Last August, a team led by a Stanford economist Erik Brynjolfsson published a An in-depth look at the impact of artificial intelligence on businessPowered by a “large-scale, high-frequency administrative dataset from ADP,” the largest payroll software provider in the United States.
The findings were striking: a significant relative decline in employment of workers ages 22 to 25 in occupations most exposed to AI since the widespread adoption of productive AI, even after controlling for other economic shocks. Critics immediately pushed back. Google economists say it’s interest ratesOthers blamed the tech sector’s overhiring, disruptions in remote work and pandemic noise. Earlier this month, Apollo Torsten Slok of Global Management went on to argue that entry-level hiring woes are a feature of low hiring and a low-fire job market, asking: “Where is the AI job crisis?“
Brynjolfsson not only continued to update the data, but also partnered with ADP Research, the economics arm of the private payroll data provider that serves nearly one in six American workers. The effect did not disappear upon closer inspection
“Whatever it is,” said Brynjolfsson. Luck“it’s not going anywhere.”
New figures have arrived Canaries Control PanelThe centerpiece of Brynjolfsson’s expanded partnership between the Stanford Digital Economy Lab and ADP Research builds on that explosive paper last summer. It draws on information on 4.6 million workers across more than 730 occupations, and the Digital Economy Lab considers it the highest-profile dashboard among a variety of freely accessible, constantly updated indicators. Artificial intelligence economic indicators It continues to track the effects of artificial intelligence on the labor market in near real time.
For Brynjolfsson, this is a partial response to every critic who said his original finding was a mistake. “We are flying blind towards one of the most important periods in world history,” Brynjolfsson said at the launch of the platform. “We need timely and reliable evidence to understand where AI creates value and where it disrupts business.”
The dashboard processes payroll data covering roughly one in six American workers. What emerges when broken down by age and level of exposure to AI is a widening fault line.
The bulk picture is deceiving
Among all workers, the numbers remain silent. The occupations most exposed to artificial intelligence contracted by only 0.2% on an annual basis as of April 2026, while the occupations least exposed to it increased by 0.1%. Since ChatGPT launches in late 2022, annual employment growth in occupations exposed to AI is actually augmented Among the least exposed, the rate increased by 1.1% compared to 2%. At headline level, the sky did not fall.
The story changes if you break down the data by career stage.
For workers ages 22 to 25, employment in occupations with high exposure to AI is currently shrinking at 3.8% per year, with early career decline sharpening after the first year; The 2.8% decline from April 2024 has reached an annual decline of over 4% since then. The average month-on-month decline averages around -0.3%, but Brynjolfsson notes that the trend is noisy compared to the annual slowdown.
The least exposed jobs in the same age group are growing at a rate of 2% per year. Mid-career workers ages 31 to 34 are also contracting, down 1.7% from the previous year. In contrast, workers between the ages of 35 and 40 are growing by 2%. Technology does not eliminate work in every field. It eliminates the rate of increase and does so with increasing precision as data accumulates.
Why are young workers bearing the brunt?
The mechanism is not mysterious. AI absorbs tasks before it absorbs jobs, and the first tasks it reaches are those that do not require many years of experience: retrieving, summarizing, planning, formatting, mechanically combining information. These are tasks disproportionately given to people at the beginning of their careers. Veteran workers still have hard-to-codify job-specific skills that buffer against displacement. Young workers have not yet done this.
ADP chief economist Nela Richardson — Brynjolfsson’s partner in the research — argued that the distinction between automation and augmentation is the key variable. Professions where artificial intelligence augments human labor show more permanent employment growth; There is a shrinkage in places where AI directly automates tasks. Early career workers, who concentrate on the most automatable layer of any profession, fall squarely into the second category.
“Overall, AI’s impact on employment remains modest,” Richardson said. June 16 blog post in the first set of dashboard data. But when AI’s impact is measured by career stage, “dramatic differences emerge,” he continued.
Given all the variables involved and the wide range of uncertainties, much of the discussion around AI and jobs is based on “predictions,” according to Richardson. Its result, like itself It was said recently Luckmore “subtle” is the fact that AI is disrupting things from the bottom up rather than from the top down. “We are seeing increased employment in occupations and career stages where AI leverages human capabilities and potential,” he wrote.
Brynjolfsson has now stress-tested the finding against all the major counter-arguments. The interest rate hypothesis points in the wrong direction; Occupations that are most sensitive to interest rates, such as construction, have the lowest AI exposure. It wiped out the entire tech sector. Remote working isolated its effects. The pattern has always held. “If you eliminate the entire tech industry, or you eliminate all technology-related occupations, or you slice it in different ways, you still get that effect,” he said.
The original document covered data through August 2025. The new dashboard extends this to April 2026; that is, almost four years of post-ChatGPT labor market data. The effect was not mean reverting. It has steadily increased by about half a point per month, month after month.
Friendly matches at the top of the field
MIT economist and Nobel laureate Daron Acemoglu has become the most prominent voice of AI skepticism in the field, and the two have been publicly sparring for months. Acemoglu’s models produce much lower productivity estimates than Brynjolfsson’s; This gap frustrates Brynjolfsson, although he has deep respect for his former MIT colleague.
Brynjolfsson said he actually “went back and forth” with Acemoglu on the morning of our interview. “We’re trying to find common ground.”
There is some. Both agree that AI should be used to complement human workers rather than replace them, and both are “trying to beat that drum,” as Brynjolfsson puts it. But the gap between them in terms of productivity is still wide. Acemoglu argued that artificial intelligence can still provide significant gains if used correctly; Brynjolfsson finds this view somewhat contradictory. “I don’t understand how you have such low productivity numbers,” he said. “I’m telling him this. But time will tell. Very soon we will see who is right.”
Acemoglu did not soften public skepticism. he said recently Luck Much of the AI productivity discourse goes something like this “stupid“- speculative to the point of fiction,” he explained, “not stupid per se.”
What’s notable is that the debate is no longer about whether AI is transformative. It’s about size and timeline. This is a change in itself. A year ago, Brynjolfsson was still trying to persuade mainstream economists to take the question seriously. Now the debate has moved into his domain, and the data he has created is designed to solve this problem.
bets
Brynjolfsson is careful about scale. He positions himself between Silicon Valley doomsayers and mainstream economists who think AI adds a tiny percent to productivity. But the middle space it occupies is still historically large. The comparison for this degradation is not the internet. It’s not even globalization.
Referring to the comparison he made in 2016 New York Times best seller Second Machine AgeThis is similar to the Industrial Revolution, the last time humanity created machines that completely changed work and productivity. “It’s automated, it’s strengthened our muscles, and now we’re doing it for our minds. How can it not be this big or bigger?” he said. “I think it will be bigger and 10 times faster.”
But early-career workers in AI-exposed professions are currently bearing a cost that has yet to appear in the headlines. The Canaries Dashboard takes its name from this logic: The canaries in the coal mines did not stop the danger. They just told you the clock is working.
Brynjolfsson shared that he made a friendly bet with himself Northwest economist Bob Gordon — A 10-year bet on longbets.com that productivity will increase significantly by the end of the decade.
“I’m already ahead,” he said. “And I always thought it was restored because of my J-curve theory. So barring war or disaster – the AI part should be positive.”
This story first appeared on: Fortune.com
