The BaseResearchStanford Digital Economy LabAugust 2026

Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence

Stanford researchers analyze ADP payroll records for 3.5 to 5 million full-time US employees through June 2026. They find no evidence of economy-wide job displacement, but employment of 22-25-year-olds in the most AI-exposed occupations is 19% below where it would be had it kept pace with less-exposed peers, mainly through reduced hiring. They call the results descriptive, not causal.

Open the source ↗Top story of the week

Why it matters

The same thing, for different readers.

What it means for you

In this payroll sample, fewer young people are being hired into jobs most exposed to AI, while people already in those jobs are not leaving at higher rates.

For lawmakers

Aggregate employment looks stable while one group falls behind: the authors say moderate aggregate changes can mask larger ones in subgroups, and publish monthly indicators to track them.

For researchers

Descriptive, not causal: the young-worker gap shrinks with an education control, partly predates ChatGPT, and is far larger in this payroll sample than in national surveys.

The question it raises

The authors float that AI substitutes for codified, school-taught knowledge and complements experience; if so, how do people gain experience when entry-level roles shrink?

In their own words

Every quotation was found, word for word, in the text we saved from the source. That shows the words are there; it does not make them true, or prove the speaker meant them as quoted.

From the source’s own text.

“Overall, we do not find widespread job displacement in occupations most exposed to generative AI. However, employment has declined substantially for early-career workers in the occupations most exposed to AI, including but not limited to software development and customer support.”

Erik Brynjolfsson, Stanford University and NBERDiscussion section: the three authors sum up what the payroll data show overall and for early-career workers.

Found word for word in the source

“Employment of young workers in AI-exposed jobs now stands 19% below where it would be had it kept pace with that of their less-exposed peers; experienced workers show no comparable gap.”

Erik Brynjolfsson, Stanford University and NBERSection 2.2, Fact 2: the paper's headline result for ages 22-25, as the three authors state it.

Found word for word in the source

“Separation rates fell for both groups and, among young workers, fell at least as much in the most exposed occupations as in the least exposed, the opposite of what displacement would imply. The divergence instead reflects hiring, with the gap in hiring rates between more- and less-exposed young workers opening after 2022.”

Erik Brynjolfsson, Stanford University and NBERSection 2.4, Fact 4: the authors on whether the gap comes from hiring or from separations.

Found word for word in the source

“Declines are concentrated in occupations where AI usage primarily substitutes for human tasks; where usage primarily complements workers, employment is flat or rising, especially for experienced workers.”

Erik Brynjolfsson, Stanford University and NBERSection 2.5, Fact 5: the authors on automating versus complementary uses of AI, measured with Anthropic Economic Index usage data.

Found word for word in the source

“These patterns attenuate when controlling for education, show some divergent trends predating generative AI, and are more pronounced in the ADP analysis sample than in national survey benchmarks, with some evidence of consistent patterns in government administrative data. We interpret these facts as early, descriptive indicators—canaries in the coal mine—rather than causal estimates”

Erik Brynjolfsson, Stanford University and NBEREnd of the abstract: the authors on what weakens the pattern and how far the six facts can be read as causal.

Found word for word in the source

How the finding was reached

In our words, from the source. Weigh the finding by how it was made.

What was done: Descriptive comparison of monthly payroll records from ADP: a balanced panel of firms with 3.5 to 5 million full-time employees a month, January 2021 to June 2026. Occupations are grouped into five bands of AI exposure (task ratings made with GPT-4, plus Anthropic Economic Index usage data in which Claude labels conversations as automating or augmenting work), and employment is tracked by age group from November 2022. Robustness checks and comparisons with national surveys (Current Population Survey, American Community Survey) follow.
What it found: No sign of widespread job losses: average employment in the sample rose about 6% between November 2022 and June 2026 (about 4% in the most exposed fifth of occupations). But employment of 22-25-year-olds in the most AI-exposed occupations is 19% below where it would be had it kept pace with less-exposed peers (15% in the July 2025 data), with no comparable gap for experienced workers. In levels, it fell about 11% in the two most exposed fifths while rising about 10% in the three least exposed. The gap runs through reduced hiring, not more separations, and is concentrated where AI use automates tasks rather than complements workers; compensation trends differ little by age or exposure.
What it does not show: The authors call these early, descriptive indicators, not causal estimates. The young-worker gap shrinks when occupational education is controlled for (the most-exposed estimate falls from -0.18 to -0.09), part of the divergence predates ChatGPT, and the gap is far larger in this sample than in the American Community Survey (-0.132 against -0.022 for 2022-24). The sample leans toward large firms and AI-exposed occupations, base pay leaves out bonuses and equity, firm-level AI adoption cannot be observed, and within-firm estimates for the most exposed fifth are sensitive to data-pipeline choices and not significant under the 2025 filters, and the codified-versus-tacit knowledge measure is scored by a language model and called suggestive. The data are proprietary, and ADP may review the paper to prevent disclosure of confidential information.
Status: A working paper: not yet peer reviewed. The authors say Census administrative data (Tucker, 2026) show consistent raw patterns by age and industry exposure. A study using online profiles and job postings (Lambert and Schindler, 2026) finds remote work attenuates the AI link; following that specification on ADP data, the authors do not reproduce it and attribute the difference to the data.

The record

Questions: What happens to work, income and the economy?
People: Erik Brynjolfsson (Wrote it, Stanford University and NBER) · Bharat Chandar (Wrote it, Stanford University) · Ruyu Chen (Wrote it, Stanford University)
What we read: We read the whole source.
Found by: A research pass: work, income and the economy, gathered October 13, 2026, via research-work (tier 1: the source’s own channel or an official record)
Cite this: Stanford Digital Economy Lab, “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence”, August 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/stanford-2026-08-canaries-coal-mine-ai-employment

Our notesWe read the paper's main text in full and the online appendix's prose; the appendix figures and tables were not checked. The paper is dated August 2026, with no day. Claude, made by Anthropic, wrote this entry, and one of the paper's six facts (automating versus complementary use of AI) rests on Anthropic's usage data; the entry is not otherwise about Anthropic.