New analysis from the “Canaries in the Coal Mine” research team finds that jobs most exposed to artificial intelligence are not the same ones most sensitive to monetary policy — and the employment decline is accelerating.
The debate over what is driving the steep decline in entry-level hiring across the American economy has sharpened in recent months. Critics of a landmark Stanford University study have pointed to the Federal Reserve’s aggressive interest rate increases as a more plausible explanation for why young workers are losing ground. Now, the authors of that study are firing back with new evidence they say undercuts the interest rate hypothesis — while acknowledging the picture is more nuanced than either side has suggested.
In a February 9 research note published by the Stanford Digital Economy Lab, economists Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen addressed two questions that have dogged their widely cited “Canaries in the Coal Mine” paper since its initial release: whether rising interest rates better explain the employment patterns they observed, and whether the timing of the decline is actually consistent with AI exposure.
Their answer on interest rates is unequivocal. Using occupational data from Zens et al. (2020), the researchers found that jobs with the highest exposure to artificial intelligence are actually among the least sensitive to interest rate fluctuations. Construction occupations, for instance, carry high interest rate exposure but low AI exposure, according to the Stanford team’s analysis. The negative correlation between the two measures, the authors argued, means that monetary tightening is a poor candidate for explaining the disproportionate decline in hiring for AI-exposed roles.
The researchers went a step further, splitting occupations into groups based on their sensitivity to interest rates and then tracking employment trajectories within each group. In both high and low interest rate sensitivity categories, occupations with the greatest AI exposure still showed steeper employment declines than average, according to the February note. The pattern held regardless of how exposed a given occupation was to the broader monetary environment.
The timing question, however, yielded a more complex answer. Brynjolfsson, Chandar, and Chen acknowledged that when they applied their most rigorous statistical controls — specifically, firm-time fixed effects that account for company-level economic shocks — the statistically significant employment decline in AI-exposed occupations did not emerge until 2024, rather than late 2022 or 2023 as earlier, less controlled analyses suggested.
This is a meaningful concession. It implies that some portion of the early decline in entry-level employment, which coincided roughly with the launch of ChatGPT in November 2022, was likely driven by other macroeconomic forces — possibly including, but not limited to, interest rate movements, post-pandemic labor adjustments, and broader hiring freezes across the technology sector.
Yet the researchers stressed that the concession does not weaken their central thesis. Once those confounding factors are stripped away, the AI-specific employment effect becomes statistically significant beginning in 2024 and has continued to accelerate. Updated data through October 2025 now shows approximately a 16 percent relative decline in employment for early-career workers aged 22 to 25 in the most AI-exposed occupations, up from the 13 percent figure reported in the team’s original paper, which used data only through July 2025, according to the Stanford note.
The underlying dataset comes from ADP, the largest payroll processing firm in the United States, which provides individual-level monthly payroll records covering millions of workers across tens of thousands of private companies. That granularity has given the Stanford team a window into labor market dynamics that public surveys cannot match. As co-author Bharat Chandar noted in an August 2025 primer on the research, publicly available datasets like the Current Population Survey sometimes contain as few as 26 observations per month for narrow demographic-occupation groups such as young software developers, making reliable trend analysis nearly impossible.
The broader landscape of research on AI and employment has grown considerably since the Stanford team first published its findings. The February note referenced more than a dozen concurrent studies — from researchers at institutions including Yale, the University of Chicago, and the World Bank — that are exploring the same questions using different data and methodologies. The authors explicitly encouraged readers to consult that wider body of work, acknowledging that no single study can definitively isolate AI’s causal impact on employment in the absence of a controlled experiment.
Still, the direction of the evidence is consistent across multiple research efforts. Entry-level hiring at the 15 largest technology firms fell 25 percent between 2023 and 2024, according to a report from venture capital firm SignalFire cited by IEEE Spectrum in December 2025. A separate SignalFire analysis found a 50 percent decline in new role starts by workers with less than one year of post-graduate experience at major tech firms and maturing startups between 2019 and 2024, as reported by CNBC in September 2025. And Challenger, Gray and Christmas tallied nearly 55,000 job cuts explicitly attributed to AI in 2025 alone, out of a total 1.17 million layoffs that year — the highest figure since the pandemic.
The Stanford team was careful to note that their findings do not support a narrative of AI-driven mass displacement across the entire economy. Overall employment continues to grow. Experienced workers in AI-exposed occupations have not seen comparable declines. And in occupations where AI augments rather than automates human labor, employment has remained stable or even increased across all age groups, according to the original “Canaries” paper.
The researchers also cautioned against reading any single factor as the sole driver. As they wrote in the February note, they do not believe AI is always and everywhere the sole determinant of employment outcomes, and they urged others not to interpret their findings that way. The sharp decline among young software developers, for instance, is likely the result of multiple forces converging simultaneously.
What the data does show, however, is that the employment gap between the most and least AI-exposed occupations is widening, not narrowing. The trend has not reversed. And the workers bearing the brunt of it are precisely the ones with the least experience and the fewest professional networks to fall back on — the youngest entrants to the American labor force.
Whether this represents the early tremors of a broader transformation or a temporary dislocation that will self-correct as the technology matures remains an open question. The Stanford team has committed to monitoring the data on an ongoing basis. Given the stakes, the rest of the research community — and the policymakers who rely on it — would be well advised to do the same.
