Two credible, current pictures of what AI is doing to entry-level jobs flatly contradict each other. Stanford's Digital Economy Lab, tracking millions of US payroll records in an ongoing project called Canaries in the Coal Mine, finds that employment for workers ages 22 to 25 in the most AI-exposed occupations is now 19% below where it would be had it kept pace with less-exposed peers -- and the gap has kept widening every time researchers have checked. A separate, larger survey of nearly 6,000 corporate executives across four countries found close to the opposite headline: more than 90% report AI has had no effect on their own firm's employment at all.
The Stanford numbers come from economists Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, reading administrative payroll data from ADP covering roughly 4.6 million workers across 730 occupations -- about one in six American workers -- through June 2026, in a paper most recently revised August 12. The gap they're tracking has grown at every checkpoint: 13% with data through mid-2025, 16% by an October 2025 update, and 19% in the current revision. "Whatever it is, it's not going away," Brynjolfsson told Fortune in June. "We are flying blind into one of the most consequential periods in world history." He has said the pattern holds even after stripping out the tech industry entirely.
The Stanford employment gap, at each measurement
The counter-evidence comes from a survey fielded between November 2025 and January 2026 by researchers at the Federal Reserve Bank of Atlanta, the Bank of England, the Deutsche Bundesbank and Macquarie University, who put identical questions to nearly 6,000 CEOs, CFOs and senior finance managers across the US, UK, Germany and Australia. Per NBER's own digest of the results, more than 90% of executives reported no effect of AI use on employment over the past three years, and 89% reported no impact on labor productivity. A separate, smaller NBER working paper surveying roughly 750 US executives, run out of the Atlanta and Richmond Fed banks, reached a similar conclusion: "little evidence of near-term aggregate employment declines due to AI."
The two findings are not actually the contradiction they look like once the resolution each one is measuring at gets accounted for. Stanford's data can isolate exactly what happened to hiring for one narrow age band in one class of occupation, because it reads it directly off payroll records. The executive survey asks a CFO to self-report a single number for their ENTIRE workforce's employment trend -- a hiring slowdown concentrated in fresh graduates, answered by someone tracking headcount in aggregate, could easily net out to "no effect," especially if senior hiring at the same firm is flat or growing. (Neither survey is wrong about what it asked. The mismatch is that a firm-wide aggregate and an occupation-and-age-specific slice answer genuinely different questions, and only one of them is built to see a narrow effect at all.) Stanford's own explanation for why young workers are hit hardest points the same direction: AI automates specific, codifiable tasks -- summarizing, scheduling, formatting, retrieving information -- before it threatens whole occupations, and those tasks are concentrated in early-career roles rather than senior ones.
For the workers actually living inside that 19% gap, the distinction between the two studies is close to academic. A 23-year-old who can't get hired into a role AI has made cheaper to automate doesn't experience an aggregate that nets to zero -- they experience a market that quietly stopped needing them at the rate it used to. The asymmetry cuts one way: senior employees in the same AI-exposed occupations aren't seeing anything like the same contraction, because the tasks that make up their jobs are harder to codify into a prompt. Whatever the executive survey's respondents believe about their own hiring, the group with the least power to contest either finding is the one the payroll data says is actually losing ground.
Two credible studies about entry-level jobs point in opposite directions because they are measuring different resolutions of the same economy, not different economies.
No regulator or agency currently tracks this specific gap the way Stanford does, which is itself notable: a university payroll-data partnership with a single payroll processor is, for now, the most granular public evidence of one of the labor market's most consequential entry points quietly narrowing. That's thinner public infrastructure than the size of the claim would suggest -- if the gap Stanford is tracking is real and keeps widening, the country will have found out about it first from one lab's dashboard, not from official labor statistics built to catch it.
Neither side of this is finished. The NBER-organized survey's own respondents expect the picture to shift: executives forecast AI will cut employment by 0.7% and raise productivity by 1.4% over the next three years -- more than double what they report having seen so far. Stanford's dashboard updates on a rolling basis as fresher ADP data arrives. For a 22-year-old job-hunting in software engineering or customer service today, the executive survey's aggregate comfort is cold comfort against a payroll-data gap that has grown every single time someone has re-measured it.
- Stanford's payroll-data study finds a 19% employment gap for young workers in AI-exposed jobs.
- That gap has widened from 13% to 19% across three measurements since mid-2025.
- A separate NBER-organized survey of nearly 6,000 executives found over 90% saw no employment effect.
- The two studies measure different things: narrow payroll data versus broad executive self-reports.
- Executives' self-reported numbers could mask a narrow entry-level effect too small to notice in aggregate.