Young Workers in AI-Exposed Jobs Are 19% Below Where They Would Be, a Stanford Study of Millions of Payroll Records Finds, and Its Authors Say That Does Not Prove AI Did It

Employment of 22 to 25 year olds in the American occupations most exposed to AI is 19% below where it would be had it kept pace with their less-exposed peers. That is the headline of a Stanford working paper revised on 12 August 2026, built on payroll records for 3.5 to 5 million US workers a month. The same paper finds no sign of job losses across the wider economy, and its three authors say the 19% describes a pattern and does not measure what AI caused.

How the 19% is worked out

Erik Brynjolfsson, Bharat Chandar and Ruyu Chen follow two groups of 22 to 25 year olds from November 2022, the paper’s starting point for the generative AI era, to June 2026. Those working in the two most AI-exposed fifths of occupations saw their employment fall by about 11%. Those in the three least-exposed fifths saw it rise by about 10%. The paper puts that 21-point difference at 19% relative to the growth of the less-exposed group.

The 19% is a shortfall against a benchmark, and it is not a count of lost jobs. All 22 to 25 year olds in the sample were only about 2% below their November 2022 level in June 2026, because growth in the less-exposed jobs offset most of the fall. Every older age group grew, by as much as 11% for workers aged 35 to 40. Across all ages, employment in the sample rose about 6%, and the most exposed fifth of occupations grew about 4%.

The data and the exposure scores

The records come from ADP, a payroll processor whose clients employ over 26 million US workers. The analysis uses a balanced panel of firms present in every month from January 2021 to June 2026, counting full-time workers under 70. Job titles are missing for roughly 30% of the sample, and the authors fill gaps where they can from the same worker’s nearest recorded title.

To rank occupations by exposure, the authors use two scores. The first, from Tyna Eloundou and colleagues, rates how well each occupation’s tasks match the abilities of large language models, the technology behind chatbots such as ChatGPT. The second comes from the Anthropic Economic Index, which sorts conversations with Anthropic’s Claude chatbot by the work task involved and by whether the user hands the task over (automation) or works alongside the model (augmentation). The Anthropic paper behind the index analysed over four million conversations and put 57% of usage at augmentation and 43% at automation. Anthropic sells the chatbot whose usage records feed this score.

Why older workers are the control group

The comparison leans on experienced workers. If something other than AI had hit every exposed occupation, older staff in those jobs would show the same gap. They do not. In the most exposed fifth, employment for 22 to 25 year olds grew about 18 percentage points less than in the least exposed fifth. For workers aged 35 to 40 the matching figure is 0.001.

The decline comes through hiring. Separation rates fell for both exposed and unexposed young workers, and fell at least as much in the exposed jobs, so layoffs do not explain the gap. The gap in hiring rates opened after 2022.

The paper also splits exposure by the kind of AI use. For 22 to 25 year olds, each standard deviation of automation share in an occupation goes with about 10 points less employment growth. Complementary use shows no such effect for the young and a positive one for workers aged 41 to 49 (plus 0.024). Base pay differs little by age or exposure, though base pay leaves out bonuses, overtime, commissions and equity, which are largest in exposed, high-income jobs.

The number has also moved. Earlier versions led with regression estimates of 13% (July 2025 data) and 16% (September 2025 data). The August 2026 version leads with a simpler comparison, which gives 15% on the July 2025 data and 19% now, so the three headline figures come from two calculations.

What the authors say it cannot show

The paper says its facts are “early, descriptive indicators” and that “this work does not estimate a causal impact of AI.” Four limits sit in the text.

Education shrinks the result. In the main sample, controlling for each occupation’s share of college graduates cuts the 22 to 25 estimate from minus 0.18 to minus 0.09, still significant at the 10% level. In a sample balanced from 2021 the figures are minus 0.19 and minus 0.06, and the second is no longer significant. The authors treat the two as a range, since education may be a rival cause or the route by which AI acts.

Part of the pattern predates generative AI. Exposed occupations boomed against the rest in the pandemic and peaked around mid-2020, then declined. Of a roughly 30-point fall since, about 13 points undo that boom and about 17 take the gap below its 2018 to 2019 level. The authors take that as a reason to weigh structural explanations over business-cycle ones, and still say it cannot be read as the effect of AI.

The ADP sample is not the country. It overrepresents manufacturing, wholesale, larger firms and occupations with high AI exposure. For 2022 to 2024, the American Community Survey, a Census Bureau survey, shows a gap of minus 0.022 (interval minus 0.055 to plus 0.011), against minus 0.132 in ADP. In professional, information and financial services the two sources agree closely, at minus 0.231 and minus 0.213. The authors write that the size of the divergence “appears specific to the ADP sample” and that its direction is consistent.

The raw data cannot be inspected. It is proprietary payroll data under a data-use agreement, ADP has the right to review the paper to prevent disclosure of confidential information, and ADP supports the lab as a corporate affiliate.

The critique: remote work

Peter John Lambert, an economist at Warwick and the London School of Economics, and Yannick Schindler of the Ellison Institute of Technology in Oxford argue that studies like this one blame AI for a decline that remote work explains better. Their paper, The Broken Ladder, dated May 2026, uses 243 million new-hire records and 407 million online job postings from the US, UK, Canada and Australia for 2017 to 2025. By 2025, they find, the share of new hires going to early-career workers was 8% to 11% below its 2019 level in all four countries.

Their point is that AI exposure and remote-work exposure fall on the same white-collar jobs, with a rank correlation of 0.77 across occupations. Taken alone, a rise of two standard deviations in either exposure predicts a drop of about 5 percentage points in the junior share of new hires. Entered together, the remote-work effect stays and the AI effect weakens sharply, often to a size statistically indistinguishable from zero. Schindler says AI may well still matter for junior hiring, and that the evidence suggests it was not the main cause of the post-pandemic slowdown.

Brynjolfsson, Chandar and Chen answer in their section 4.3. They replicate Lambert and Schindler’s specification as closely as they can on payroll data and find that AI exposure still predicts a falling junior share of new hires once remote work is accounted for. They put the difference down to data, since Lambert and Schindler use online professional profiles and job postings. The Stanford paper says future work is needed to separate the effects of AI and remote work, because firms that adopted remote work were also more likely to adopt generative AI.

Interest rates are another rival cause, which the Stanford paper cites to Zanna Iscenko and Fabien Curto Millet among others. Occupations with high interest-rate exposure tend to have low AI exposure, and adding that control moves the 22 to 25 estimate from minus 0.179 to minus 0.178.

Two other readings of the same years

A Federal Reserve Bank of Dallas note by Tyler Atkinson and Shane Yamco, published on 6 January 2026, used a different source, the Census Bureau’s Current Population Survey of about 100,000 people. It found that the share of young workers’ employment in the most exposed occupations slipped from 16.4% in November 2022 to 15.5% in September 2025, with young workers defined as ages 20 to 24. Layoffs did not rise for that group, and the authors judged the effect on the overall unemployment rate small.

The Yale Budget Lab’s tracker, published on 16 July 2026 and updated on 15 September 2026, reaches a flatter conclusion. It reports that the occupational mix is not yet changing in ways that clearly align with AI’s arrival, and that AI usage measures show no connection to changes in employment or unemployment.

What it says about India

The Stanford paper covers only the United States. The closest Indian document is NITI Aayog’s Roadmap for Job Creation in the AI Economy, released on 10 October 2025 with NASSCOM and the Boston Consulting Group, and it models scenarios without measuring payrolls. In its business-as-usual case, India’s tech workforce of 7.5 to 8 million in 2023 falls to 6 million by 2031, and the customer-service workforce of 2 to 2.5 million falls to 1.8 million. If the country acts on the roadmap’s recommendations, the tech figure rises to 10 million and the customer-service figure to 3.1 million. The report expects entry-level quality-assurance and support roles in IT services companies to shrink, the same rung the Stanford paper tracks in the US. India is also building AI models in its own languages.

The Stanford and ADP team publishes occupation-level results monthly on its public dashboard, which was last updated on 23 September 2026.