masetti · 2026-08-21

The Trillion-Dollar Question Nobody Can Answer

Something puzzling has happened at the intersection of AI and economics. The world has witnessed the fastest adoption of a transformative technology in recorded history — ChatGPT reached 100 million users in two months, a pace that took the internet four years and television thirteen. Hundreds of millions of people now use AI tools regularly. Trillions of dollars have been invested in AI infrastructure, with no sign of slowing.

Major corporations have restructured entire business functions around AI capabilities. And yet, when economists and statisticians look for this transformation in the data that is supposed to capture economic productivity — GDP growth, total factor productivity, output per hour worked — they largely cannot find it.

This is the productivity paradox, and it is not a new phenomenon. In 1987, Nobel Prize-winning economist Robert Solow noted that "you can see the computer age everywhere but in the productivity statistics" — a wry observation that became famous as the Solow Paradox and that haunted the economics of computing for nearly two decades until the late 1990s, when computer-driven productivity growth finally became unmistakable in the data. That lag was approximately 25–35 years from the commercialisation of computers.

Now economists are asking, with growing concern, whether AI has created an identical phenomenon. The evidence suggests it has — but the question of whether the lag this time will be shorter, longer, or different in character is where expert opinion diverges sharply.

The Evidence for Non-Impact

Let us be precise about what the data actually shows, because the claim "AI has not moved the productivity needle" is stronger than it might initially appear.

The United States Bureau of Labor Statistics measures non-farm labour productivity — output per hour worked — at quarterly intervals. From 2022 through 2024, the period of maximum AI deployment acceleration, US labour productivity grew at an average annual rate of approximately 2.0% — unremarkable by historical standards and certainly not suggestive of a transformative technology-driven shock. In the European Union, productivity growth during the same period was even lower, in the range of 0.5–1.0%.

The series has not broken since. In the year to the second quarter of 2026, non-farm business productivity rose 2.2%, with output up 2.5% and hours worked up 0.2%; the quarter itself came in at a 1.4% annualised rate. Measured across the whole current business cycle, from the fourth quarter of 2019 through the second quarter of 2026, productivity has grown at an annualised 2.1% — the same as the long-term average since 1947, and above the 1.5% of the previous cycle.

That is the shape of the problem in one number. Four years into the fastest technology adoption on record, the productivity series looks like the productivity series.

A 2025 working paper from the National Bureau of Economic Research, surveying 6,000 chief executive officers and chief financial officers across the United States, United Kingdom, Germany, and Japan, found that the vast majority reported minimal impact of AI on their organisations' overall performance. The exceptions — a subset of technology-intensive firms in specific sectors — were consistent with normal early-adopter dynamics rather than economy-wide transformation.

The Yale Budget Lab modelled the question in 2026 and declined to forecast. It ran the moderate-adoption scenario from a survey of economists, in which labour productivity grows 2.5% a year from 2025 to 2030 against an average of 1.8% from 2015 to 2025, and described its own output as "illustrative demonstrations of stylized scenarios, rather than authoritative forecasts". Industry advocates have suggested that AI could add 10–20% to global GDP over the same period. The McKinsey Global Institute has estimated AI's economic potential at $6.1–7.9 trillion annually once fully diffused — a number that sounds large until you note that it represents roughly 6–8% of current global GDP, spread over 10–20 years of diffusion.

Academic assessments from leading economists are sobering. Daron Acemoglu at MIT published a provocative 2024 analysis arguing that the type of tasks AI is currently automating — primarily routine cognitive tasks, administrative work, customer service, basic information processing — are not the highest-value tasks in the economy. His modelling suggested that current AI technologies, even if fully diffused across the economy, would add 0.5–1.0% to productivity growth over a decade: meaningful, but not transformative.

Acemoglu's analysis attracted intense pushback from AI optimists who argued that his model was insufficiently dynamic — that it failed to account for AI capabilities improving rapidly, for new applications emerging from current tools, and for the long-run productivity gains from investments currently being made. The debate, unresolved, captures the essential uncertainty: the productivity impact of AI depends fundamentally on assumptions about the trajectory of AI capabilities and the pace of economic adjustment, neither of which is known.

Erik Brynjolfsson and the J-Curve Theory

The most influential framework for thinking about the AI productivity paradox comes from Erik Brynjolfsson, whose career has been shaped by this problem. In 1993, Brynjolfsson — then at MIT Sloan School of Management, now at Stanford's Digital Economy Lab — published "The Productivity Paradox of Information Technology," which systematically examined the failure of IT investment to show up in productivity statistics in the 1970s and 1980s. His analysis identified the structural reasons for the lag and correctly predicted that productivity gains would eventually materialise.

Brynjolfsson is now advancing the J-curve hypothesis for AI: the idea that general-purpose technologies like AI follow a characteristic trajectory in which measured productivity initially declines or stagnates — because organisations must invest in complementary assets (skills, workflows, management structures, organisational redesign) before they can capture the technology's full potential — before eventually rising sharply once the complementary investments are in place.

The J-curve metaphor comes from the observation that structural transitions look like a J when plotted over time: initial decline or flat performance, then a sharp upswing. The key claim is that the flat part of the J is not evidence that the technology does not work; it is evidence that the economy has not yet made the complementary investments needed to harness it.

Brynjolfsson and co-authors Daniel Rock and Chad Syverson published an influential 2019 paper, "The Productivity J-Curve," which developed this framework systematically. They argued that the current AI investment wave is creating what they call intangible assets — organisational capital, human capital, and process redesign — that are economically valuable but not captured in standard GDP accounting. The result is a systematic underestimate of AI's current contribution to economic value, because the measurement framework is designed for tangible physical assets and does not capture intangible value creation.

This is an important point. When Amazon redesigns its entire warehouse logistics system around AI-guided robotics, the investment creates enormous value — but much of that value is in the redesigned process, the trained workforce, and the organisational knowledge of how to manage the new system. These intangibles do not show up in the national accounts in the same way that the physical robots do.

The Electrification Analogy: A 40-Year Wait

The most instructive historical precedent is electrification — the process by which electric power replaced steam power in factories and workshops during the late nineteenth and early twentieth centuries.

Electric motors became commercially available in the 1880s. By 1910, a majority of US factories had access to electricity. And yet the productivity gains from electrification did not appear in the aggregate productivity data until the 1920s — a gap of roughly 30–40 years. Paul David, the economic historian, documented this delay in a classic 1990 paper that provided much of the intellectual foundation for Brynjolfsson's IT productivity paradox analysis.

Why the delay? The answer, which is now well-understood, is that electrification required a wholesale redesign of the factory to realise its potential. Steam power drove a central shaft through the factory, which distributed power to individual machines via belts and pulleys. The layout of a steam-powered factory was determined by the geometry of power transmission — machines had to be clustered around the shaft, arranged in the order that made mechanical sense, not the order that made production sense.

Electrification made it possible for each machine to have its own electric motor. This was not just a substitution of one power source for another; it was an opportunity to redesign the entire factory layout around production logic rather than power-transmission logic. But most factory managers in 1890 did not redesign their factories when they switched to electricity. They replaced the steam shaft with an electric motor driving the same shaft, because that was the familiar organisational form. They got some efficiency gains from the cleaner, more reliable power source, but nothing like the productivity revolution that was theoretically available.

The full productivity potential of electrification was only realised in the 1920s, when a new generation of factory managers — who had grown up in an electrified world and had no prior commitment to the shaft-and-belt paradigm — designed factories from scratch around the new technology. These new factories, with individually motorised machines arranged in optimal production sequences, with conveyor belts and assembly lines made possible by flexible power delivery, were dramatically more productive than their predecessors.

The parallel to AI is suggestive but not perfect. The productivity revolution awaiting AI may not come from existing businesses gradually improving their processes, but from new businesses designed from the ground up around AI capabilities — businesses whose organisational structures, workflows, and management practices are native to AI in the way that new factories in the 1920s were native to electrification.

What Firm-Level Data Actually Shows

While aggregate productivity statistics show little AI impact, firm-level data — examining individual organisations that have adopted AI — is more nuanced and, in some cases, more encouraging.

A 2023 field experiment by researchers from Stanford, MIT, and the National Bureau of Economic Research examined the impact of an AI-assisted coding tool on software developers at a major technology company. Developers randomly assigned to use the AI coding assistant completed tasks 56% faster than the control group — a substantial productivity gain by any measure. A parallel study on AI-assisted customer service agents at a Fortune 500 company found that agents using AI assistance resolved issues 14% faster and received substantially higher customer satisfaction scores.

These are real productivity gains. The question is whether and how they scale up to the macroeconomy.

The gap between strong firm-level results and weak aggregate results suggests several things simultaneously: that AI does work in specific applications, that the benefits are highly unequally distributed, and that economy-wide diffusion and complementary investment take time.

Research by Brynjolfsson and his collaborators at the Stanford Digital Economy Lab has found systematic differences between firms that realise AI productivity gains and those that do not. The key differentiator is not the AI technology itself — it is whether the firm has made the complementary investments in organisational redesign, workforce training, and process change that allow the technology to be used effectively. Firms that simply add AI tools to existing workflows tend to see modest gains. Firms that redesign workflows around AI capabilities — restructuring job roles, retraining workers, redesigning processes — see much larger gains.

This is the organisational capital argument: the technology is necessary but not sufficient. The real investment is in the human and organisational adaptation.

The Measurement Problem: Is GDP Blind to AI Value?

A distinct but related argument questions whether standard economic statistics can actually capture AI's contribution to economic welfare, even when that contribution is real.

GDP measures market transactions — the exchange of goods and services for money. It systematically misses or underweights several categories of value that AI creates.

First, consumer surplus from free services: many of the most valuable AI applications — search, translation, information synthesis, creative assistance — are provided at zero or near-zero marginal cost to the consumer. The economic value consumers derive from these services may be enormous, but GDP captures only what is paid for them, not the consumer surplus.

A landmark 2018 study by Brynjolfsson, Avinash Collis, and Felix Eggers attempted to measure consumer surplus from free digital goods using "willingness to accept" surveys — asking people how much they would need to be paid to give up access to various digital services for a month. Their estimates were striking: the median respondent required $17,530 to give up internet search for a year, $8,414 to give up email, and $3,648 to give up maps and navigation. None of this value appears in GDP statistics, because these services are provided free.

As AI improves these free services — and creates new ones — the gap between GDP-measured economic output and actual economic welfare will grow. GDP will understate AI's contribution to welfare increasingly severely.

Second, quality improvements in existing goods and services: AI-improved products — better medical diagnoses, more personalised educational content, more efficient logistics — create value that GDP partially misses because it measures output volume rather than quality-adjusted output. If an AI-assisted doctor diagnoses twice as many conditions correctly per hour, the GDP impact is the same number of doctor-hours billed; the improvement in patient outcomes is invisible in the statistics.

Leading Indicators: What to Watch

Brynjolfsson and his team have identified a set of leading indicators that, historically, have preceded aggregate productivity gains from general-purpose technologies. These indicators are more encouraging than the aggregate statistics suggest.

Corporate restructuring and job redefinition: surveys of major US employers find rapid and accelerating changes in job descriptions, workflow designs, and organisational structures driven by AI — the kind of complementary investment that historically precedes productivity gains.

Capital investment in AI-adjacent infrastructure: data centre construction, GPU investment, and software spending in the US reached record levels in 2024 and 2025, and went further in 2026. The five largest hyperscalers guided to roughly $775–800 billion of AI infrastructure spending for the year, against about $429 billion in 2025. Epoch AI, measuring the US portion of the build-out, put AI-related data centre construction, compute hardware and networking equipment at about 0.8% of US GDP in the first quarter of 2026, which lifted computing infrastructure as a whole to about 1.5% of GDP, roughly double the 2015–2022 average of 0.7%.

Intangible investment growth: measures of intangible capital formation — training, R&D, software, organisational development — are growing faster than tangible investment, consistent with the J-curve theory that the current period is one of complementary investment rather than measured output growth.

Firm-level performance divergence: the gap between the most productive firms in each sector and the median firm is widening in industries where AI adoption is highest. That is consistent with early-stage technology adoption, where pioneers gain advantage before the technology diffuses to laggards.

The Indicator That Cuts Both Ways

The capital-spending indicator has grown large enough to argue against itself. One analysis of the Bureau of Economic Analysis advance estimate for the first quarter of 2026 attributed roughly half of the quarter's 2% annualised real growth to two line items: investment in computers and peripheral equipment, growing at a 67% annualised rate, and software investment, growing at 23%. Strip the AI build-out out of the arithmetic and the quarter looks closer to 1%.

A J-curve reading says that spending is complementary capital being laid down ahead of the gains, exactly as the theory predicts. A less comfortable reading says the build-out is the growth: that AI's measurable macroeconomic contribution so far is the act of buying the equipment, not anything the equipment does, and that a rational correction in capital expenditure would subtract from GDP immediately while the productivity payoff stayed wherever it has been hiding.

The two readings are not mutually exclusive, and the aggregate data cannot yet separate them. What can be said is that the J-curve is no longer the only serious explanation on the table for why the numbers look the way they do, and that the longer the flat part of the J runs, the more of measured growth depends on the flat part continuing.

The Honest Assessment: Uncertainty Above All

Anyone who tells you with confidence that they know when AI will deliver macroeconomic productivity gains, or how large those gains will be, is overstating their knowledge. The honest assessment is one of deep uncertainty bounded by historical analogies and structural reasoning.

What the historical analogies suggest is that the lag is real, measured in years to decades rather than quarters, and that it will end when organisational adaptation catches up with technological capability. The electrification analogy suggests a 30–40 year lag; the computing analogy suggests 25–35 years. If these analogies hold, we are in the early middle of the wait, not near its end.

What structural reasoning suggests is that the AI case may differ from these historical precedents in important ways. AI capabilities are improving much faster than either electricity or computing did at comparable stages of deployment. The potential applicability across the economy is broader. And the speed of information diffusion — through which successful AI adoption strategies spread from pioneers to followers — is much faster than in previous eras.

These factors argue for a shorter lag than historical precedent would suggest. But the organisational adaptation challenge remains real: changing how millions of businesses and billions of workers incorporate AI into their daily work is a social and institutional process, not just a technical one, and social and institutional change has its own pace that technology cannot simply accelerate.

The productivity revolution from AI is, on the available evidence, real but delayed. The question is not whether it will come, but when, how equitably it will be distributed, and whether our economic institutions are prepared to manage the disruption it will entail.