← All articles

The Productivity Paradox Is Back

In 1987, Robert Solow said you could see the computer age everywhere but in the productivity statistics. Nearly 6,000 executives just said the same thing about AI. Same paradox, different machine.

Bert Carroll ·
LinkedIn RSS

The Paradox, Named

In 1987, Robert Solow wrote in a New York Times Book Review piece that “you can see the computer age everywhere but in the productivity statistics.”1 The line landed as an aside inside a review of Cohen and Zysman’s Manufacturing Matters. The fuller sentence is sharper:

"What everyone feels to have been a technological revolution, a drastic change in our productive lives, has been accompanied everywhere, including Japan, by a slowing-down of productivity growth, not by a step up."
— Robert Solow, 1987

Not a plateau. A deceleration.

Solow never wrote the data version of that argument. He was a theorist, and the piece was a book review. The empirical spine came six years later from Erik Brynjolfsson, who coined the phrase “productivity paradox,” credited Solow directly, and ran the numbers: 1970s and 1980s IT investment against BLS and BEA output data.2 His four-part framework — mismeasurement, lag, redistribution, and mismanagement — still anchors every serious revisit of the question.

The paradox resolved in the late 1990s. Not because the technology got better. Because organizations learned to manage what the computers produced. The tools were mostly already there. The workflows, the management practices, the editorial discipline, those had to be built.

We are back here. Same paradox, different machine.

The 2026 Data

In February 2026, the National Bureau of Economic Research published findings from a survey of nearly 6,000 senior executives across the US, UK, Germany, and Australia.3 The headline: roughly 90% of firms report AI has had no measurable impact on productivity or employment over the past three years. Not some firms. Not laggards. 69% of firms actively use AI; more than two-thirds of executives use it personally, averaging 1.5 hours a week.

Adoption is real. Impact is not. Not yet.

A parallel ManpowerGroup survey of workers across 19 countries4 captured the mirror image. Regular AI use jumped 13 percentage points in 2025, to 45% of workers. Confidence in using the technology fell 18%, the first decline in three years. The steepest drops were among older cohorts. More usage. Less trust. 43% now expect automation to replace their job within two years, up five points from 2025.

Put together, the two surveys describe the same organization from two sides. Leadership is deploying faster than results are arriving. Workers are using the tools more and trusting them less. The productivity line is flat.

This is Solow at a different scale.

Activity Is Not Productivity

The top reported uses of AI in the NBER survey are text generation, visual content creation, and data processing. The most visible thing AI does is produce output. Documents, decks, summaries, reports, emails, analyses. All of it faster and cheaper than anything a human team could generate.

This looks like productivity. It is not. It is activity.

Productivity is output per unit of input: revenue per employee, units shipped, decisions made, problems solved. A meeting recap nobody reads is not productivity. A forty-slide deck that replaces a conversation is not productivity. A five-paragraph email conveying one sentence of actual information is not productivity. All three look like work. All three consume review time downstream. None of them move the number.

When AI lowers production cost to near zero, organizations do not produce the same amount of work faster. They produce more work. Volume expands to fill the available tool. And when volume expands faster than the organization’s capacity to evaluate it, the net effect on productivity is negative. Not because AI is bad at its job, but because activity got cheaper while evaluation did not.

Brynjolfsson’s mismeasurement explanation from 1993 has a modern twist. The metrics are probably fine. What they are measuring is mostly not productivity.

Usage Up, Confidence Down

The ManpowerGroup confidence drop has a mechanism. Models are trained on feedback signals that reward fluent, validating, comprehensive-sounding output. What you get at scale is surface competence with hollow substance, confidently delivered. Use any tool like that daily and trust erodes exactly as usage rises. You notice the fluent summary said nothing. You notice the analysis rephrased the input. You notice how often “it depends” and an elegant tradeoff table replaced a recommendation.

Dell’Acqua’s 2023 BCG study quantified the downstream cost.5 Consultants working with AI on tasks outside the model’s capability frontier performed worse than the control group, and did not recognize the degradation. Inside the frontier, they gained 40%. Outside it, they lost ground silently. The organizational-level version of that is the NBER finding: adoption climbing, output climbing, productivity flat, because some of the volume is additive and some of it is subtly erosive, and nobody is sure which.

Executive adoption does not show the same confidence decline. Leadership use is still climbing. That is the divergence to watch. The people producing AI output trust it less the more they produce. The people commissioning it are still scaling up.

The Silent Cost

The NBER executives use AI roughly 1.5 hours a week on average. Adoption is up. Output is up. Whether they understand what got produced is a separate question, and MIT’s Media Lab gave that question a 2025 answer.6

In an EEG study across writing sessions, participants who relied on an LLM from the start showed systematically reduced neural connectivity across prefrontal and occipito-parietal regions. Participants who did the work manually first and then switched to AI showed strong multi-band connectivity. The people who built understanding before automating retained it. The people who skipped straight to automation never built it.

That is the individual version of a problem that scales. Work gets produced. Understanding does not. The debt sits in the operators, not the output. Measured across 6,000 executives producing volume they do not fully read, on decisions they cannot fully reconstruct, the productivity flatline stops being mysterious.

Brynjolfsson’s fourth explanation, mismanagement, fits here exactly. The problem is not the technology. It is what the technology is doing to the humans running the organization.

The Paradox Is Transitory, But

Solow was right in 1987. He was also, eventually, wrong. IT productivity gains arrived in the late 1990s, roughly a decade after the paradox was named.

But the 1990s resolution was not uniform, and it did not hold. Acemoglu, Autor, Dorn, Hanson and Price revisited the question in 2014 with firm- and industry-level US manufacturing data7 and found a sharper story. IT-intensive sectors did show faster productivity growth, but output actually declined in those sectors relative to other manufacturing. Their summary: “productivity increases, when detectable, result from the even faster declines in employment.” Productivity per worker went up because employment went down faster than output did. That is not the productivity gain the original paradox asked about.

The lesson for 2026 is not “the paradox always resolves, so wait.” It is “the paradox can resolve in ways that look good in the aggregate and feel bad in the firm.” AI gains may show up as labor shedding before they show up as output growth. Some firms will genuinely capture the technology. Others will capture the appearance of it through headcount reduction while output stays flat or declines.

The NBER executives expect both: a 1.4% productivity gain alongside a 0.7% employment reduction over the next three years.3 Ratios worth watching.

Judgment as Infrastructure

The 1990s resolution came from a specific kind of organizational learning. Information gating. Not less technology. More discipline about what got through.

This time the gate is harder to build because the volume is higher, the content is more fluent, and the signal is harder to distinguish from the noise. An agonizingly detailed report printed on paper at least looked like work. A fluent, well-structured AI summary that confidently says nothing looks like insight until you are twenty minutes into reading it.

The executives who close the NBER gap fastest will not be the ones who deploy the most AI. They will be the ones who build the evaluation layer. Who treat editorial judgment as infrastructure, not overhead. Who measure what shipped, not what was generated. Who require compression before distribution and assign ownership to the filter, not just the output.

That shift is a management problem, not a technology problem. It always was.

Where This Sits

The data explains the frustration. Executives are using AI, watching volume climb, and seeing the productivity number refuse to move. That is not failure. That is the expected early-cycle behavior of a general-purpose technology before the organizational layer catches up.

The Solow paradox resolved. This one will too. It will not resolve on its own. It will not resolve faster with more output. It resolves when someone owns the gate.

Sources

  1. Solow, R.M. "We'd Better Watch Out." New York Times Book Review, July 12, 1987, p. 36. Review of S.S. Cohen & J. Zysman, Manufacturing Matters: The Myth of the Post-Industrial Economy. NYT does not host the piece (pre-digital archive); http PDF mirror is the widely-cited working copy. See also Brookings overview and Standup Economist citation research. ↑
  2. Brynjolfsson, E. "The Productivity Paradox of Information Technology." Communications of the ACM, Vol. 36, No. 12, December 1993, pp. 67–77. DOI: 10.1145/163298.163309. (ACM Digital Library currently gates automated access via Cloudflare; the 1996 expanded version of this argument is openly mirrored at MIT CCS.) Coined the term "productivity paradox," credited Solow's 1987 quip, and laid out the four-explanation framework still used to interpret IT-productivity gaps: mismeasurement, lag, redistribution, mismanagement. ↑
  3. Yotzov, I., Barrero, J.M., Bloom, N., Bunn, P., Davis, S.J., Foster, K.M., Jalca, A., Meyer, B.H., Mizen, P., Navarrete, M.A., Smietanka, P., Thwaites, G., & Wang, B.Z. "Firm Data on AI." NBER Working Paper 34836, February 2026 (revised March 2026). NBER. Survey of nearly 6,000 senior executives across US, UK, Germany, and Australia. Nine in ten report no measurable AI impact on productivity or employment over the past three years; forward expectations of 1.4% productivity gain, 0.8% output gain, 0.7% employment reduction over the next three years. Surfaced in popular coverage by Fortune, which connected the findings back to Solow's 1987 paradox. ↑
  4. ManpowerGroup. Global Talent Barometer 2026. January 20, 2026. Survey of workers across 19 countries. Full report PDF. Regular AI usage up 13 percentage points to 45% of workers; confidence in using technology fell 18% (first decline in three years); 43% expect automation to replace their job within two years (+5 pts from 2025). ↑
  5. Dell'Acqua, F., et al. "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality." Harvard Business School Working Paper 24-013, 2023. HBS. 758 knowledge workers. AI boosted performance 40%+ on tasks inside the capability frontier; degraded performance on tasks outside it, with workers failing to recognize the degradation. ↑
  6. Kosmyna, N., Hauptmann, E., Yuan, Y.T., Situ, J., Liao, X-H., Beresnitzky, A.V., Braunstein, I., & Maes, P. "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task." MIT Media Lab, 2025. arXiv:2506.08872. EEG study across writing sessions. Participants who relied on an LLM from the start showed systematically reduced neural connectivity across prefrontal and occipito-parietal regions; those who built manual proficiency first and then switched to AI retained strong multi-band connectivity. ↑
  7. Acemoglu, D., Autor, D., Dorn, D., Hanson, G.H., & Price, B. "Return of the Solow Paradox? IT, Productivity, and Employment in US Manufacturing." NBER Working Paper 19837, January 2014. NBER. Firm- and industry-level US manufacturing data. IT-intensive sectors showed faster productivity growth, but output declined in those sectors relative to other manufacturing. "Productivity increases, when detectable, result from the even faster declines in employment." ↑