Is AI Really Shrinking Entry-Level Hiring? The 2026 Data

At HR Tech 2026 in Las Vegas this month, Experian's chief economist Joseph Mayans told a room of HR leaders that employers' AI expectations are slowing white-collar hiring. It's a tidy headline, and it landed in a year when white-collar hiring has genuinely been weak. But Mayans was stepping into an argument that labor economists have been having, in public, for most of 2026 — and the two sides aren't actually describing the same thing, even though they sound like they are.

One side points to specific, data-heavy research with real numbers behind it. The other side points to a broader hiring slowdown and says AI is getting more credit (or blame) than it deserves. Both are right. The disagreement is about scope, not about whether anything is happening — and the distinction matters a lot more to a hiring team than "is AI taking jobs" does, because it tells you where to actually look.

The specific, measurable effect

The most-cited research behind the "AI is doing this" camp is "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence," a paper by Stanford's Erik Brynjolfsson along with Bharat Chandar and Ruyu Chen, first released in August 2025 and updated with a follow-up note in February 2026. It uses monthly ADP payroll data covering millions of workers and tens of thousands of firms, from January 2021 through mid-2025.

The finding is narrow and specific, which is exactly what makes it credible: workers aged 22 to 25 in the occupations most exposed to generative AI saw roughly a 13% relative decline in employment compared with workers in less-exposed occupations. Employment for young software developers specifically was down almost 20% from its late-2022 peak by mid-2025. And critically, older workers in the same occupations did not show the same drop — some held steady, some grew. That asymmetry is the paper's strongest piece of evidence, because it rules out "this occupation is just shrinking" as the full explanation. If the job itself were disappearing, you'd expect the decline to show up across age groups, not concentrated in the newest entrants.

The authors' February 2026 follow-up specifically tested the obvious alternative explanation — that this is really about interest rates and the broader tech-hiring pullback, not AI — and reported that interest-rate timing doesn't line up well with when the entry-level gap actually opened. They also found the effect only becomes statistically clear once you include the broadest set of controls, which is a more honest way to present a finding than researchers who lead with their headline number. The team has since partnered with ADP Research on an ongoing "Canaries Dashboard" tracking 4.6 million workers across more than 730 occupations, so this isn't a one-off study — it's a result someone is actively trying to break.

A second, independent data source says something similar

If the Stanford/ADP finding stood alone, it would be worth noting and nothing more. What makes it worth building an argument on is that a completely different research group, using a completely different dataset and method, found a compatible pattern.

The Federal Reserve Bank of Dallas, in a research note published in September 2026 by economists Samuel Dodini and Tucker Smith, analyzed millions of online job postings (via Lightcast) against an occupation-level measure of how automatable each occupation's tasks are by generative AI — built from observed patterns of how people actually use Anthropic's Claude. No payroll data, no age cohorts, no overlap in method with the Stanford paper. The result: postings for more AI-exposed occupations were about 5% lower than less-exposed occupations by the end of 2023, and that gap had grown to roughly 8% by the first quarter of 2025. At the state level, the Dallas Fed estimates total Texas job postings were 1.8% lower in 2024 and 2.6% lower in 2025 because of AI exposure alone.

A related Dallas Fed note from January 2026 adds the piece that connects most directly to entry-level hiring specifically: among 20-to-24-year-olds searching for jobs in highly AI-exposed fields in 2025, roughly 14.4% landed an offer within a month — down from about 17.6% two years earlier. Meanwhile the employment share for young workers in low-exposure jobs (the Dallas Fed's examples: cashiers, laborers) held steady over the same period. Two different institutions, two different data sources, and they converge on the same shape: a real effect, concentrated in young workers, concentrated in the occupations where generative AI can plausibly do more of the task list.

The case for "slow down before you blame AI"

None of that settles the broader claim Mayans was making at HR Tech — that AI expectations are driving the general white-collar hiring slowdown, not just a specific entry-level squeeze in exposed roles. And on that broader claim, the pushback from other labor economists has been pointed.

At a Deel-hosted economics event in New York in May 2026, Glassdoor chief economist Daniel Zhao argued directly against treating AI as the primary driver, pointing instead to policy uncertainty and energy prices — the same month Glassdoor's own Employee Confidence Index hit a record low, which the company attributed largely to geopolitical conditions rather than automation anxiety. That echoes a line Indeed economist Cory Stahle used back in 2025 to describe the same broader white-collar weakness: "This is more of an economic story and less of an AI-disruption story, at least so far."

Indeed's own Hiring Lab put numbers to the uncertainty in August 2026, surveying more than 100 professional forecasters: asked whether AI would have a net negative effect on US employment over the next year, 52% said it would be at least a mild drag — meaning the panel was essentially split down the middle, with a bare majority even willing to go that far. That is not the result you'd expect if the "AI is driving the whole slowdown" reading were obviously correct; it's the result you'd expect if the true effect is real but narrower than the headline version of the story.

Both things are true, and the scope is the whole story

Put the two bodies of evidence next to each other and they stop looking like a disagreement. The rigorous, data-heavy research — Stanford/ADP, the Dallas Fed — finds a specific, bounded effect: entry-level and early-career workers in a defined set of highly AI-exposed occupations, measured in both payroll data and posting volume, independently. The skeptical economists are responding to a different, bigger claim — that AI explains the broad white-collar slowdown across the board — and on that bigger claim, they have a point: attributing the general slowdown entirely to AI overreaches past what the entry-level-specific research actually shows.

For a hiring team, the useful takeaway isn't "AI is coming for white-collar jobs" or "it's all macro noise, carry on as before." It's narrower and more actionable: if your organization hires into occupations that sit high on an AI-task-exposure measure, the data says to expect real pressure specifically at the entry-level rung of that pipeline — not a case for freezing hiring everywhere, and not a case for ignoring the data either. Treating every function and every seniority level the same, in either direction, is the mistake both camps would agree you shouldn't make.

The research that holds up isn't "AI is replacing white-collar workers." It's "AI is reshaping who gets the first job in a narrow set of exposed occupations" — a much smaller claim, and a much more useful one.

That's also an argument for hiring processes that evaluate candidates stage by stage rather than compressing a decision into one opaque pass. NiceHire's pipeline is built as a configurable, multi-stage flow — resume vetting, an AI screening interview, live human interviews, and more, each a distinct, separately logged checkpoint rather than a single pass/fail AI interaction before an employer decision. And within the screening stage specifically, auto-rejecting candidates who fall below a passing threshold is off by default; unless a team deliberately turns it on, a below-threshold result routes to pending human review instead of silently closing the file. Neither of those is a claim about solving the entry-level compression the research above describes — it isn't, and we're not asserting otherwise. It's a narrower point: when the pressure on junior hiring is real but concentrated, as the data suggests it currently is, a pipeline that keeps a human decision point in the loop by default is a more honest way to apply that pressure than a single automated gate would be.