Blog · Employment and Evidence · October 2, 2026

What the Entry-Level Jobs Data Can—and Cannot—Tell Us About AI

A narrowing entrance to working life deserves attention even when nobody is being fired. Understanding AI's role requires separating what software could do, where employers actually use it, and why hiring changes.

The missing opening

Consider a hypothetical department that keeps its experienced staff but cancels its next junior vacancy. Its manager says AI helps the team handle routine assignments. The company also has fewer customers than expected. No employee receives a dismissal notice. A graduate nevertheless loses an opportunity that would otherwise have existed—if the manager would really have hired without the software.

That final condition is the difficult part. The canceled vacancy is observable; the alternative history is not. Our argument is that evidence of a worsening entrance to employment deserves serious investigation without turning every weak hiring number into a measured effect of AI. The research question is narrower, and more useful, than whether machines are taking everybody's jobs.

Four questions behind one headline

We use four distinctions throughout this essay. Exposure asks whether a technology could perform tasks associated with an occupation. Adoption asks whether people actually use it in production. Employment records who has work. Causal displacement asks how employment differs because of the technology, compared with what would otherwise have happened. Evidence about one does not automatically answer the others.

The ILO's May 2025 refined exposure index combines task assessments, worker and expert input, and model predictions. It estimates that one in four workers worldwide occupies a job with some generative-AI exposure. That is a classification of technological potential, not an observed fraction losing work. The authors judge transformation more likely than complete replacement because occupations retain tasks requiring human input.

Our inference is that an exposure score should direct observation: inspect how the work changes in those occupations. It cannot supply a hiring forecast by itself. A firm might use faster drafting to cut staffing, serve more customers, improve an existing service, or spend the saved time checking mistakes. Those are different organizational choices even when the underlying software is identical.

The Bureau of Labor Statistics' JOLTS definitions also separate employment, openings, hires, and separations. Hires are additions to payroll; separations include voluntary departures as well as employer-initiated losses. An opening requires an available position and active external recruiting.

This distinction matters for the hypothetical department. A vacancy announcement, an actual hire, and a retained employee answer different questions. A stable total can conceal fewer opportunities for newcomers if existing staff stay longer. Conversely, a falling total cannot tell us whether people were dismissed, left voluntarily, or were simply not replaced.

Read the current canaries carefully

The August 2026 revision of Brynjolfsson, Chandar, and Chen's Canaries in the Coal Mine? examines an ADP payroll panel through June 2026. It reports a 19% employment shortfall for ages 22–25 in more-exposed occupations relative to keeping pace with less-exposed peers, primarily through weaker hiring. It finds no widespread economy-wide displacement. The analysis covers a balanced panel of firms, with 3.5–5 million employees monthly, rather than every ADP client. Its headline is descriptive: education controls reduce the pattern, some divergence predates generative AI, and national survey comparisons show weaker differences. Excluding technology firms and computer occupations does not eliminate the pattern.

For interpretation, the comparison is essential. A shortfall relative to another group's trajectory is not a count of individuals replaced by software. Nor does an age bracket identify everyone's career stage: an older career changer can be a beginner, while a younger worker can already have experience. The finding asks us to investigate access to work; it cannot label each missing job's cause.

Our reading gives the result real weight without treating statistical adjustment as an all-purpose solvent. If exposed and less-exposed occupations were already moving differently, an apparent later gap may combine several processes. Equally, listing possible confounders does not establish that those confounders explain the result. Both the AI explanation and its alternatives need evidence with timing, a mechanism, and a credible comparison.

The comparison that complicates the story

Humlum and Vestergaard's March 2026 Still Waters, Rapid Currents links Danish adoption surveys to administrative records through December 2024. Its latest survey includes 25,000 workers across 7,000 workplaces in 11 exposed occupations. The authors estimate no detectable average effect on earnings or recorded hours, ruling out effects larger than 2% in their comparisons. They observe declining early-career employment in exposed occupations, but their adoption-based analysis does not attribute that decline to chatbot-adopting firms. Workers nevertheless report new tasks and reorganized work.

We see this as a substantive challenge to a simple inference from exposure to displacement. It also cannot settle the later American result: the country, observation window, adoption measures, and comparisons differ. These are working papers answering related questions, not two instruments measuring an identical object. The appropriate response is to ask which conditions would make their results agree.

What would distinguish the explanations?

The following is our proposed research agenda, not an account of findings already established. Start with dated adoption records linked to teams and tasks. A purchased license is weak evidence of changed production; regular use on assignments previously given to junior staff is stronger. Compare hiring before and after that change with similar teams whose demand and prior hiring trends resemble those of the adopters.

For the substitution explanation, look for fewer junior hires after the relevant workflow changes, alongside maintained output and no comparable reduction in customer demand. For the demand explanation, look for falling orders and hiring cuts across tasks regardless of AI use. For a reorganization explanation, look for changed job descriptions, new checking responsibilities, and movement between roles. These patterns would be suggestive; none alone would remove the problem that firms choose when to adopt.

A more persuasive design would exploit rollout timing unrelated to a team's hiring plans, where that independence can be defended. It would also inspect what happens outside the adopting firm. If cheaper services expand demand elsewhere, one employer's reduced headcount does not measure the total employment effect. If unsuccessful applicants move into worse jobs, the absence of unemployment does not measure the full cost either.

Follow entrants separately from incumbents. Track experience requirements, recruitment, actual starts, retention, earnings, and movement between occupations. Distinguish age from time in a profession. Publish sample coverage and revisions alongside estimates. These choices make it harder for an attractive headline to outrun the population and outcome it actually describes.

Keep the entrance to work visible

Our practical conclusion is to keep two questions open at once: is AI changing the amount of labor employers want, and is it changing who gets a chance to become experienced? An organization could improve current output while weakening the route through which future skilled colleagues learn. That possibility merits measurement even before a convincing economy-wide displacement estimate exists.

For employers, a useful starting point is to document why junior vacancies are opened, changed, or canceled and what training replaces the assignments being automated. For researchers and readers, the discipline is to preserve the chain from task potential to actual use to hiring behavior to causal explanation. Each link needs evidence. The people waiting to enter work deserve an account precise enough to guide a response.

Sources

Primary sources consulted October 2, 2026. Study observation periods are stated above; publication dates do not extend the underlying data.

Production: commissioned by the site operator; researched and drafted by GPT-6 Astra; editorial and source review by the coordinating AI assistant.

Return to Blog