“You’re (not) Hired”: AI and Early-Career Hiring

What does “You’re (not) Hired” actually establish?

The April 2026 U.S. Census Bureau working paper “You’re (not) Hired: Artificial Intelligence and Early Career Hiring in the Quarterly Workforce Indicators” reports a 12% regression-adjusted decline in employment among workers aged 22–24 in the highest AI-exposure quintile of industry-state groups over the ten quarters after ChatGPT’s release. The paper identifies reduced hiring as the main proximate channel behind the later employment decline. This is a warning about early-career access in those groups. It is an association in observational data, however: exposure is assigned at group level, and the study does not link an individual worker’s AI use or employer’s adoption to that worker’s employment outcome. It does not establish that AI alone caused the decline or predict whether a particular applicant will be hired. The distinction matters because “exposure” describes the kinds of tasks concentrated in a group, not a record of which firms deployed a tool or how their work changed. The paper uses matched employer-employee administrative records and compares early-career employment across industry-state cells with different measured exposure. Its abstract also reports that hiring rates largely recovered by early 2025 against a smaller employment base. A rate can rise while the number of jobs and entrants remains lower; neither measure is a personal displacement probability. The paper is a Center for Economic Studies working paper, dated April 2026, not a final Census Bureau finding. It reports earlier trend shifts around the COVID pandemic and discusses remote work and rising educational attainment among workers in exposed occupations as possible complications. These qualifications complicate interpretation of timing, but do not make the result irrelevant. The useful conclusion is narrower than either “AI has already eliminated these jobs” or “nothing has changed”: the study documents a serious group-level early-career hiring signal whose cause is not settled. For someone weighing an entry path, it is reason to examine openings and requirements in reachable markets, not a personal forecast or an instruction to abandon a field.

Sources: CREAT paper profile: You’re (not) Hired: Artificial Intelligence and Early Career Hiring in the Quarterly Workforce Indicators; You’re (not) Hired: Artificial Intelligence and Early Career Hiring in the Quarterly Workforce Indicators

What does the Census paper actually measure?

“You’re (not) Hired: Artificial Intelligence and Early Career Hiring in the Quarterly Workforce Indicators” combines linked employer–employee administrative records with an estimate of how exposed an industry’s typical mix of jobs may be to AI. It compares employment and hiring among U.S. workers aged 22–24 across industry-state groups. It does not observe an individual worker using AI or link a named employer’s adoption decision to that worker’s outcome. The 12% result is a regression-adjusted decline in employment in the highest-exposure quintile over the ten quarters after ChatGPT’s release, relative to the paper’s modeled comparison. The paper separately tracks hires and identifies reduced hiring as the main proximate source of the decline. This is evidence about a group-level pattern, not a 12% chance that a particular worker will be replaced.

The unit matters.“More Information about the QWI” explains that Quarterly Workforce Indicators tabulate jobs and workforce flows. The paper uses industry-by-state cells, rather than occupation-by-state cells, because QWI can link workers to employers and report age but lacks the occupation detail used here. It sorts these cells into five exposure quintiles, weighted by all-age employment in 2022; the top quintile is the fifth of industry-state groups with the highest assigned average exposure. It does not mean every young worker there performs highly exposed tasks. The analysis follows ages 22–24 through 2025 Q2,. QWI employment is indexed to 2022 Q4, while the adjusted result compares changes across exposure groups. It is distinct from the raw change in the top group.

To construct exposure, the study starts with an occupation-level task measure and maps it through American Community Survey microdata into employment-weighted industry averages, then QWI industry-state cells. The underlying measure assesses whether large language models might reduce time needed; it is not a record of tools used at work. Aggregation allows comparison with employer workforce outcomes but erases within-industry variation: two people in one cell may have different occupations, duties, and exposure. Quintiles simplify comparisons but hide differences within each group. The study’s strength in linking worker demographics to employer outcomes therefore comes with less precision about each worker’s tasks.

The paper checks whether this proxy corresponds to observed adoption. Comparing exposure with Census Business Trends and Outlook Survey (BTOS) responses, it reports that a one-standard-deviation increase in exposure is associated with 6.7 percentage points higher reported AI adoption; it predicts about 47% of observed adoption variation in April 2026. This supports exposure as an indicator of where adoption is more common. It does not show adoption caused the employment result: BTOS responses are not linked to workers or firms in the outcome analysis, and some subsector responses are suppressed. The paper also notes BTOS changed its adoption questions in late 2025. This is aggregate construct validation, not a treatment record for QWI workers.

Timing also limits attribution. The working paper reports an immediate hiring decline, but its triple-difference analysis finds evidence of earlier shifts around COVID in employment, hiring, and separations. The author discusses remote-work growth and increased educational attainment among workers in more exposed occupations as possible explanations. These patterns complicate a simple before-and-after account: industries may differ for reasons that predate generative AI. The author says the abrupt hiring discontinuity remains notable, while longer-run changes are difficult to separate from alternatives. It is a Center for Economic Studies working paper, not a final Census Bureau finding; its opening pages state CES papers have not undergone the review accorded Census publications.

Read the estimate as a signal about employment and hiring among groups assigned higher average exposure,. It can justify closer attention to entry routes in a target market. It cannot show that all workers in those groups face the same task changes, that their employers adopted AI, or that AI alone produced the observed difference. “You’re (not) Hired” provides linked outcomes and an aggregate adoption check, but neither turns predicted applicability into observed use. Keep that distinction beside the headline when judging a particular role.

Sources: You’re (not) Hired: Artificial Intelligence and Early Career Hiring in the Quarterly Workforce Indicators; CREAT paper profile: You’re (not) Hired: Artificial Intelligence and Early Career Hiring in the Quarterly Workforce Indicators

Why can hiring weaken before a workforce visibly shrinks?

Hiring is a flow into work; employment is the stock of jobs present at a point in time. An employer can take on fewer newcomers while keeping most current staff, so weaker hiring may show before separations rise or total employment falls sharply. For early-career workers, that can mean fewer routes into a field even while experienced colleagues remain. The Census paper reports weaker hiring among 22–24-year-olds in more exposed industry-state groups and identifies reduced hiring as the main proximate explanation for the subsequent employment decline. That is a group-level finding, not a count of people rejected or proof that a newly automated task explains each missing hire. The Quarterly Workforce Indicators (QWI) distinguish these flows. Census defines a new hire as an accession who was not employed by the same employer during the previous four quarters. A separation is recorded when someone employed by a business in one quarter is absent from it the next. This measure does not distinguish a voluntary departure, such as changing jobs, from an involuntary exit. “More Information about the Quarterly Workforce Indicators” defines both measures. Employment counts also depend on the data frame. “QWI Measure Descriptions” explains that the indicators link workers with employers through quarterly administrative records and count jobs, rather than employed people as household surveys such as the Current Population Survey (CPS) do. A person can hold multiple jobs, and self-employment and independent contracting are outside its ordinary frame, so totals are not exactly comparable with CPS. The flow-stock distinction has a direct implication for opportunity. When hiring slows but separations do not rise, current staff can remain in place while fewer newcomers enter. In its counterfactual exercise, the Census paper estimates that maintaining earlier hiring would have left early-career employment near its pre-release level in early 2025; lower separations offset part of the observed decline. This is a modeled comparison of group paths, not a list of people denied jobs or proof that AI caused each decision. The paper cautions against attributing the whole pattern to AI. The result nevertheless makes access worth examining alongside layoffs. A further career-ladder effect is plausible, but remains an interpretation. Entry roles may include repeatable assignments with feedback, such as preparing drafts or checking records. If employers assign fewer such tasks to beginners, they might retain experienced staff while offering fewer chances to practise under supervision. Less practice could affect how newcomers build judgment. The Census study does not measure mentorship, learning-by-doing, entry-level task content, or whether a career ladder has weakened. Its hiring result raises a question to investigate; it does not establish that later chain of effects. So “Were people laid off?” is only one question for someone considering a field. Also ask whether employers still recruit for entry roles, whether those jobs include work that lets beginners demonstrate competence, and whether responsibility grows through review and feedback. A national industry-state association cannot answer for one employer or city. The QWI guide notes that hiring often reflects worker turnover rather than net employment growth, so a flat or shrinking headcount alone can miss replacement openings. The practical conclusion is that fewer hires can restrict access before incumbent employment visibly shrinks, but this finding does not mean every exposed occupation has stopped hiring or that a reader should leave a field. Check recurring entry openings, stated requirements, and opportunities to take on reviewed responsibility in the market where you can work. Treat concerns about lost practice or mentorship as questions to verify locally, not outcomes established by the Census paper.

Sources: You’re (not) Hired: Artificial Intelligence and Early Career Hiring in the Quarterly Workforce Indicators; More Information about the QWI; QWI Measure Descriptions

How much does the later payroll evidence add?

Stanford’s revised August 2026 analysis adds a later signal that early-career outcomes in exposed work deserve attention. It does not verify the Census estimate by repeating the same design. “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence” describes high-frequency ADP payroll records covering millions of U.S. workers through June 2026. In that sample, employment for workers ages 22–25 in AI-exposed occupations stood 19% below the path it would have followed if it had kept pace with less-exposed peers. This is a relative gap, not a count of jobs eliminated, an absolute fall in all young-worker employment, or an individual’s chance of losing work. Stanford reports no comparable gap for experienced workers and no evidence of widespread economy-wide displacement. The Stanford authors say the pattern operates primarily through reduced hiring of younger workers, rather than increased separations. That matters because weaker entry can appear while existing employees remain on payroll. But the studies are not interchangeable. Census follows workers ages 22–24 in industry-state cells, assigning exposure through the industries’ occupational mix and measuring employment and hiring in Quarterly Workforce Indicators. Stanford analyzes ages 22–25 by occupation in ADP payroll data and compares their employment trajectory with less-exposed peers. Their units, classifications, time windows, data coverage, and comparison paths differ. Similar direction makes the question harder to dismiss, but it is not two independent experiments isolating AI’s effect. The revised Stanford summary says the divergence widened from the first version and that declines concentrate in occupations where AI use primarily substitutes for human tasks. Where use primarily complements workers, employment is flat or rising, especially for experienced workers. The results do not establish that every exposed occupation is shrinking or that complementarity guarantees growth. The authors call the results early descriptive indicators, not causal estimates. They report that some divergence predates generative AI, patterns attenuate when controlling for education, and results are more pronounced in the ADP sample than in national survey benchmarks. They also report checks excluding technology firms and computer occupations, and controls for interest-rate exposure and remote work. These checks address alternatives but do not erase the identification limits. A gap that predates widespread generative AI complicates attributing the full later difference to the technology. Extending the observation period gives a longer view; more observations do not by themselves turn a descriptive comparison into proof of cause. Read alongside “You’re (not) Hired,” the update raises the practical salience of early-career access while leaving its source unresolved. Both analyses report weaker relative hiring or employment for younger workers in more-exposed work. Neither links a worker’s own AI use to that worker’s outcome. Exposure categories describe potential applicability at occupational or group level; they are not employer adoption measures. Reduced hiring alone does not show whether firms changed task design, faced different demand, or responded to other labor-market shifts. The defensible reading is that early-career outcomes merit monitoring, particularly where entry tasks are changing, while the studies cannot tell a reader whether a specific employer will hire fewer juniors. For someone considering a first role or early-career move, 19% is a reason to inspect entry routes, not abandon a field based on a national estimate. Check whether openings in the market you can reach still seek junior candidates, what tasks they assign, and whether requirements are changing. Compare those observations with the role’s workflow: which tasks are becoming assisted or bundled, and where review, context, or accountability remains part of the work. If openings and requirements remain viable, a focused upgrade may preserve options. If several concrete signals point the same way, compare adjacent roles against income needs, location, credentials, study time, and family or health constraints. The update makes that check more salient; it does not decide the career question for you.

Sources: Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence; You’re (not) Hired: Artificial Intelligence and Early Career Hiring in the Quarterly Workforce Indicators

Why do vacancy findings complicate the hiring story?

The New York Fed’s May 2026 analysis of Lightcast job postings complicates the claim that generative AI caused a new collapse in entry-level hiring. It finds that advertised vacancies in occupations with higher measured AI exposure were already declining relative to less-exposed occupations before ChatGPT’s late-2022 release. Researchers find no clear additional break afterward, and no clear relative decline in junior postings compared with senior postings inside high-exposure occupations. This does not erase the Census working paper’s finding of weaker hiring among 22–24-year-olds in highly exposed industry-state groups. It shows that evidence depends on which labor-market stage, unit, and period a study observes.

In Do Job Postings Show Early Labor-Market Effects of AI?, the New York Fed researchers combine Lightcast listings with an occupation-level exposure measure built from O*NET tasks and observed AI use, including whether tasks could theoretically be completed by AI and whether use looks more like automation than augmentation. Occupations scoring at least 0.2 are classified as high exposure; the authors report similar results with alternative cutoffs. This measures potential task exposure, not the number of employers that adopted AI or workers whose jobs changed.

The first event study tracks posting differences between higher- and lower-exposure occupations around ChatGPT’s release. The relative vacancy decline begins before 2022 and shows no clear further change afterward. Because the gap predates release, it weakens the claim ChatGPT initiated it. The difference stabilizes after 2023. They conclude: postings provide little indication of a distinct AI-driven decline in labor demand, and the pattern is difficult to attribute directly to AI adoption. This complicates a causal story; it does not establish that AI had no effect in any occupation or employer.

The junior-versus-senior comparison asks a narrower question. Within high-exposure occupations, Lightcast classifies postings by role level using information in each listing. Researchers compare the relative movement of those groups around late 2022. The series fluctuates without a clear trend, so the analysis does not find a distinct junior-posting decline relative to senior postings. This weighs against the claim that the slowdown was concentrated at the entry rung. But it concerns advertised vacancies and constructed role labels; it cannot show that every employer’s entry requirements or training assignments stayed constant.

Why can this coexist with the Census result? Census uses linked administrative job records, aggregates exposure to industry-state cells, and reports hires and employment for workers aged 22–24. The New York Fed uses postings grouped by occupation and role level. A listing is an advertised opening, not a completed hire; it may remain unfilled or disappear without a hire. These measures can move differently, and neither records why a particular opening or hire changed. Different age bands, exposure measures, and periods make these complementary views, not replications.

For someone weighing an early-career path, inspect current local postings and changes in entry requirements. Look for repeated signs: fewer openings for the role you can reach, duties shifting from junior work, or qualifications changing across employers. A national vacancy pattern cannot settle those local questions, just as a group-level exposure association cannot forecast one person’s prospects. The New York Fed findings weaken a simple AI-caused hiring-break explanation; Census findings still make reduced access worth watching. Together, they support attention to the entry route, not a verdict that the field is closed.

Sources: Do Job Postings Show Early Labor-Market Effects of AI?; You’re (not) Hired: Artificial Intelligence and Early Career Hiring in the Quarterly Workforce Indicators

What does an exposure measure say about AI and a job?

Exposure measures potential task applicability: whether technology could affect parts of work under stated assumptions. It does not establish employer adoption, reliable production performance, workflow change, falling demand, or worker displacement. Exposure is a prompt to inspect work, not an individual forecast.

Technical capability concerns what a system can do under specified conditions; exposure maps capability to tasks. Observed use means workers use a tool, while organizational adoption means an employer incorporates it into a process. Workflow redesign shifts responsibilities: tasks may be substituted, assisted, checked, or unchanged. Labor demand concerns how many workers employers seek or retain as output, costs, and choices shift. Displacement is an outcome, not a synonym for these earlier stages. A task may be exposed without adoption, and adoption may assist workers.

The International Labour Organization’s *Generative AI and Jobs: A 2025 Update* offers global context. Its refined assessment combines task data, expert input, and AI predictions across nearly 30,000 tasks, assigning occupations four exposure gradients. It estimates one in four workers worldwide is in an occupation with some exposure and concludes most jobs are more likely transformed than made redundant, given continuing need for human input. These global occupation estimates are not U.S. hiring or employer deployment data, nor individual job-loss risks. A role may persist while its tasks change.

Consider an illustrative reporting workflow, not a measured task ranking. A system might help draft a first summary from supplied figures. That possibility does not tell us whether a team permits the tool, whether its draft is accurate enough, or whether workload changes. Someone may still need to check source completeness, investigate an anomalous variance, choose an appropriate comparison, and explain a recommendation. Those steps involve context, verification, and accountability. An employer could preserve responsibilities while using assistance, redesign the process and move review elsewhere, or make no operational change. Capability alone cannot distinguish these outcomes.

The Census working paper *You’re (not) Hired: Artificial Intelligence and Early Career Hiring in the Quarterly Workforce Indicators* shows the measure’s value and limit. It crosswalks occupation-based task exposure into industry-state groups and compares employment and hiring patterns, then separately checks exposure against reported AI adoption. A one-standard-deviation increase in subsector exposure is associated with 6.7 percentage points higher observed adoption; the measure predicts about 47% of variation in April 2026. This supports an aggregate association with adoption patterns. It does not identify workers’ use or establish that adoption caused employment findings.

The adoption observations are not linked to the worker and employer records in the employment analysis, and some survey responses are unavailable or suppressed. A proxy can be informative without directly measuring treatment. The Census result therefore does not mean an exposed worker’s job was automated; an adoption response likewise cannot establish that a specific task was replaced. The group-level association is a reason to investigate mechanisms, not to collapse exposure, use, adoption, demand, and displacement into one claim.

For an early-career decision, inspect work and hiring in the market you can reach. Ask which tasks are changing, who checks outputs, where accountability sits, and whether entry openings still build experience. Review role descriptions, current openings, and credible accounts of implemented changes; treat exposure labels as prompts for questions. This cannot yield a personal replacement probability. It can help decide whether to continue exploring, test a focused upgrade, or compare an adjacent option against evidence and constraints.

Sources: Generative AI and Jobs: A 2025 Update; You’re (not) Hired: Artificial Intelligence and Early Career Hiring in the Quarterly Workforce Indicators

Which signals would change the decision about a target path?

A useful decision should respond to evidence about the entry route you can actually pursue, not to an exposure label alone. Treat three signals as a practical checklist, not a validated score or a numerical cutoff: whether relevant openings and their requirements are changing; whether the work itself is being reorganized; and whether organizations in the target market have moved from experimenting with AI to using it in a defined workflow. First, inspect advertised openings for the role, level, and geography that fit your circumstances. Compare entry-level titles with similar titles and note stated requirements, not headline counts alone. Note whether openings recur, whether a degree or credential is required, and whether employers now expect tasks that used to be learned on the job. Track changes across more than one posting cycle where possible, since a quiet week can reflect ordinary timing rather than a durable shift. Postings count advertisements, not completed hires, and may not describe daily work. The New York Fed’s “Do Job Postings Show Early Labor-Market Effects of AI?” illustrates why timing matters. Its Lightcast analysis found that vacancy declines for more AI-exposed occupations began before ChatGPT’s late-2022 release, with no clear extra break after it; within highly exposed occupations, junior and senior postings moved broadly in parallel. This complicates a simple AI-caused junior-hiring story but cannot establish whether a particular employer still opens beginner paths. Second, look for observable changes in the task bundle. Ask what a new hire would actually do in the first months: which routine drafts, summaries, searches, or data transformations are still assigned; what checking, exception handling, customer context, or explanation is now expected; and who is accountable when output is wrong. If a role’s entry work has been removed, combined with a more senior job, or changed so that verification requires knowledge beginners have not yet acquired, that is more directly relevant to your path than a general occupation exposure score. These are questions for current role descriptions, work samples, or informed conversations; the Census study did not measure mentorship or training. “You’re (not) Hired” reports fewer early-career hires in its most exposed industry-state groups and identifies reduced hiring as the main proximate channel. It does not show that first assignments disappeared in every role or that fewer hires mean a broken ladder. Third, seek evidence of adoption tied to the market and workflow you are considering. A public statement that an organization is exploring AI, a trial, or a worker’s occasional use is not the same as a deployed process that changes staffing or entry requirements. Look for the task handled, whether use is routine, what review remains, and whether duties or hiring criteria changed. Keep intention, implementation, and employment consequences separate. The Census paper’s exposure measure predicts some variation in subsector adoption, but its worker outcomes are not linked to adoption at each worker’s employer. Adoption is a local question, not a causal conclusion from the group result. Weak or mixed signals support monitoring and a low-cost upgrade, not an abrupt exit. Choose one recurring task, learn its AI workflow, and make a small sample showing judgment and verification. Ask for feedback from someone familiar with that work. If openings repeatedly contract, requirements shift, and workflow evidence points in the same direction, compare an adjacent role before making a larger change. Compare options against income needs, location, credentials, time, health, care responsibilities, and reversibility. A person who already has direct evidence of shrinking local openings or newly changed entry requirements may need to compare sooner. The practical question is whether evidence justifies a reversible experiment or a more serious alternative.

Sources: You’re (not) Hired: Artificial Intelligence and Early Career Hiring in the Quarterly Workforce Indicators; Do Job Postings Show Early Labor-Market Effects of AI?; QWI data

What is a proportionate next move?

Do not leave a field because of the Census working paper alone. Its group-level finding is a reason to inspect access to the work you want, not a forecast about your prospects. The ILO’s *Generative AI and Jobs: 2025 Update* also treats exposure as a signal about tasks that may change, not a count of people certain to lose work. Decide using evidence closer to your intended role and your actual constraints.

Name one target role and list its recurring tasks: what is produced, what needs checking, and where judgment or accountability sits. Then review a manageable set of current openings in places you could realistically work. Note entry requirements, repeated duties, and whether employers describe changed workflows or ask for experience once gained through junior assignments.

Pick one change to investigate. If a task is becoming tool-assisted, try a low-cost course, guided practice, or small work sample, then ask someone familiar with the role to assess it. Bound the experiment by income needs, location, credentials, time, health, and care responsibilities. Varied openings and transferable experience may favor an upgrade; repeated narrowing entry routes may justify comparing an adjacent role before a larger change.

The free task checker can organize that first review: “Check my task exposure.” It gives task-level change-pressure signals, not a validated probability of displacement. If you need a structured comparison of staying, adjacent options, and a larger change against your constraints, the paid roadmap offers scenario planning and a 30/60/90-day plan; it does not guarantee employment or income. Move when direct evidence and your circumstances justify it, not because one estimate appears to decide your future.

Sources: You’re (not) Hired: Artificial Intelligence and Early Career Hiring in the Quarterly Workforce Indicators; Generative AI and Jobs: A 2025 Update

Questions readers ask