AI-exposed jobs are shrinking for young US workers. What the September payroll data really show.

New US payroll figures released on October 21 by ADP Research and Stanford's Digital Economy Lab sharpen a disturbing question for people entering the workforce: is the first rung of the career ladder becoming harder to reach as AI spreads? The latest Canaries Dashboard update reports that employment in occupations rated highly exposed to AI shrank 0.5% in September compared with a year earlier. Jobs in the least-exposed occupations grew 1.0%.

The difference was sharper among employees aged 22 to 25. Within this age bracket, the high-exposure group saw a 3.7% year-over-year decline, while the least-exposed group edged up 0.2%. Across all occupational exposure levels, employment among the same young workers was down 2.2%. The dashboard reports 36 consecutive months of year-over-year declines for the young, high-exposure group, starting in October 2023.

These are measures drawn from ADP's private-sector payroll data. They describe differences between categories of employment, not a controlled experiment in which comparable workers were randomly assigned access to AI. That distinction matters because the same period contains economic uncertainty, industry-specific hiring swings, graduate-supply changes and decisions to pause hiring rather than dismiss existing workers.

An occupation can be 'exposed' to generative AI when a substantial share of its tasks can plausibly be performed or assisted by models. Exposure is not a measure of actual AI adoption by every employer, and it is not the share of positions already automated. Software development and customer support are useful examples: a model may draft code or resolve a standard ticket without handling security, responsibility or complicated customer needs.

Stanford's October 21 explanation of its AI Economic Indicators reports a parallel pattern in software and customer-service employment and warns against over-reading subgroups that may not be representative. The researchers' August paper, using data through June, found young workers in highly exposed roles 19% below a comparative employment trend. That 19% level gap and the new 3.7% annual contraction are two different statistics. They must not be added together.

If employers automate routine entry-level assignments while retaining experienced staff to supervise systems, hiring can weaken before headline unemployment rises sharply. But the payroll data alone cannot tell how much of the divergence comes from substitution, complementarity, offshoring, cyclical changes or shifts in demand. A credible account needs matched employer adoption and occupational hiring data, not an automatic AI-causation label.

Why should readers care? Entry-level work often teaches the tacit knowledge that later makes a person valuable: handling exceptions, writing for real audiences, diagnosing unfamiliar problems and understanding what a customer actually needs. If fewer people receive those opportunities, the long-term consequences could reach beyond one year's hiring figures. The present data raise that question; they do not settle it.

The most useful next indicators are independent vacancy series, hiring rates by age and occupation, pay, separations and the number of employers using AI intensively. Analysts should compare several data providers and measure whether exposed industries diverge after accounting for business cycles. A simple correlation, however attention-grabbing, is not the same as proving that an algorithm replaced a specific worker.

For prospective employees, a reasonable response is to learn how AI changes the tasks within a chosen occupation while developing domain-specific judgment and evidence of real work. For Cortist's AI & Jobs research hub, the standard is even stricter: any future probability about workforce displacement must define exactly which workers, which outcome, what deadline and what independent dataset would resolve the forecast.