Michael Strain Maps the AI Boom Onto an Old Lag

Michael Strain told a Georgetown audience that generative AI has not yet lifted U.S. aggregate productivity, even after several years of intense adoption. The McCourt School professor of practice, who also directs economic policy studies at the American Enterprise Institute, said some firm studies look promising while the national accounts stay quiet.

The gap matches the one Robert Solow named in 1987, when computers were already on desks and still hard to find in output-per-hour figures. Strain is not calling the tools fake. He is saying the country is in the lag, and that Silicon Valley’s civilizational claims are running ahead of the Bureau of Labor Statistics.

The Productivity Accounts Have Not Caught the Boom

Strain put the measurement point in plain language at the campus event. Investors and executives expected a burst in output per hour. He said economists largely agree they have not seen it in the totals, even if a handful of companies look better in small studies.

When will it translate into productivity gains. First of all, it hasn’t yet, and I think there’s pretty broad agreement among economists on that point. There are some studies of specific firms or small groups of firms that do show that, but, when you look at the aggregate productivity data, it’s just not there.

Michael R. Strain, professor of practice, Georgetown McCourt School event

The bureau’s own prints do not show zero. They also do not show a 1990s-style computer boom. In the second quarter of 2026, nonfarm productivity rose 1.4 percent at an annual rate, as output rose 1.7 percent and hours worked rose 0.3 percent. For all of 2025, labor productivity in that sector rose 2.1 percent. Total factor productivity, which nets out capital and other inputs as well as labor, rose only 0.8 percent in 2025, with output up 2.6 percent and combined inputs up 1.7 percent.

U.S. PRODUCTIVITY IN THE AI YEARS

A 2.1 percent year is faster than the slog after the global financial crisis and slower than the peak of the 1990s digital wave Strain has used as a benchmark in his own essays. The 0.8 percent total-factor reading is the weaker signal, and it is the one closest to his claim that the boom is not yet in the accounts. The next productivity print is scheduled for November 5, 2026.

Adoption and output are still different stories. In an August 2025 essay, Strain cited OpenAI figures that 28 percent of employed U.S. adults used ChatGPT at work, up from 8 percent in 2023. A St. Louis Fed tracker built by Alexander Bick, Adam Blandin, and David Deming found that the share of workers using generative AI rose from 28.2 percent in 2024 to 39.2 percent by mid-2026, while the share of work hours saved rose only from 1.6 percent to 2.2 percent. Faster drafts can fill the extra minutes with more drafts. National output does not have to move.

The Dynamo Lag That Lasted Forty Years

Strain asked the audience to put chatbots next to older leaps, not next to a product launch. He named the domestication of plants and animals, which let people stay in one place and rebuild culture around settled life. He named the printing press, without which he said it is hard to imagine the American Revolution. Then he scaled the claim back.

“I think it is going to be a big deal, but maybe not as big of a deal as some folks in Silicon Valley say,” he said. He added that it is too early to know whether the technology is overhyped. “I think nobody really knows the answer to that question.”

That historical habit is the part of the talk that travels. Solow’s 1987 line still does the work: you can see the computer age everywhere but in the productivity statistics. Economic historians have described a similar wait after commercial electric power, when motors spread slowly and the big factory gains arrived only after plants were rebuilt around the new drive, well into the 1920s. Office computing followed the same shape. Machines showed up first. Measured output per hour showed up later, once firms changed how they filed, billed, and scheduled work.

HOW LONG PAST LEAPS TOOK TO SHOW UP

1. Late 19th century: Commercial electric power arrives; factories still run on older drive systems for years.
2. 1920s: U.S. manufacturing productivity jumps after plants are redesigned around the electric motor.
3. 1987: Robert Solow writes that computers are visible everywhere except the productivity statistics.
4. Late 1990s: Measured U.S. productivity finally rises with information technology, after a long reorganization.
5. November 2022: ChatGPT launches and consumer adoption races ahead of workplace redesign.
6. 2025: U.S. labor productivity rises 2.1 percent; total factor productivity rises 0.8 percent.
7. Second quarter of 2026: Nonfarm labor productivity rises 1.4 percent at an annual rate.

Strain has used that same long view on jobs. He has written that about 60 percent of the jobs workers held in 2018 did not exist in 1940, because new tools create tasks no one could have listed in advance. The optimism and the caution come from the same file. New work appears late. The scare arrives early.

Goldman Now Puts the Job Hit at 16,000 a Month

The campus recap still leaned on a 2023 Goldman Sachs estimate that AI could affect or displace more than 300 million jobs worldwide. That global exposure figure is the one that filled headlines. It is not the figure Goldman’s own labor work is now putting on the U.S. payroll.

On April 24, 2026, Goldman Sachs Research economist Elsie Peng wrote that AI had produced a 16,000-job monthly payroll drag over the prior year and had raised the unemployment rate by 0.1 percentage point. In occupations where the tools help people rather than replace them, the same team found a gain of about 9,000 jobs a month. Peng noted that even the net drag likely overstates the hit, because the estimates do not fully count hiring for data-center construction or extra demand from cheaper output.

Those are small numbers next to a 300 million global headline. They sit closer to Strain’s line that AI has not meaningfully moved the unemployment rate. They also match the pattern in the productivity file: substitution shows up in pockets, help for workers shows up in other pockets, and the national total barely twitches. Peng’s note added that the weaker job creation has been falling largely on younger, less-experienced workers, which is a distribution story, not an end-of-work story.

On September 10, 2026, two days before the student recap of his remarks, the Commission on AI and the Future of the American Workforce named Strain among 20 commissioners. The group, co-hosted by AEI and the Urban Institute and co-chaired by Gina Raimondo and Paul Ryan, launched on June 11, 2026. Strain is arguing in public that the aggregates are calm while sitting on a body built to write policy playbooks if they are not.

Why Young Adults Still Expect Fewer Jobs

Young adults now sound more like the 300 million headline than like the 16,000-a-month print. A Pew Research Center survey of 3,488 U.S. adults, fielded June 22-28, 2026, found that 73 percent of adults under 30 think AI will lead to fewer jobs in the United States over the next 20 years, up from 61 percent in 2024. Among all adults the share is 71 percent, up from 64 percent. Only 5 percent think AI will mean more jobs.

Worry about daily use has climbed with the job fear. Fifty-two percent of adults now say they are more concerned than excited about AI in daily life, up from 37 percent in 2021. For the first time in that series, a majority of adults under 30, 55 percent, say the same. Excitement among that age group has faded to 11 percent.

The split is the Solow pattern in public opinion. People can see the tools. They cannot see a national hiring boom, and they cannot see a crash either, so the story they tell is the one that arrived first. Strain told the audience his own worry has gone the other way. “My level of concern is shrinking, not growing,” he said. “I really approached this question with a historical mindset, and concerns about the end of human work have been with us for a long time.”

Georgetown Built Gemini Into Campus Anyway

The university hosting the skeptic has spent the past year putting the tools in students’ hands. That is not a contradiction in Strain’s terms. He has said the technology will be a big deal. It is a contradiction in pacing. The campus is building for a general-purpose shift while one of its own economists is telling students the printing press still has a stronger claim on history.

WHAT GEORGETOWN HAS ROLLED OUT

• Enterprise Gemini: After a 2025 faculty and staff pilot with a 100 percent recommendation rate, the university issued Google’s assistant to faculty and staff in early March 2026, with students following, under an agreement that campus chats are not used to train the models.
• AI Fellows: The President’s Office program pairs faculty with undergraduate AI Fellows on a course, paying $2,500 for about 130 hours; applications are due September 28, 2026, with selections on October 5.
• New teaching stack: The College of Arts and Sciences is adding an undergraduate certificate in artificial intelligence, the McDonough School of Business is adding an AI core requirement for the MBA, and the School of Continuing Studies already launched a master’s in artificial intelligence management in 2025.

A planned university-wide framework sits behind those pieces. Strain’s talk does not tell the university to cancel them. It tells students not to confuse a software rollout with a measured jump in national output per hour. Universities can teach people to use a tool for years before the tool rewrites the productivity accounts. That is what happened with personal computers on campus in the 1980s.

Lower Prices Beat a 40 Percent Job Collapse

The part of the talk that is easy to miss, if the headline is “marginal impact,” is the price channel. Strain said firms would not use the tools unless they cut production costs, and that cheaper production puts downward pressure on market prices. In that frame, the first national effect worth watching is not a spike in the unemployment rate. It is whether a cost cut ever shows up on a receipt.

“AI will lower production costs, or it wouldn’t be used,” he said. He then walked through the imagined disaster that still dominates student questions, a drop from widespread prime-age work to a country where most adults do not have jobs.

If we go from a society where 85 percent or 90 percent of prime-age adults had a job to a society where 40 percent of prime individuals had a job, then our choice is going to either be that people are starving to the death on the streets or we’re going to have to really ramp up our safety net spending.

Michael R. Strain, professor of practice, Georgetown McCourt School event

He does not think that is the path. “I don’t think it’s going to take all our jobs. I don’t think it’s going to kill all people on Earth,” he said. “I don’t think any of those things are going to happen, but I think it will be a big deal.”

A big deal that is not the printing press, and not a 40 percent employment collapse, is a narrower claim than either camp likes. It leaves Strain arguing that the safety-net fight can wait for the accounts to move, while young adults, reading empty job boards in white-collar entry roles, already vote with Pew’s 73 percent. The historical file says both can be true for a long time. Computers were ordinary long before they were visible in output per hour, and the panic about them did not wait for Solow to be proven wrong.