Solow's paradox, AI edition

The line that came back

In 1987 Robert Solow wrote that you could see the computer age everywhere but in the productivity statistics. On 17 February Fortune ran the line again, this time about AI, on the back of a new National Bureau of Economic Research working paper. The paper, by Ivan Yotzov, Nicholas Bloom, Steven Davis and colleagues, surveyed close to 6,000 executives at firms in the United States, the United Kingdom, Germany and Australia. Around 90 percent said AI had no effect on their firm's productivity or employment over the past three years.

The same survey found that 69 percent of the firms use AI, that adoption is higher at younger and more productive firms, and that executives who use it average about 1.5 hours a week. Asked about the next three years, they expect AI to raise productivity by 1.4 percent, raise output by 0.8 percent and cut employment by 0.7 percent. Those are small numbers from people who, in the same breath, say they are spending on the technology. Fortune quotes a Financial Times count of 374 S&P 500 companies mentioning AI on earnings calls in the year to September 2025, and corporate AI investment above 250 billion dollars in 2024.

Three kinds of evidence that disagree

The reason the paradox feels sharp is that the studies people cite are measuring different things. The first kind is the randomised task experiment. Fortune cites 2023 MIT work reporting that AI could raise a worker's performance by close to 40 percent, and a Stanford Institute for Economic Policy Research study of browsing data from 200,000 households that found efficiency gains on online tasks of between 76 and 176 percent. These studies give a worker a bounded task, a control group, and a stopwatch. They are clean and they measure the ceiling.

The second kind is the self-report. The NBER survey is one. ManpowerGroup's 2026 Global Talent Barometer, covering nearly 14,000 workers in 19 countries, found regular AI use up 13 percent in 2025 while workers' confidence fell 18 percent. A Boston Consulting Group survey of 1,488 full-time US workers found productivity rising with up to three AI tools and falling when people used four or more. Self-reports tell you what people believe happened, aggregated over everything they did, with no control group.

The third kind is the aggregate. The Federal Reserve Bank of St. Louis reports about 1.9 percent of excess cumulative productivity growth since late 2022 in its State of Generative AI Adoption report, and Fortune notes US productivity rose 2.7 percent in 2025. Daron Acemoglu's 2024 estimate, the one most often cited by sceptics, puts the total effect at roughly a 0.5 percent productivity gain over the coming decade. The aggregate is what Solow was talking about, and it is the number that lags.

Why the numbers do not add up to each other

A 40 percent gain on a task is consistent with a zero at the firm if the task is a small share of the job, if the time saved is absorbed rather than redeployed, or if the output of the task was never the constraint on the firm's output. Most knowledge work has a bottleneck somewhere other than the drafting of text. Speeding up a step that was not the bottleneck moves the aggregate very little, and an executive who is asked whether the firm became more productive will answer, honestly, no.

There is also a selection effect in who gets asked. Randomised experiments recruit workers whose tasks can be given to a model and timed. Surveys ask executives about the whole firm, including the parts nothing has touched. The 1.5 hours a week figure matters here. An executive who uses a tool for ninety minutes a week and sees a small fraction of the workforce use it at all is describing the truth from where they sit.

And the 1987 version of the paradox eventually resolved. Productivity growth in the United States fell from 2.9 percent a year over 1948 to 1973 to about 1.1 percent afterward, and the computers Solow could not find in the statistics did eventually show up in them. Whether that pattern repeats is an open question, and the survey's own respondents seem to think the gains are ahead of them rather than behind.

What the 90 percent figure can and cannot tell you

It cannot tell you that the task-level gains are fake. A firm-level self-report has no power to detect a 5 percent improvement in one department, and nine in ten executives saying no effect is exactly what you would expect if effects are real, small, and concentrated. It also cannot tell you the gains are coming, because expectation surveys have a long record of being wrong in both directions.

What it can tell you is that the adoption story and the measurement story are out of step. Firms are paying for a technology whose payoff they cannot yet see in their own accounts, which is the situation Solow described. The honest reading is that we are early in the reorganisation phase, and the honest caveat is that this is also what people said in every year of the previous paradox until it ended.

What would settle it

Payroll and output data at the establishment level, matched to adoption dates, would do more than another survey. The NBER paper is a step in that direction because it asks the same executives about the past and the future and can be rerun. What we would want next is a panel: the same firms, a year from now, asked the same question, so the 90 percent becomes a trajectory rather than a snapshot.

Until then we read the three kinds of evidence as three different instruments pointed at the same thing from different distances. The task experiments tell you what is possible. The self-reports tell you what has been noticed. The aggregate tells you what has been paid for. Right now those three are far apart, and the interesting research question is how long it takes for them to converge and in which direction.