Digiko: Market Anarchy

Work

AI hit young workers first. Read the entry-level data.

ADP data shows 22-25 workers in AI-exposed jobs are down 13% versus less-exposed peers, while older workers in the same roles stayed flat or grew.

The Editors · 7 min read ·

White corner desk

Three independent studies converge on the same finding in US labor data through late 2025. AI's first measurable effect lands on entry-level hiring in AI-exposed occupations. Aggregate unemployment for those occupations has not moved.

The Stanford Digital Economy Lab finds entry-level hiring fell 13% in AI-exposed jobs versus less-exposed jobs at the same firms, with a 16% relative employment drop for 22-25 year-olds in those roles. Older workers in the same occupations grew or held steady. Anthropic's own labor study, built on what people actually use Claude for, found a 14% drop in job-finding rates for 22-25 in highly exposed occupations since late 2022, and called the finding "just barely statistically significant." The NY Fed shows unemployment for recent college grads at 5.6% in March 2026, up from 3.6% pre-COVID, with remote work explaining the largest single share.

The story is consistent enough to take seriously and confounded enough that you should read it carefully before acting on it.

The 13% number, in plain English

The 13% comes from "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence" by Brynjolfsson, Chandar, and Chen at the Stanford Digital Economy Lab, October 2025.

The team used ADP payroll data, which covers tens of thousands of US firms. Inside each firm, they compared hiring of 22-25 year-olds in AI-exposed roles against hiring of 22-25 year-olds in less-exposed roles at the same firm. Entry-level hiring in AI-exposed roles came in 13% lower than the within-firm baseline. The same paper finds early-career workers in AI-exposed occupations saw a 16% relative employment decline once firm-level shocks are controlled for.

The Dallas Fed, using a different cut, confirms the direction. Between mid-2024 and September 2025, early-career workers in AI-exposed roles saw 16% slower employment growth than the least-exposed young workers. Two methods, two datasets, same sign.

Who is hit and who is not

The Stanford paper flags software development, customer service, and clerical work as the heaviest-hit occupations. Software developers aged 22-25 fell about 20% from a 2022 peak by July 2025.

The same occupations grew or held steady for older workers. The Stanford result holds when the tech sector is excluded and holds when remote workers are excluded. The age gradient is the finding.

Anthropic's March 2026 paper, measured directly from Claude usage rather than from task-list exposure scores, lands on the same three categories: computer programmers, customer service representatives, and financial analysts have the highest measured exposure.

What "AI-exposed" actually measures

Two different definitions, same answer. Stanford uses occupation-level exposure scores built from task lists. Anthropic measures what people pay Claude to do, then maps it to occupations. Both flag white-collar, screen-based, document-and-code work first.

The convergence is the strongest part of the evidence. If only one method showed the pattern, the result would read as a measurement quirk. Both methods showing it makes the pattern hard to wave off.

One distinction worth keeping straight. The exposure score measures whether AI can do the job that used to require you. It does not measure whether you yourself use AI. A junior developer who codes with Cursor every day is in a high-exposure role. The question is what the role required before, and what is left after.

The remote-work confound

The single most important caveat sits at the NY Fed. In their March 2026 update, recent college graduate unemployment held at 5.6%, against a 3.6% baseline from March 2019. Their research argues that remote work can account for 64% of the recent grad unemployment increase: employers find it harder to train new hires on distributed teams, so they hire fewer.

That is the same labor force, the same data window, and a different mechanism. The honest reading: some of what looks like AI is the lingering effect of remote work on the apprenticeship ladder.

The Stanford team checked this. Excluding remote-eligible workers, the age gap inside AI-exposed occupations still holds. The two effects can both be real and stack.

What the Anthropic study adds, and what it does not

Anthropic's own write-up is unusually candid for a tech-company release. They report the 14% decline in job-finding rates for 22-25 in exposed roles since ChatGPT launched, then call the finding "just barely statistically significant" and list three alternative explanations: people staying in their current roles longer, shifting to adjacent occupations, returning to school.

The paper also notes that aggregate unemployment for highly exposed workers shows no clear rise since late 2022. The headline finding sits inside a body of caveats most coverage strips out.

If you want the strongest single source, the Stanford paper is firmer. If you want the broadest exposure measurement, Anthropic's is the only one built from real usage. Read both before you take a position.

A Danish counterexample

A 2024 working paper by Humlum and Vestergaard studied AI adoption in Denmark across 11 exposed occupations and found no detectable change in earnings, hours, or job mobility from generative AI use. Their study covers an economy with stronger labor protections and a different industry mix. That might explain part of the gap.

It also means the "young workers first" story is consistent in US data and absent in at least one peer economy. Read it as a US effect for now. The universal claim does not have peer-economy support.

What to do if you are 22-25

The studies agree on the active part. Employment is falling in roles where AI automates the work and rising in roles where AI augments it. Automation means the AI does the task end to end and the worker is no longer in the loop. Augmentation means the worker uses the AI to do more or higher-quality work per hour.

The career bet sits with roles where AI augments the worker. Roles where AI does the work end to end will keep losing entry slots until firms find a different way to ramp juniors.

That pushes against the obvious "learn to use AI" advice in a useful direction. The exposed occupations (computer programmer, customer service rep, financial analyst) are full of people who already use AI tools every day. Tool use no longer carries the differentiation. The differentiation sits in work the tool cannot finish on its own.

Two adjacent reads from our archive: AI freelance pay in 2026: what Upwork and Fiverr actually report covers the rate-card side of the same shift, where platforms still pay for production AI work and the floor is moving up. How to get paid what you're worth is the prerequisite once you have picked the work.

What to watch next

Three releases will move this story in the next quarter.

  1. The NY Fed's college labor market update refreshes quarterly with new 22-27 unemployment and underemployment numbers. Watch whether the 5.6% rate keeps climbing or rolls over.
  2. The BLS Employment Situation report publishes age- and education-specific unemployment monthly. The long-duration unemployment line is the one to track: median duration of unemployment hit 11.6 weeks in May 2026, the longest since November 2021.
  3. Stanford's Canaries dashboard updates the ADP cut as new payroll data flows in.

If the entry-door pattern keeps widening after the remote-work effect should have stabilized, the AI read gets stronger. If it flattens, remote work wins more of the share. Either way, the people whose career is being decided by the answer are graduating right now.

Sources

This is not financial advice.

More from Work

See all →
Two people shaking hands in front of a laptop

How to get paid what you're worth

The raise tends to go to the person who asks with evidence, not the one who waits.

The Editors · 4 min read