AI Productivity Gains Surface in Economic Data

For decades, economists have warned about the productivity paradox: transformative technologies that don't show up in the numbers. Robert Solow's famous 1987 observation that "you can see the computer age everywhere but in the productivity statistics" became the template. New technology arrives, everyone insists it changes everything, and the data stubbornly refuses to budge.

This time might be different.

Stanford economist Erik Brynjolfsson argues in a recent Financial Times piece that AI's productivity impact is now visible in macro data. The evidence comes from an unlikely source: the Bureau of Labor Statistics revising 2025 job figures downward by about 400,000 positions. What looked like 584,000 new jobs is now believed to be 181,000.

Meanwhile, GDP growth remained robust—provisional Q4 figures show 3.7 percent growth, with the Atlanta Fed's forecast even higher at 5.4 percent. The math is straightforward: productivity equals GDP divided by workers. Fewer workers producing the same output means productivity went up. Brynjolfsson estimates 2025 productivity growth at 2.7 percent, nearly double the past decade's average.

The timing matters. Technology transitions typically take years to register in productivity statistics. We saw this with computers, with the internet, with mobile computing. The lag exists because companies don't just install new technology—they restructure around it. They experiment, fail, learn, reorganize. Brynjolfsson has written about this before, calling it the productivity J-curve. Measured productivity initially drops as resources flow into intangible investments: training, process redesign, organizational learning. Only later do the gains materialize.

"General purpose technologies from the steam engine to the computer do not deliver immediate gain," Brynjolfsson wrote. "During this phase, measured productivity is suppressed as resources are diverted to investments. The updated 2025 US data suggests we are now transitioning out of this investment phase into a harvest phase."

If accurate, this represents a significant break from historical patterns. But the interpretation isn't unanimous.

Economist Guy Berger examined the revised statistics and found the job losses concentrated in government positions, mining, logging, transportation, and manufacturing. Not exactly the AI-exposed white collar sectors Brynjolfsson's earlier research highlighted. "I'd be careful about drawing this inference from that data point," Berger posted. "May turn out to be true, but it's very thin evidence."

The white collar picture remains murky. The Kobeissi Letter reports that professional and business services now have just 1.6 job openings per 100 employees—the lowest in eleven years. Total openings in the sector have dropped 1.4 million since March 2022. The hiring rate sits at 4.2 percent, matching 2008 financial crisis levels.

But correlation isn't causation. Interest rates rose sharply in 2022. Consumer behavior shifted. Multiple factors could explain hiring slowdowns. Brynjolfsson and colleagues addressed this in a follow-up to their "Canaries in the Coalmine" paper, concluding that while interest rates affect overall employment, "existing evidence does not suggest they are a good explanation for the disproportionate decline in entry-level hiring in AI exposed occupations."

They also conceded something important: when you add comprehensive controls to the data, the employment decline in AI-exposed occupations only becomes statistically significant in 2024. Earlier declines likely stem from multiple factors, not just AI.

This is where the conversation moves from numbers to people. Andrew Yang has been amplifying concerns about white collar displacement, recently writing about a family member using AI to build a website in minutes—work that previously required days from a designer. "How many roles essentially consist of processing information and then presenting it to someone to make a decision?" Yang asked. "Now not only the process and report will be automated, but perhaps the decision as well."

Politicians are taking notice. Republican Jay Obernolte, who holds a master's degree in AI and spent three decades in tech, told reporters: "There will be job displacement. We need to reskill the workers that are in industries with that job displacement and equip them with the skills that they need to succeed in other industries. We are going to need a social safety net because there will be people that fall through the cracks."

Senator Elizabeth Warren expressed similar concerns from a different political angle: "I'm deeply concerned about AI and what it's going to mean when people go out one day for lunch and come back and their jobs aren't there anymore."

The bipartisan worry is notable. So is the uncertainty. Nobody can definitively map cause and effect here. The productivity signal Brynjolfsson identifies could be real. It could also be statistical noise, a temporary artifact of how we count workers and output during a period of unusual economic turbulence.

What's clear is that we're still operating with incomplete information. The Yale-Stanford "Canaries in the Coalmine" study showed hiring slowdowns but couldn't prove causation. Brynjolfsson's productivity argument relies on a single data revision. Other research shows AI users spending more time on tasks, not less—hardly the displacement story.

The challenge is that by the time we have definitive data, the transition will already be underway. We've been here before with every major technology shift: arguing about what's happening while it happens, trying to measure what can't yet be fully measured, debating whether this time is truly different.

History suggests the pattern holds: technology creates disruption, some jobs disappear, new ones emerge, and the adjustment period causes real pain for people caught in the middle. The question isn't whether AI will follow this pattern—it likely will—but whether the speed and scale differ enough to demand different policy responses.

Brynjolfsson believes we're transitioning "from an era of AI experimentation to one of structural utility." If the productivity data confirms that shift, we'll need more than anecdotes about family members building websites. We'll need systematic evidence about which workers, which industries, which regions face the sharpest adjustments. And we'll need that evidence before the displacement happens, not after.

The data is starting to arrive. Whether we're reading it correctly remains an open question.

—Bob Reynolds