Stanford’s Digital Economy Lab has updated its Canaries in the Coal Mine? research, and the headline number moved in the wrong direction. Employment among workers aged 22 to 25 in highly AI-exposed occupations now sits about 19% below where it would be had it kept pace with same-aged workers in less-exposed occupations. A year ago that gap was 15%.
The study, led by Erik Brynjolfsson and colleagues, runs on high-frequency administrative payroll data from ADP covering millions of US workers, updated through June 2026. It is not a survey of what executives think AI is doing. It is what actually cleared payroll.
And the same dataset says there is no evidence of widespread, economy-wide displacement. Experienced workers show no comparable gap. Older cohorts in AI-exposed occupations are flat or expanding. The damage is real, measurable, and almost entirely confined to people at the start of their careers.
Our take: This is why the aggregate jobs number keeps looking fine while everyone under 26 says the market is brutal. Both are true. AI is not producing a wave of layoffs — it is producing a quiet, distributed hiring freeze on the bottom rung, and a hiring freeze does not generate a press release, a WARN notice, or a line in the Challenger report. It generates nothing at all. That is precisely what makes it hard to argue about and easy to under-react to.
The mechanism is doors that never opened
The most useful finding in the update is not the size of the gap but how it is produced. The divergence runs almost entirely through reduced hiring of young workers, not through increased separations. Firms are not walking junior staff out. They are declining to bring the next cohort in.
That distinction has consequences. Layoffs are visible, legally scaffolded, and politically expensive. Non-hiring is invisible, costs nothing, requires no announcement, and can be reversed silently if it turns out to be wrong. If you are a company that suspects AI covers some fraction of your entry-level workload but is not certain, freezing the funnel is the cheap option in every direction. Enough firms take the cheap option and you get a 19% hole with nobody responsible for it.
Substitution versus complement is the whole ballgame
The declines are not spread across all AI-exposed work. They concentrate in occupations where observed AI usage primarily substitutes for human tasks. Where usage primarily complements the worker, employment is flat or rising.
That is the actionable line in the entire paper, and it cuts across job titles rather than along them. Two people can hold the same role, in the same industry, and land on opposite sides of it depending on whether the AI in their workflow does the task or helps them do a bigger one. The question is not “is my job AI-exposed.” Almost everything is. The question is whether the exposure in your specific workflow replaces output you were producing or multiplies it.
For anyone early in a career, that reframes the defensive move. Learning the tool is table stakes and no longer differentiating. Positioning yourself where the tool is an amplifier — work with judgment, client contact, ownership of an outcome rather than a task — is the thing the data actually rewards.
What to watch
- Whether the gap keeps compounding. 15% to 19% in a year is the trend that matters more than either level. A third annual reading tells you whether this is a transition or a new floor.
- The age gradient. Modest declines already show up in the next-youngest cohorts. If the effect creeps up the age curve, the “experienced workers are fine” finding has a shelf life.
- Graduate hiring commitments for 2027. Campus recruiting decisions get made a year ahead. Whatever firms do this autumn is already priced into next year’s data.
- The gap between what executives say and what payroll shows. Survey after survey has CEOs reporting AI has had no headcount effect. ADP’s records disagree for one specific cohort. Trust the payroll.
The framing to retire is “will AI take jobs.” On this evidence it is not taking jobs in aggregate. It is taking the first one — the underpaid, over-supervised, genuinely educational job that everyone currently senior used to learn on. Nobody has yet explained where the next generation of experienced workers is supposed to come from.