Entry-Level Hiring Fell Where AI Automates, Not Everywhere

What Did the Canaries Study Actually Measure?

Researchers at Stanford's Digital Economy Lab pulled a balanced monthly payroll panel of 3.5 to 5 million ADP records running from January 2021 through June 2026 and asked a narrow question: did employment for 22-to-25-year-olds move differently in occupations rated highly exposed to AI than in ones rated less exposed. It now sits 19% below where it would be had it kept pace with less-exposed peers, with no comparable gap for experienced workers in the same jobs ¹.

The authors are explicit in their own abstract that this is a descriptive pattern, not a causal estimate of what AI did. Four slightly different versions of this number circulate across earlier drafts and later coverage, from 13% to 19%, depending on which data vintage a given article is quoting, so treat any single figure as a snapshot rather than a fixed law. The mechanism matters as much as the headline number: separation rates fell for both more- and less-exposed young workers over the same window, so the gap opened through employers hiring more slowly rather than through people getting pushed out ².

If you are already in an AI-exposed entry-level job, this data does not say your position is at risk. It says the door behind you is opening more slowly than it used to for people applying now, which is a search-length problem, not a layoff warning.

Which Kind of AI Exposure Actually Predicts the Decline?

Coverage of this paper almost never asks what "AI-exposed" means, and the answer is where the useful information lives. A separate labor-market analysis of job postings found that skills exposed to automation, meaning AI can substitute directly for the task, were 16% more likely than a baseline group of skills to see posting demand decline. Skills exposed to augmentation, meaning AI extends what a person does rather than replacing it, were 7% more likely to see demand increase ³.

The two are not opposites happening to different jobs. The same institute found that occupations experiencing the most automation are often the same ones experiencing the most augmentation, at the skill level rather than the job level, so the split runs inside a role as much as between roles.

That split is the actionable part of the entry-level finding. A role built mostly from tasks a model can complete unsupervised, formatting a report, transcribing a call, producing a first-pass summary someone else will rewrite, sits on the automating side. A role built from tasks where AI output still needs a person to catch what it got wrong, triage which client actually needs a callback, decide which finding matters, sits on the augmenting side. The occupation-level decline in the Stanford data tracks that distinction more closely than it tracks any single industry.

Read the Counter-Evidence Honestly

Two qualifications matter most. The regression estimate behind the headline roughly halves once the researchers control for the occupation's typical level of formal education, from -0.18 to -0.09 in the paper's own Table 1, with the smaller coefficient significant only at the 10 percent level. Those coefficients are a different quantity from the 19% figure, and the paper says plainly that it cannot fully separate an AI effect from a pandemic-era disruption to how educated labor markets absorb new graduates ⁴.

The pattern is also not confined to technology. Dropping computer occupations and technology firms from the sample barely moves the estimate, so this is not simply the tech hiring slowdown wearing an AI label ⁵.

The most useful counter-evidence comes from outside the paper entirely. LinkedIn's own January 2026 labor market report, built from a different population and a different method, states outright that entry-level roles have not been disproportionately affected relative to experienced roles, and that the entry-level hiring share is returning toward historical norms after a post-pandemic high ⁶. The largest payroll dataset and the largest professional network disagree about the same period, which is a reason to weight evidence from your own field over any national headline, including this one.

What Should a New Grad Actually Do With This?

Aim at the mechanism, not the fear. The paper's own proposed explanation is that generative AI substitutes best for codified knowledge, the formal, documented content a textbook teaches, and complements tacit knowledge, the kind of judgment built by watching someone experienced handle a real, messy situation. Occupations scoring higher on codified knowledge show slower entry-level growth in the data; ones scoring higher on tacit knowledge show faster growth for the people already working in them ⁷.

The authors call this a raw, non-causal gradient rather than a settled mechanism, but it points toward something you can act on: target roles and employers that still put a junior next to someone experienced, rather than roles built entirely around independent, documentable output. A few concrete moves follow from that:

• Read job postings for who you would sit next to, not just the title, and ask in an interview how a new hire's first ninety days actually run
• Weight your own field's hiring signal over a national headline, since the Stanford and LinkedIn numbers genuinely disagree at the aggregate level
• Look specifically at structured entry points built for supervised reps. Choosing between an apprenticeship, an internship and a rotational program is largely a question of how long the role takes to become useful in, and that structure matters more here than the job title does

The practical response to a slower door is spending your search time on the version of the field where someone experienced is still doing the training, because that is exactly the condition the data says is holding up.

See a sample report >