Young workers just starting their careers in jobs most exposed to artificial intelligence are losing ground at a measurable clip. Employment among 22-to-25-year-olds in highly AI-exposed occupations is shrinking about 4% a year, according to data from Stanford University’s Digital Economy Lab and ADP Research. The finding, drawn from payroll records mapped to federal occupation codes and scored by task-level AI exposure, raises a pointed question: are large language models already replacing the entry-level cognitive work that once served as a career on-ramp for recent graduates?
Why the 22-to-25 age band is the sharpest warning signal
The contraction hits hardest at the point where workers have the least bargaining power. People aged 22 to 25 are typically in their first or second professional role, still building tenure, industry contacts, and specialized skills. A sustained annual decline of that size compounds quickly. Over a five-year stretch, it would erase roughly a fifth of the positions available to that cohort in the affected job categories, narrowing the pipeline through which early-career workers gain the experience employers later reward with higher pay.
The Stanford Digital Economy Lab tracks these shifts through its AI indicators dashboard, built in collaboration with ADP Research. The project treats early-career cohorts as leading indicators, the labor-market equivalent of canaries in a coal mine. Because younger workers concentrate in roles heavy on routine cognitive tasks, they are the first group whose employment levels register the effects of AI adoption. Older workers in the same occupation codes have not shown the same rate of decline, suggesting the pressure is not simply a broad hiring slowdown but something specific to the tasks that junior employees perform.
One way to test whether the drop reflects genuine task substitution rather than normal cyclical softness is to compare month-to-month job-posting rates for AI-exposed occupation codes against non-exposed codes within the same age band. If postings for AI-exposed roles are falling faster while non-exposed postings hold steady or grow, the substitution explanation gains strength. Publicly available job-board data could support that comparison, though no published analysis has yet isolated the effect at that level of granularity for this specific cohort.
Another check would be to examine wage trajectories inside the same occupations. If employers are quietly automating away lower-level tasks while retaining more senior staff, average wages in those occupations could rise even as headcounts for younger workers fall. That pattern would be consistent with a move toward smaller, more experienced teams supported by AI tools rather than a broad collapse in demand for the occupation itself.
How researchers score occupations for AI exposure
The 4% figure rests on a layered methodology. Researchers start with the Standard Occupational Classification system, which assigns every U.S. job a numeric code tied to detailed task descriptions maintained in the O*NET database. Each occupation’s tasks are then scored for alignment with the capabilities of large language models.
That scoring approach traces back to a 2023 paper by Tyna Eloundou, Sam Manning, Pamela Mishkin, and Daniel Rock, available as arXiv preprint 2303.10130. The paper defined exposure as the degree to which an LLM could reduce the time needed to complete a given task by at least half, then aggregated those task-level scores up to full occupations. Roles heavy on writing, data analysis, and routine information synthesis scored highest, while hands-on or physically intensive work tended to show lower exposure.
Anthropic refined the framework in a separate study titled “Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations,” released as an arXiv study. Rather than relying solely on expert ratings, the Anthropic team analyzed real usage logs to map actual LLM interactions to O*NET tasks and occupations. The study distinguished between conversations that augmented a worker’s output and those that automated a task outright. Automation-heavy interactions became the key risk factor in the Stanford dashboard’s exposure scoring, because they point to tasks that employers might eventually assign to AI systems instead of entry-level staff.
By combining payroll counts from ADP with these exposure scores, the Stanford team can track employment trends for specific age bands within high-exposure occupation groups. The result is a near-real-time signal that traditional government surveys, which report with longer lags, do not yet capture at the same resolution. For employers, that signal can serve as an early warning about where talent pipelines may be thinning; for policymakers, it offers a way to spot where training or transition support might be needed before dislocation shows up in headline unemployment rates.
What the numbers might mean for early-career workers
If AI is eroding the number of entry-level roles in exposed occupations, the long-run consequences could extend well beyond a single age cohort. Many midcareer professionals in fields like marketing, finance, and law started in junior analyst or assistant positions that involved exactly the kinds of routine information processing now vulnerable to automation. Fewer such roles today could mean fewer workers with the experience needed to move into more complex responsibilities a decade from now.
Some young workers may already be adapting. Graduates who once would have aimed for traditional office roles may be shifting toward fields with lower measured AI exposure, such as healthcare support, skilled trades, or in-person services. Others may be trying to ride the wave by learning to use AI tools as force multipliers, positioning themselves as “AI-native” hires who can oversee or complement automated workflows rather than compete directly with them.
Education and training systems face their own adjustment challenge. If the tasks that used to justify a first job in an AI-exposed occupation can now be done by software, internships, project-based learning, and simulated work environments may need to shoulder more of the burden of giving students practical experience. Colleges and bootcamps could respond by weaving AI tool use and prompt design into standard curricula, while employers experiment with new forms of apprenticeship that focus less on rote tasks and more on judgment, communication, and domain expertise.
Gaps in the data and what to watch next
Several pieces of the puzzle are still missing. The dashboard does not publish downloadable quarterly employment counts broken out by single-year age band and AI-exposure quintile, making independent replication difficult. No firm-level or individual longitudinal records link specific hires coded to particular SOC occupations with documented Claude or GPT task usage inside those firms. That means the connection between the payroll decline and actual AI deployment at the employer level remains inferred from occupation-level patterns rather than confirmed by direct observation.
The U.S. labor statistics series confirm broad age-band employment trends but add little detail on which exact job titles within high-exposure categories are disappearing fastest. No public statement from BLS or Department of Labor statisticians has addressed how the 4% annual figure aligns with their own Occupational Employment and Wage Statistics microdata. Without that cross-check, the possibility that compositional shifts, such as young workers migrating to gig or contract roles not captured in standard payroll data, account for part of the decline cannot be ruled out.
Researchers will be watching several indicators over the next few years. If AI-exposed occupations continue to shed young workers while overall employment in those occupations holds steady, it would strengthen the case that employers are leaning on automation for lower-level work while preserving more senior roles. If, instead, both young and older workers in those jobs start to see similar declines, the story may shift toward broader restructuring or industry-specific downturns rather than targeted entry-level displacement.
For now, the 22-to-25 age band serves as an early, if imperfect, signal that the diffusion of large language models is beginning to reshape the bottom rungs of the career ladder. Whether that shift ultimately produces a more productive economy with new kinds of opportunity, or a more polarized labor market with fewer stepping-stone jobs, will depend on how quickly institutions, employers, and workers themselves respond to what the canaries are already telling them.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
