The study Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence by Brynjolfsson, Chandar and Chen (August 2025) provides an extensive look into how the rapid adoption of generative AI is reshaping the American labor market, especially for young and entry-level employees. Drawing on high-frequency payroll records from millions of workers across thousands of firms, the researchers offer a near real-time perspective on recent labor market developments and how exposure to AI is correlated with employment trends.
One of the central findings is that entry-level workers ages 22 to 25 in the most AI-exposed occupations, such as software developers and customer service representatives, have suffered substantial declines in employment since late 2022 even after adjusting for firm shocks and various alternative explanations. The decline is measured at about 13 percent relative to their older peers or to young workers in less exposed jobs. In contrast, older workers in the same occupations and workers of all ages in less-exposed fields have seen their employment remain steady or even grow. This suggests a strong age-based disparity where it is primarily early-career talent bearing the brunt of new technology’s disruptive effects.
Overall employment in the United States during the study period continued to grow, signaling a resilient economy, but that topline number masks stagnation and even decline among the youngest cohort in AI-exposed roles. Early-career workers in the highest exposure categories experienced employment losses while those in the middle or lowest exposure groups experienced steady gains. Underneath these trends, the study reveals a crucial distinction: not all uses of AI yield the same consequences. When AI tools are mainly automating tasks by taking over work previously done by humans, young workers face job losses. But when AI provides augmentation by helping employees do more rather than replacing them, job prospects for entry-level workers are much stronger or even improved.
The research utilizes two major occupational exposure methodologies. The first, based on work by Eloundou et al., employs a measure of how much AI, specifically large language models, can substitute for tasks in each occupation. The second approach uses data from millions of real-world conversations with Claude, Anthropic’s generative AI, to estimate the share of tasks per occupation being automated or augmented. These methods together demonstrate a robust pattern: automation-heavy AI applications drive declines for entry-level workers while augmentation-heavy roles show little decline or even some gains.
These employment trends persist even after controlling for firm-level shocks, sectoral trends, overall industry health or changes following the COVID-19 pandemic. The declines are not limited to technology or information sectors and also appear in both remote-eligible and on-site roles. The study further demonstrates that the impact is seen more acutely in headcount rather than in average annual compensation. Salaries have not fallen sharply, potentially due to stickiness in pay or offsetting forces such as greater productivity requirements or wage adjustment lagging headcount changes. The authors argue that these findings clarify widespread confusion in public debate where aggregate employment trends appear healthy, yet striking changes are underway for specific segments of the workforce.
These facts are consistent across different analytical samples, excluding technology firms, stripping out part-time and temporary staff, accounting for educational background and disaggregating by gender. This consistency highlights the resilience of the pattern. The study counters the idea that declines among entry-level workers might be explained purely by COVID-driven changes, educational disruptions or transient shocks to tech hiring. Evidence also suggests that non-college workers in lower-exposure occupations do not experience the same negative consequences, whereas those in high-exposure non-college roles see trends similar to college-educated peers but less insulation from experience.
The mechanism suggested by the authors is that AI is more effective at replacing tasks based on codified knowledge, the type of work often assigned to new graduates or early-career employees, and has a much harder time replacing labor that depends on tacit knowledge or experience-based nuance, which is more prevalent among seasoned workers. As AI increasingly masters standardized tasks but struggles with the complex and often ambiguous challenges handled by experienced staff, the result is that fresh entrants to the workforce in exposed roles become especially vulnerable.
For legal professionals and policymakers, the study’s findings have significant implications for anti-discrimination law, employment law, workforce planning and the design of future regulation around AI-driven workplace transformation. Real-time payroll data such as those used in this study provide a clearer and more nuanced picture of occupational risk than traditional labor surveys and highlight the urgent need to monitor adjustment pathways for displaced entry-level talent. The adjustment period that typically follows a major technological revolution may be especially wrenching for younger workers, making careful attention to reskilling, retraining and the design of new entry pathways critical.
This research stands as early and large-scale evidence that the generational impacts of AI are not simply speculative but are already appearing in payroll systems nationwide. It suggests that future research should focus on even finer-grained firm-level AI adoption data, longer-term wage effects and the development of concrete policy responses to prevent the hollowing out of the workforce’s youngest tier as AI becomes ever more integral to business operations.