"Canaries in the Coal Mine" — Evidence That AI Is Cutting Entry-Level Jobs, and the Case Against

Of the studies that tried to test "AI will take our jobs" against real employment data, the one that drew most attention was "Canaries in the Coal Mine?" by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen of the Stanford Digital Economy Lab. Their working paper, dated 26 August 2025, analysed payroll data from ADP, the largest US payroll software provider, and reported that since the widespread adoption of generative AI, workers aged 22–25 in the most AI-exposed occupations had experienced a 13% relative decline in employment, even after controlling for firm-level shocks.

In the same occupations, employment of more experienced workers was stable or growing, and young workers in less exposed occupations were not losing ground. The title asks whether, like canaries that sicken first in a mine, young workers are where AI's effects show up first. Other analyses, however, found no visible change across the economy as a whole, and by autumn 2026 the evidence remained contested.

What the paper found

The August 2025 version set out six facts. The main ones:

Occupational exposure was measured with indices including that of Eloundou et al., whose authors include OpenAI researchers. As to why young workers, the authors offered one possibility: AI replaces codified knowledge, the "book-learning" at the core of formal education, and is less able to replace the tacit knowledge that accumulates with experience; young workers supply relatively more of the former.

How the numbers changed with revisions

The authors kept updating the data.

The August 2026 revision switched the headline measure from the regression estimate to the simpler "kept pace" comparison. On that measure the gap was 15% with July 2025 data and widened to 19% by June 2026. The decline works mainly through less hiring of young workers rather than more separations. Overall employment in the ADP sample rose about 6% from November 2022 to June 2026, and even the most exposed occupations grew about 4%. On the point that jobs are not disappearing across the economy, the paper has been consistent.

Objections and limits

Questions came early. The most common was whether the sharp rise in interest rates from 2022, or the unwinding of pandemic-era over-hiring in tech, could explain the pattern.

In a February 2026 note the authors argued that while interest rates affect overall employment, they do not explain the disproportionate decline in entry-level hiring in AI-exposed occupations. They acknowledged, though, suggestive evidence that with the broadest set of controls the relationship with AI exposure starts in 2024 rather than late 2022 or 2023. The August 2026 version listed further limits: the pattern weakens when controlling for education, some divergent trends predate generative AI, and it is more pronounced in the ADP sample than in national survey benchmarks. It described the findings as early, descriptive indicators, "canaries in the coal mine", rather than causal estimates.

There is analysis on the other side. On 1 October 2025 the Budget Lab at Yale reported that the broader labour market had not experienced a discernible disruption in the 33 months since ChatGPT's release, undercutting fears that AI automation was eroding demand for cognitive labour across the economy. The occupational mix was changing somewhat faster than in the past, but the difference was not large and predated the widespread use of AI at work. Its update of 14 October 2026 still found that changes in the occupational mix did not clearly align with AI adoption and that measures of AI usage showed no connection to changes in employment or unemployment.

Other evidence

On the forecasting side, Anthropic chief executive Dario Amodei told Axios on 28 May 2025 that AI could wipe out half of all entry-level white-collar jobs and spike unemployment to 10–20% in the next one to five years. That is a prediction by the head of an AI developer, not a measurement. Companies' actual behaviour has not run in one direction either: IBM said in 2023 it would slow back-office hiring, yet in 2026 it was reported to be tripling US entry-level hiring.

What about Japan

As of October 2026 no study applying the same method to young workers in Japan was found. One reference point: a survey by the education and labour ministries found that 98.0% of university students graduating in March 2026 were employed as of 1 April, unchanged from the year before. That figure is the share of job-seeking graduates who found work, and says nothing about whether hiring is falling in AI-exposed occupations.

Why the paper matters

The "AI will take jobs" debate long revolved around estimates of technical feasibility such as Frey and Osborne's 47%. This paper was one of the first large-scale attempts to check the question against payroll records, and it put forward the hypothesis that any effect would show up first at the entrance to careers, as young people not hired rather than workers laid off. If entry-level work shrinks, so do the places where experience is gained; the same worry about where future managers will come from was the reason IBM's HR chief gave for hiring more young people.

The authors' own caution, that this is not a causal estimate, and the continuing work to separate AI from interest rates, education and other factors, are part of the story. The wider sequence is on the AI and society timeline.