Research update · 28 September 2026
Relative pay for workers who stayed fell in AI-exposed U.S. occupations. Posted pay rose.
A new U.S. working paper finds a split between wages in job ads and wages earned by people who stayed in their jobs. In more LLM-exposed occupations, real posted wages were 3.2% higher per exposure standard deviation by 2026Q2, while incumbent real wages were about 0.9% lower in the later post period. These are relative associations, not proof that AI caused either change.
What was measured
Kevin Rinz compares U.S. occupations by large-language-model task exposure, using Lightcast full-time job ads and Current Population Survey records through 2026Q2. The paper tracks AI language in ads, posted pay, hires, separations and realized weekly earnings. The job-ad series continues to August 2026.
The pay and hiring split
Ads in more-exposed occupations mentioned AI more often. Relative posting volume stabilized after an earlier decline. Hires and separations both rose; the estimated net flow out was not statistically significant. Real pay for new hires had a positive point estimate near 1%, but its confidence interval included zero. Continuously employed workers’ real pay fell about 0.9% in the later period. Unemployed workers per new ad rose about 9% per exposure standard deviation.
What it cannot show
The exposure score measures tasks, not which firms actually deployed AI. Many AI mentions in ads were generic or short-lived. Occupational comparisons can reflect other labor-market changes. Posted wages are not accepted wages; more hires and separations are not a net jobs gain or loss. The paper is preliminary and does not give a causal AI wage effect.
Why the evidence map changes
The new result narrows a tempting claim from earlier posting data: higher advertised pay in AI-exposed work does not mean every worker in that work got a raise. It also does not erase the entry-level warnings. The paper finds less hiring among younger workers relative to older workers in more-exposed occupations, but still cannot isolate AI from other causes. The broad U.S. monitor’s lack of a clear economy-wide AI footprint remains compatible with a narrower pay or entry-route change.