The U.S.-Iran War and Its Economic Fallout | The High Top
July 20, 2026 • 10:00 – 10:15 am EDT
America’s Unproven Productivity Boom, and the Cost of Misreading the Data
Photo: Wallap Puraya/Adobe Stock. Image was generated with AI.
Last week, tech stocks tumbled as market participants seemed to reconsider valuations “underpinned by a rosy, AI-enabled expectation for the future that is yet to be realized.” But equity markets are not the only place AI optimism is being priced in. Increasingly, projections of aggregate growth treat the gains from AI as already material, based on one reading of measured labor productivity. AI may yet revolutionize the economy, but that verdict is not in, and the current data is too thin to settle the question.
At the aggregate level, that optimism rests largely on a single number. Measured U.S. labor productivity grew about 1.3 percent per year from 2013 to 2019; since 2022 it has nearly doubled, to 2.4 percent. Technology optimists read the jump as an early dividend of new technologies diffusing through the economy. The evidence is far less clear: On balance, the sectoral pattern looks less like a technology dividend than a routine labor-market adjustment, with firms unwinding previously hoarded labor—the workers they kept on payroll through 2020 and 2021 in anticipation of a demand rebound that was slow to arrive. That is not to prejudge technologies that may yet prove transformative, but to caution against an overly optimistic reading of the current data, one that could lead to serious policy missteps.
One way that premature optimism translates into misguided policy is through government revenue projections. Forecasters have already begun to bet on the optimists’ reading. In February 2026, the Congressional Budget Office for the first time added a productivity adjustment to its baseline, an assumption that AI will measurably raise future output per worker. Those gains are now being counted on to offset self-imposed drags on growth, from trade frictions to immigration restrictions. If the dividend fails to materialize, those policies will prove far more expensive than advertised. Should the increased labor productivity prove transitory, the error could add roughly $2.2 trillion to federal borrowing over the next decade. Given those stakes, and the cost of guessing wrong, policymakers should be wary of spending against optimism and should invest now in the data infrastructure needed to answer what we still don't know about AI's economic impact.
What the Data Shows
Labor productivity is measured by output per hour. It rises when the economy produces more, and it also rises when the same output is produced with fewer hours. While both patterns can occur due to technological change, the second pattern can just as easily be a symptom of a labor market adjusting after a cyclical disruption: as firms shed the excess hours they carried out of the downturn, output per hour rises with no new technology behind it. Simply observing high productivity growth is not enough to tell whether technical change is behind it.
Begin with the economy-wide arithmetic. Relative to its 2013–2019 pace, labor input growth slowed by about 1.3 percentage points per year through 2025. Output growth, meanwhile, barely moved: firms kept output growing at its earlier pace while adding far fewer hours of work. That alone could be consistent with new technology reducing the need for labor in production. But capital-input growth also fell (see Figure 1A), which is difficult to square with large-scale investments in labor-augmenting technology unless intangible capital is being dramatically mismeasured. The decline in the growth of labor inputs is large enough to account for the entire measured pickup in multifactor productivity, leaving the technology residual slightly negative (see Figure 1B). Whether this represents a durable gain or a passing cyclical adjustment is therefore central to the outlook for U.S. productivity and output.
Remote Visualization
An independent benchmark points the same way with different data. The Federal Reserve Bank of San Francisco publishes a utilization-adjusted measure of total factor productivity that removes the effect of simply working existing workers and equipment harder. Stripped of that effect, almost nothing is left: over the four quarters ending in the first quarter of 2026, utilization-adjusted total factor productivity grew just 0.07 percent, against measured labor productivity growth of roughly 2.5 percent over the same period. Utilization is a different cyclical margin from the hours channel emphasized above, and it yields the same conclusion. What looks like a technology boom on the headline number does not survive either adjustment.
This Has Happened Before
While the aggregate accounting and utilization evidence is suggestive, it still leaves many questions open. A natural place to look for more clarity is past periods of technology-induced productivity booms.
Start with the clearest recent technology boom, the internet build-out, a decade-defining stretch of investment and adoption that cemented U.S. leadership in the digital economy. The current data looks nothing like that episode.
The simplest way to tell the two cases apart is to look at output growth and productivity growth together (see Figure 2). In the 1990s boom, output itself accelerates. From 1995 to 2000, output grew about 5 percent per year, well above its 3 percent norm, with productivity rising alongside. By contrast, productivity spikes after the 2001 and the 2009 Global Financial Crisis (GFC)—the latter famously led to a “jobless recovery” and saw output grow at a more tempered pace, while measured productivity growth came from labor inputs falling faster than output. From an accounting perspective, these are both productivity enhancements, but from an economic view, they are extremely different. Today’s productivity growth looks like 2009, not the 1990s.
Remote Visualization
One might object that the right analogy for 2022–2025 is the quiet early period of the personal computer, before the gains surfaced in aggregate data. That comparison does not rescue the optimistic reading. From 1987 to 1995, personal computers were already common in offices, output grew about 3 percent per year, productivity about 1.5 percent, and hours were still rising. Today hours are falling. The early-innings analogy gets the direction of the labor input backwards.
The 1990s acceleration was led by output from its first year; this one is still led by falling hours three years in. What has persisted is the headline productivity number, which slower hiring inflates.
Where the Gains Aren’t
Aggregate decompositions and historical analogies can only go so far. To get a better sense of the underlying drivers, we need to look beneath the aggregate: which sectors the productivity growth is coming from, and whether the pattern matches technological adoption or some other story.
If AI were driving the aggregate numbers, the industries that use it most intensively should show the largest gains. That is not the case.
Of six measures of industry AI exposure, including a usage-based index built from Anthropic’s data on roughly two million real Claude conversations, none predicts which industries compressed labor after 2022 (see Figure 3), nor which saw measured productivity accelerate through 2024, the latest year with published industry detail.
Remote Visualization
If AI is not the driver, what is? The cross-industry evidence points to hoarded labor being released (see Figure 3). If the post-2022 compression in hours is firms shedding workers they had over-retained, it should be deepest where over-retention was deepest. It is. The same cross-industry design used for the AI exposure measures, now with a hoarding measure in place of AI, yields what none of those measures produced: a clear, positive, and statistically significant relationship between an industry’s labor over-retention and its subsequent labor compression.
This implies that the industries that held onto workers through 2021 are the ones whose hours growth compressed most from 2022 on. This is the opposite of what a simple fire-and-rehire story predicts, in which firms that shed labor in the downturn and rehired would show a rebound, not a contraction. A contraction concentrated in the biggest hoarders is the signature of over-retention being unwound. If this is the driving mechanism behind the increase in labor productivity experienced since 2022, there is no reason to believe these gains will be persistent, as they are the result of a cyclical adjustment in the U.S. labor market.
Further evidence comes from a horse race: when an industry’s over-retention and each of the six AI exposure measures are entered in the same regression, over-retention remains a strong predictor of which industries compressed labor, while no AI measure does (see Figures 3 and 4).
While a null result like this does not prove that AI is not having significant firm-level impacts that the aggregate data cannot yet see, what it can rule out are effects large enough to move whole industries by 2024, well into the acceleration of labor productivity, and that is precisely the effect the aggregate forecasts now assume.
Remote Visualization
Many questions about the source of current productivity growth, and its future path, remain open. But the evidence assembled here does not support treating a technology-driven boom as the working assumption. The prudent course is not to spend as though it were, but to invest in the better data that would let us know.
The Cost of Miscalculation and the Value of Better Measurement
Whether that boom is real is not an academic dispute. Growth is the resource base for U.S. economic power and statecraft. The productivity assumptions embedded in official economic forecasts dictate the projected revenue path, which in turn sets the bounds for what the United States can comfortably afford across defense, industrial strategy, social services, and sovereign debt service.
According to the Congressional Budget Office’s published sensitivity analysis, a sustained 0.1-percentage-point shift in productivity growth moves the cumulative federal deficit by roughly $300 billion over a 10-year horizon: faster productivity narrows the deficit, slower productivity widens it. If half of the post-pandemic labor compression unwinds as hiring normalizes, the implied forecasting error is about $1.1 trillion in additional borrowing; a full unwind of the labor-driven part of the acceleration runs to about $2.2 trillion.
Over the coming decade, the United States will face heavy capital demands to rebuild infrastructure, secure critical supply chains, and resource a more contested global security environment. These initiatives will require trillions of dollars in upfront funding, much of it financed by federal borrowing. The long-term sustainability of the country’s debt burden depends heavily on the economy’s underlying structural growth rate. While these new technologies may provide material assistance in paying for these investments, the productivity growth that many are banking on is not yet supported by the data.
One cheap hedge against this expensive error would be better measurement. The federal statistical system that produces these estimates is deteriorating, and many industries that will be exposed to new technology have far less granular data available. The AI DATA Act, a bipartisan Senate bill, would broaden federal data collection on AI and require an annual report on the technology’s effect on the labor force. Against a plausible forecasting error measured in the trillions, funding the statistics that would resolve it is among the highest-return investments available.
Until the data starts to provide a clearer story, policymakers would be wise to consider the jury still out.
Note: Unless otherwise specified, all figures are based on author analysis of economic and AI sector data from various sources. Details and replication code available upon request.
Philip Luck is director of the Economics Program and Scholl Chair in International Business at the Center for Strategic and International Studies in Washington, D.C.