What was published today
Erik Brynjolfsson, Bharat Chandar and Ruyu Chen of the Stanford Digital Economy Lab revised "Canaries in the Coal Mine?" today, extending it through June 2026 on ADP administrative payroll data covering millions of American workers.
Employment of 22-to-25-year-olds in AI-exposed occupations now sits 19% below where it would be had it kept pace with their less-exposed peers. The November 2025 version put the same gap at 16%.
That is worth pausing on before anything else. The finding is not a snapshot that has been confirmed; it is a series, and the series is widening. A reader who saw the headline nine months ago and files this one as "same story again" has missed the only thing that changed.
| Version | Data through | Relative employment gap |
|---|---|---|
| November 2025 | mid-2025 | 16% |
| Revised 12 August 2026 | June 2026 | 19% |
The four sentences that will not survive the headline
A 19% employment decline reads as a mass firing. The paper says the opposite, in its own words, four times over.
It reports no evidence of widespread, economy-wide job displacement. The divergence operates primarily through reduced hiring of young workers rather than increased separations. Adjustment is occurring through employment rather than base compensation. And the declines concentrate in occupations where AI usage primarily substitutes for human tasks, while where usage primarily complements workers, employment is flat or rising.
Put plainly: the people already inside are not being pushed out and are not being paid less. The door is opening less often, in some occupations and not others. That is a slower, quieter and more selective event than a wave of layoffs, and it calls for a completely different response from anyone at the start of a career.
| The question | What the paper says |
|---|---|
| Are young workers being laid off? | No. The divergence operates through reduced hiring, not increased separations. |
| Is pay falling for them? | No. The adjustment is in employment, not in base compensation. |
| Is this the whole economy? | No. The authors report no evidence of widespread, economy-wide displacement. |
| Is it every AI-exposed job? | No. Declines concentrate where AI substitutes for tasks; where it complements, employment is flat or rising. |
Substitute or complement is the only line that tells you what to do
Every part of this finding is a description except one. "AI substitutes for the tasks" versus "AI complements the worker" is the distinction that separates the occupations where the gap opened from the occupations where employment held or grew — and it is the one line in the paper a twenty-three-year-old can act on this week.
It is also the line that gets cut. It survives poorly in a headline, it resists a single number, and it requires saying that the effect is uneven, which is less satisfying than saying it is universal.
The practical form of the question is not "is my field being automated". It is narrower and more answerable: in the specific job I am about to spend an evening applying for, does the tool do the task instead of me, or does it make me faster at it? The paper cannot answer that for your shortlist. It can tell you the question is the right one.
The authors publish the limit of their own evidence
Two caveats sit in the paper rather than in the coverage, and both are the authors' own.
They interpret the findings as early, descriptive indicators — canaries in the coal mine — rather than causal estimates. The paper is not claiming to have proved that AI caused the gap; it is claiming the gap is there and is shaped like something AI would cause.
And they note that the effects are more pronounced in the ADP analysis sample than in national survey benchmarks. That is a research team telling you, unprompted, that its own dataset runs hotter than the national picture. Very little of what you will read about this study will carry that sentence, and it is the single most useful thing in it for judging how much weight the 19% can bear.
What this changes about looking for work, and what we can and cannot help with
If the mechanism is absent hiring rather than dismissal, then the scarce resource for a young candidate is not resilience. It is information about which doors are still open, early enough to spend the evening on those instead of the others.
This desk belongs to a company that builds a job app, so the useful thing is to be precise about what that does and does not solve. We score every role on six published dimensions and show all six before you apply, which means you can see where a role is weak for you without spending the evening finding out. That helps with allocation of effort. It is the right shape of help for a market where the constraint is which applications are worth making.
What we cannot do is tell you whether a given job is one where AI substitutes or complements. Nothing in our data carries that, and the honest answer is that the Stanford occupation-level exposure work is a better guide to it than anything on our site. We would rather say so than let a useful distinction get quietly absorbed into a product claim.
The six dimensions this desk's company scores every role on, and the audit that tests them, are at babzituna.com/bias. Neither tells you whether a job is automated or augmented; the paper above is the better guide to that.