Last August, a team led by Stanford economist Erik Brynjolfsson published a deep look at the impact of AI on jobs, boosted by a “large-scale, high-frequency administrative dataset from ADP,” the largest payroll software provider in the United States.

The findings were stark: a significant relative decline in employment for workers ages 22 to 25 in the most AI-exposed occupations since the widespread adoption of generative AI — even after controlling for other economic shocks. Critics pushed back immediately. Google economists said it was interest rates, while others blamed tech-sector overhiring, remote work distortions, pandemic noise. Earlier this month, Apollo Global Management’s Torsten Slok continued to argue that entry-level hiring woes are a feature of the low-hire, low-fire job market, asking “where is the AI jobs crisis?”

Not only did Brynjolfsson keep updating the data, but he partnered with ADP Research, the economics arm of the private payroll data provider, which serves roughly one in six American workers. The effect hasn’t faded after a closer look

“Whatever it is,” Brynjolfsson told Fortune, “it’s not going away.”

The new numbers come from the Canaries Dashboard, the centerpiece of an expanded partnership between Brynjolfsson’s Stanford Digital Economy Lab and ADP Research, building on that bombshell paper last summer. It draws on information about 4.6 million workers across more than 730 occupations, and the Digital Economy Lab considers this to be its highest-profile dashboard among the several freely accessible, continuously updated AI economic indicators that it maintains to track AI’s effects on the labor market in near-real time.

For Brynjolfsson, it is a partial answer to every critic who said his original finding was a blip. “We are flying blind into one of the most consequential periods in world history,” Brynjolfsson said at the platform’s launch. “We need timely, trusted evidence to understand where AI is creating value and where it is disrupting work.”

The dashboard processes payroll data covering roughly one in six American workers. What it shows, broken down by age and AI exposure level, is a widening fault line.

The aggregate picture is deceptive

Across all workers, the numbers remain muted. The most AI-exposed occupations contracted just 0.2% year over year as of April 2026, compared to 0.1% growth for the least-exposed roles. Since ChatGPT’s introduction in late 2022, annual employment growth across AI-exposed occupations has actually increased by 1.1%, compared to 2% for the least-exposed. At the headline level, the sky hasn’t fallen.

Cut the data by career stage, and the story changes.

For workers ages 22 to 25, employment in highly AI-exposed occupations is now shrinking at 3.8% per year and the early-career decline sharpened after year one — 2.8% decrease to April 2024, growing to a more than 4% decline per year since. The average decline on a month-to-month basis averages about −0.3% but Brynjolfsson notes that trend is noisy, compared to the year-over-year deceleration.

The least-exposed jobs in that same age group are growing at 2% annually. Mid-career workers ages 31 to 34 are also contracting, down 1.7% year-over-year. Workers ages 35 to 40, by contrast, are growing at 2%. The technology isn’t eliminating work across the board. It’s eliminating the on-ramp — and it’s doing so with increasing precision as the data accumulates.

Why young workers bear the brunt

The mechanism isn’t mysterious. AI absorbs tasks before it absorbs jobs, and the tasks it reaches first are the ones that don’t require years of experience: retrieving, summarizing, scheduling, formatting, the mechanical assembly of information. These are disproportionately the tasks handed to people at the beginning of their careers. Senior workers have accumulated the hard-to-codify, job-specific skills that still buffer against displacement. Junior workers haven’t yet.

ADP chief economist Nela Richardson — Brynjolfsson’s partner on the research — has argued the distinction between automation and augmentation is the key variable. Occupations where AI augments human work show more enduring employment growth; those where AI automates tasks outright show contraction. Early-career workers, concentrated in the most automatable layer of any occupation, sit squarely in the second category.

“In the aggregate, AI’s impact on jobs remains modest,” Richardson stressed in a June 16 blog post on the first batch of dashboard data. But when AI’s impact is measured by career stage, she continued, “dramatic differences emerge.”

To Richardson, much of the debate around AI and jobs comes down to “guesswork,” given all of the variables involved and the huge range of uncertainties. Her conclusion, as she recently told Fortune, is that the reality is more “nuanced,” with AI disrupting tasks from the bottom up, not jobs from the top down. “In occupations and career stages where AI amplifies human abilities and potential,” she wrote, “we see employment growth.”

Brynjolfsson has now stress-tested the finding against every major counter-argument. The interest rate hypothesis points the wrong direction — the most rate-sensitive occupations, like construction, have the lowest AI exposure. He removed the entire tech sector. He isolated remote-work effects. The pattern held every time. “If you take out the entire tech industry, or take out all tech-related occupations, or you slice it different ways, you still get this effect,” he said.

The original paper covered data through August 2025. The new dashboard extends that to April 2026 — nearly four years of post-ChatGPT labor market data. The effect hasn’t mean-reverted. It’s grown by roughly half a percentage point per month, consistently, month after month.

Friendly fights at the top of the field

Daron Acemoglu, the MIT economist and Nobel laureate, has become the most prominent voice of AI skepticism within the field — and the two have been publicly sparring for months. Acemoglu’s models produce far lower productivity estimates than Brynjolfsson’s, a gap that frustrates Brynjolfsson even as he maintains deep respect for his former MIT colleague.

Brynjolfsson said that in fact he had just been “going back and forth” with Acemoglu on the morning of our interview. “We’re trying to find some common ground.”

There is some. Both agree that AI should be deployed to complement human workers rather than replace them — and both have been “trying to beat that drum,” as Brynjolfsson put it. But on the productivity question, the distance between them remains wide. Acemoglu has argued that if AI is used correctly, it could still deliver significant gains — a position Brynjolfsson finds somewhat contradictory. “I don’t get how he has such low productivity numbers,” he said. “I tell him that. But time will tell. Pretty soon we’re going to see who’s right.”

Acemoglu, for his part, hasn’t softened his public skepticism. He recently told Fortune that much of the AI productivity discourse is “brainless” — speculative to the point of fiction, he clarified, not stupid per se.

What’s notable is that the argument is no longer about whether AI is transformative. It’s about the magnitude and the timeline. That, in itself, is a shift. A year ago, Brynjolfsson was still convincing mainstream economists to take the question seriously. Now the debate has moved to his turf, and the data he’s building is designed to settle it.

The stakes

Brynjolfsson is careful about scale. He positions himself between Silicon Valley catastrophists and mainstream economists who see AI adding fractions of a percent to productivity. But the middle ground he occupies is still historically large. His comparison for this disruption isn’t the internet. It isn’t even globalization.

Alluding to the comparison he made in his 2016 New York Times bestseller The Second Machine Age, this is like the Industrial Revolution — the last time humanity built machines that changed work and productivity completely. “That one automated, augmented our muscles, and now we’re doing it for our minds. How can that not be as big or bigger?” he said. “I think it’s going to be bigger and 10 times faster.”

But early-career workers in AI-exposed occupations are, right now, bearing a cost that doesn’t yet appear in the headline numbers. The Canaries Dashboard takes its name from that logic: canaries in coal mines didn’t stop the danger. They just told you the clock was running.

Brynjolfsson shared that he has a friendly wager with Northwestern economist Bob Gordon — a 10-year bet on longbets.com that productivity will be significantly higher by the end of the decade.

“I’m already ahead,” he said. “And I always figured it was backloaded because of my J-curve theory. So barring war or catastrophe — the AI part should be positive.”

This story was originally featured on Fortune.com