Stanford update widens the AI gap for entry-level workers

A revised version of the Stanford Digital Economy Lab's 'Canaries in the Coal Mine?' paper, published in August with a larger dataset, finds that employment among workers aged 22 to 25 in highly AI-exposed occupations is now roughly 19% below where it would be had it kept pace with similar-aged workers in less exposed occupations. A year ago the same measure stood at about 15%. The analysis draws on payroll data covering some 4.6 million workers across more than 730 occupations.

Two findings do more work than the headline number. The first is that the decline runs almost entirely through reduced hiring rather than increased separations. Nobody is laid off in this story. Positions that would have been opened are not opened, and the people affected are those who never got the job — a group that generates no severance filing, no redundancy notice and no identifiable victim. It is the least visible way a labour market can contract, and the hardest to legislate against.

The second is the split by experience. Within the same occupations, older workers show no comparable decline. The paper attributes this to the difference between codified knowledge, which a model can reproduce, and tacit knowledge built through judgment and accumulated cases. The distinction is plausible, and it carries an uncomfortable implication: tacit knowledge is acquired by doing junior work. If the junior rungs of software development, customer service and clerical work are being removed, the pipeline that produces the experienced workers currently insulated is being removed with them.

Correlation is doing real work here, and the authors are careful about it. AI exposure correlates with other things that moved over the same period, and no natural experiment separates them cleanly.

But the direction is the finding. The gap widened from 15% to 19% across a year in which the effect was widely expected to prove a measurement artefact. It has not behaved like one.