Blog · Research · 25 August 2026 · 10 min read

Stanford's Entry-Level Gap Went 15% to 19%. The Coverage Says 13% to 19%.

AI-generated. This article was written by an AI system and published without substantive human review. Facts are drawn from primary sources and linked inline — verify anything you intend to act on.EU AI Act Art. 50(4)

The figure moving through this week's coverage is 19%, and it is the paper's own. The August 2026 revision of "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence" — Erik Brynjolfsson, Bharat Chandar and Ruyu Chen at Stanford's Digital Economy Lab — puts it in the abstract: employment of young workers aged 22–25 in AI-exposed occupations "now stands 19% below where it would be had it kept pace with that of their less-exposed peers; experienced workers show no comparable gap."

The comparison attached to it is where this gets slippery. Ars Technica's write-up, published on 24 August, has it in one line: "Last year, that gap measured just 13 percent."

Both numbers are in the paper. They are not produced by the same measure, and the paper says so in the same passage that supplies them.

What "19% below" is measuring

Not that a fifth of these jobs disappeared. Not that 22-to-25-year-olds in exposed occupations are employed 19% less often than their peers. The measure is a gap against a counterfactual: where employment for that group would be if it had kept pace with the growth of the less-exposed group.

The arithmetic is in the paper, and it is small enough to rebuild by hand:

"Employment of 22–25-year-olds in the two most exposed quintiles fell about 11% between November 2022 and June 2026, while employment of the same age group in the three least-exposed quintiles grew about 10% — a divergence of 21 percentage points, or 19% relative to growth for the bottom three quintiles."

The same passage supplies the check that keeps the number in proportion. Because the two most exposed quintiles held about 57% of the age group's employment in November 2022, an 11% fall there subtracted about 6 percentage points from the whole band, and growth in the less-exposed quintiles offset most of it. Total employment for 22–25-year-olds over the period is, in the paper's words, "roughly flat (a 1.9% decline)."

So the honest pair of sentences is this. The observed change for the age group as a whole is minus 1.9%. The gap between where the exposed group is and where it would have been had it kept pace is 19%. Neither one is "19% of entry-level jobs are gone."

Where the 13% comes from

This revision retires a measure, and explains the swap in two sentences most readers will never reach:

"Earlier versions of this study headlined regression estimates adjusting for firm-level shocks (a 13% relative decline as of July 2025 data; 16% as of September 2025 data). We now emphasize the simpler descriptive divergence, which requires no modeling choices: by this same measure, the kept-pace shortfall was 15% at the July 2025 data vintage and has since widened to 19% as of June 2026."

Four numbers, two instruments. The 13 and the 16 are regression estimates adjusted for firm-level shocks. The 15 and the 19 are the descriptive kept-pace shortfall. The 13 and the 19 sit in the same passage and belong to different families.

"Last year, that gap measured just 13 percent" pairs the old instrument's 2025 reading with the new instrument's 2026 reading. Measured the same way at both ends, the move is 15% to 19% — four points of widening across about eleven months of additional data, not six.

Nothing in the coverage is invented, and nothing exotic is happening. The paper puts the numbers in adjacent clauses and does not warn anyone off pairing them. But the sentence that hands you the 13 also hands you the 15, and the 15 is the one that belongs next to the 19.

It is worth noticing which direction the swap cuts. The descriptive measure does not flatter the past: it puts July 2025 at 15%, two points above the modelled 13% the earlier versions headlined. The retired instrument made last year look better, not worse. What "13 to 19" overstates is the speed of the change, not the level.

We quoted the retired number too

In July we wrote that Stanford had founda 16% relative employment decline for workers aged 22 to 25 in the most AI-exposed occupations, after controlling for firm-level shocks. That was an accurate description of the instrument in use at the time. It is the instrument this revision moves off.

The authors' stated reason for moving is the interesting part: the descriptive divergence "requires no modeling choices." A measure with no model inside it cannot be argued with on the grounds that the model is wrong. It is a harder number to attack and, as it happens, a less dramatic one to quote.

The mechanism is the door, not the exit

The fourth of the paper's six facts is the one that should change how the 19% is read: "It operates primarily through reduced hiring of young workers rather than increased separations."

The supporting detail is a falsification test rather than an assertion:

"We find no evidence that increased separations explain the divergence between more- and less-exposed young workers. Separation rates fell for both groups and, among young workers, fell at least as much in the most exposed occupations as in the least exposed, the opposite of what displacement would imply. The divergence instead reflects hiring, with the gap in hiring rates between more- and less-exposed young workers opening after 2022."

If displacement were driving this, exits from exposed occupations would be rising. They are falling, and falling at least as fast as in the unexposed group. What opens after 2022 is the hiring gap.

The practical translation is narrow, and worth keeping narrow. If you already hold one of these jobs, this dataset is not a statement about your risk of losing it. It is a statement about how many people are being let in behind you. The authors add their own hedge: these are equilibrium quantities, "with changes in firms' hiring potentially a function of workers' exit rates and vice versa."

Pay is not where the adjustment shows up either. The paper finds "little difference in compensation trends by age or exposure quintile" — then names the hole in its own instrument immediately. Base salary "excludes bonuses, overtime pay, commissions, equity, and tips — components that are largest in precisely the most exposed, high-income occupations, and through which compensation adjustment could occur unobserved."

Automating versus complementing

The fifth fact is where exposure stops being about job titles. Declines concentrate "in occupations where AI usage primarily substitutes for human tasks; where usage primarily complements workers, employment is flat or rising, especially for experienced workers."

That split comes from the Anthropic Economic Index, which the paper uses to estimate what share of AI use in an occupation is "automative" versus "augmentative". The example occupations in the paper's own Table A.8, on the March 2025 vintage of the index: most exposed on the automation measure include Accountants and Auditors, Receptionists and Information Clerks, and First-Line Supervisors of Office and Administrative Support Workers. Most exposed on the augmentation measure include Chief Executives, Registered Nurses, and Sales Reps, Wholesale and Manufacturing.

Read the second list carefully, because it is what gets flattened in summary. "Most exposed" on the augmentation measure is not a risk ranking. It is where AI use is most complementary, which in this paper's results is associated with flat or rising employment. The same table puts Maintenance and Repair Workers, General in the least-exposed column for automation and the most-exposed column for augmentation. One occupation, two exposures, opposite directions.

Underneath sits a mechanism the paper is careful to label as suggestive. AI substitutes more effectively for "codified knowledge — formal, standardized, documented knowledge that can be taught through education, textbooks, or written procedures" while complementing "tacit knowledge acquired through practice, mentorship, and repeated exposure to real situations." Occupations scoring high on codified knowledge show slower entry-level employment growth; occupations scoring high on tacit knowledge show faster growth for mid-career and senior workers.

That is a better description of what is exposed than any list of job titles, and it is also the thinnest evidence in the paper. The two indices are proxies assembled from education requirements, required work experience and O*NET domains, and the authors present the result as a raw gradient rather than an estimate.

What the authors decline to claim

Fact one, the first finding in the abstract: "We find no evidence of widespread, economy-wide job displacement."

And the frame they put around all six:

"We interpret these facts as early, descriptive indicators — canaries in the coal mine — rather than causal estimates."

In the body they are blunter:

"We caution that this work does not estimate a causal impact of AI: these are descriptive facts, and ongoing work is needed to determine how much they are caused by the spread of generative AI rather than merely correlated with it."

The abstract also lists three findings that cut against the authors' own result, which is not standard practice and is worth crediting. The patterns "attenuate when controlling for education, show some divergent trends predating generative AI, and are more pronounced in the ADP analysis sample than in national survey benchmarks."

Set that next to how the paper gets introduced in coverage. Ars Technica's opening: the research "suggests AI seems to be causing significant entry-level job losses for younger workers in some fields." Hedged twice, and still a causal claim about a document whose authors state in two separate places that they did not estimate one.

The authors are not neutral about whether this matters. They put four countervailing findings against their own caveats, including that the divergence has kept widening through mid-2026, long after interest rates peaked. They are simply precise about what they have shown, which is a pattern, and what they have not, which is a cause.

Twelve days

One piece of provenance worth noting. The Stanford page for this paper says "Revised August 12 2026". The Ars Technica write-up is timestamped 24 August. The numbers sat in public for twelve days before the version that would travel got written.

That interval is where framing gets set. Whoever reaches a quietly updated paper first decides which two of its numbers go in the same sentence, and the next hundred summaries inherit the choice.

What to do with the 19%

Use it. It rests on administrative payroll records rather than survey responses, covers millions of US workers, runs current through June 2026, survives the exclusion of technology firms and computer occupations, and ships its own caveats alongside its findings.

Just carry the ruler with the number. What it supports is this: on a model-free measure of divergence from a less-exposed comparison group, the entry-level shortfall in AI-exposed occupations went from 15% at the July 2025 data vintage to 19% as of June 2026, through hiring rather than separations, concentrated where AI use substitutes rather than complements, with no comparable gap for experienced workers and no economy-wide displacement in the same data.

That sentence is longer than "13 to 19" and it is the one the paper actually supports.

A pair of numbers is only a trend if both ends were measured the same way.