Stanford: a 19% gap now tracks AI exposure for the youngest workers

Stanford's August 2026 update of its "Canaries in the Coal Mine?" working paper finds workers aged 22 to 25 in the most AI-exposed occupations are now 19% below their peers in less-exposed fields, up from roughly 15% a year earlier. The drop shows up in hiring, not layoffs.

The paper draws on anonymized payroll data from ADP, an HR and payroll company whose records cover millions of US workers. The team compares employment in occupations rated as having a high share of tasks that current AI tools can plausibly do against less-exposed peers. The researchers measure that exposure two ways: a "potential labor market impact" index from earlier research, and the Anthropic Economic Index, a public report that scores occupations by how heavily they use Anthropic's Claude AI assistant in real work. The 19% gap holds up across both measures.

The working paper finds the adjustment is operating primarily through reduced hiring of young workers into exposed occupations, with no widespread economy-wide job displacement. In complementary-use occupations, where AI augments rather than substitutes for human work, employment is flat or rising, especially for experienced workers.

The deeper split is between codified and tacit knowledge. Codified-knowledge work is the formal, standardized, teachable material that can be put in a textbook: tax preparation, basic legal drafting, customer-service scripting. Tacit-knowledge work is practice and mentorship-based: skilled trades, senior client management, complex negotiation. The release finds the codified-knowledge tier is shrinking for the youngest workers. The tacit-knowledge tier is growing, but mostly for older workers.

The release asks, "Is AI causing these changes?" and answers, "We cannot yet answer that question definitively." Alternative explanations, including interest rates, remote work, education shifts, and the entry and exit of firms, are examined and aren't fully sufficient, but they aren't ruled out. The pattern is descriptive, not a causal estimate. The quote circulating in coverage — that the labor market could "keep overall employment level while quietly closing the on-ramp for new entrants" — comes from a Washington Post interview reported by Ars Technica, not from a direct conversation with a Type0 reporter. The data also come with a sample caveat: ADP's payroll universe is broader than the Current Population Survey in some ways and narrower in others, and its coverage gaps are larger, so the 19% figure describes ADP's universe of workers rather than every young worker in the US economy.

Two other findings matter for how the data lands. Higher education is a partial buffer: gaps between more- and less-exposed occupations are muted in fields with more college graduates. Women face greater AI exposure on average, which the authors flag as a heterogeneity worth watching.

If the gap is hiring-driven and concentrated in codified-knowledge entry-level work, the obvious move is to redesign the bottom rung. Apprenticeships, internships, and on-the-job training programs can shift new entrants toward tacit-knowledge roles where the data so far shows the older cohort is gaining. Employers can audit which codified tasks in an entry-level job are now being done by AI and replace them with tasks that require judgment, client contact, or physical presence. Universities and bootcamps can stop teaching the codified material that AI now does and start teaching the human work that sits next to it.

A short watch list for a reader's own field: is the entry-level cohort shrinking in headcount, or being re-tiered into different roles? Are new hires doing the same tasks as three years ago, or spending more time on exceptions, escalations, and client-facing work? Is the gap driven by hiring or by attrition?

The Stanford Canaries dashboard is updated each month with ADP data. The next data point will show whether the 19% gap is still widening, has plateaued, or has begun to narrow as the labor market digests the new shape of entry-level work. For now, the data say the adjustment is narrow, named, and happening at the bottom.