What Wage and Employment Data Actually Show About AI and Work

What do wage and employment data actually show about AI and work?

The current evidence shows uneven adjustment, with a sharper warning for some people entering work than a broad claim that AI is eliminating jobs. In *Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors (CES-26-56)*, a September 2026 U.S. Census Bureau working paper reports that graduates from the most AI-exposed decile of college majors saw an adjusted five-percentage-point decline in initial employment likelihood and a 13% decline in initial full-quarter earnings. Initial employment means getting work at labor-market entry; full-quarter earnings are earnings recorded across a complete quarter. Those estimates materially strengthen the evidence of pressure at entry for this defined graduate cohort, qualifying any earlier reading based mainly on broad employment totals or average wages. They make early-career access a serious part of the current picture. They are regression-adjusted associations around ChatGPT's introduction, not proof that ChatGPT caused the changes, a measure of economy-wide displacement, or a forecast for any individual worker. Other age, wage, productivity, and employer-adoption measures answer different questions and need to be read on their own terms.

Sources: Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors (CES-26-56)

Which outcome is each labor-market number counting?

Before comparing findings, identify what the number counts. Employment is a stock: how many people hold jobs at a point or over a defined period. Hiring and separations are flows into and out of that stock. A headcount can stay level when fewer people are hired if fewer people also leave; stable employment therefore does not rule out a narrower route into work. Conversely, a hiring slowdown alone does not establish that current workers were dismissed. The comparison group and time window matter: a count for all workers can conceal a change concentrated among new entrants, while an age or occupation subgroup cannot automatically describe the whole labor market.

Earnings and wages also have different denominators. Initial full-quarter earnings asks how much a defined group earned during a full quarter near labor-market entry; it can reflect whether people found work as well as what employed people were paid. Hourly wages divide pay by hours, while annual earnings cover a year and may vary with hours, employment duration, and job changes. Wage growth compares pay across time for a specified population or unit. A fall in entry-cohort quarterly earnings is consequential, but it is not interchangeable with a decline in the hourly wage of continuously employed workers. Always ask whether the statistic includes people with no earnings and how it handles part-year work.

Productivity is output relative to inputs such as labor time. A worker completing a task faster is a task-level time saving; it becomes measured output only if usable work is produced and quality, rework, and other inputs are counted. Even higher firm output per hour does not specify how much workers are paid or how many workers the firm employs. Those are separate outcomes whose relationship depends on what is measured and over what period.

A projection is a modeled estimate of a future period, not an observation of what has already happened. An occupation-wide projection can incorporate many expected forces and does not isolate AI's causal contribution. It may inform a longer-range view, but it cannot settle whether hiring has changed this quarter in a particular place. Read later numbers by naming their numerator, denominator, comparison population, and period: these determine whether a result describes employment access, pay among workers, output, or an expected future. Several measures may jointly describe adjustment, but combining them into one displacement figure would erase what each can establish.

What does the new graduate evidence change?

The September 2026 U.S. Census Bureau working paper *Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors (CES-26-56)* finds a clear entry-stage warning in linked administrative records on college graduates. For graduates from the most AI-exposed decile of majors, the regression-adjusted likelihood of initial employment was five percentage points lower, while initial full-quarter earnings were 13% lower. These are two related signals about the first step into the labor market, but they count different outcomes. The employment estimate concerns whether graduates entered work; the earnings estimate concerns the amount recorded over a complete quarter. Together they show why an early-career problem can be easy to miss in overall employment totals: access to a first job and the earnings attached to that early entry can weaken in a defined group even if a broad headcount does not show widespread contraction. The result changes the article’s balance from a general reading of wages and employment toward a more specific concern about the terms on which some graduates begin work.

The paper’s earnings decomposition also clarifies what the 13% figure represents. Roughly half of the decline is attributed in the abstract to lower earnings within employing sectors, and roughly half to a shift toward lower-wage industry mix. The first component means that the difference is not explained solely by graduates sorting into different sectors: earnings were also lower within sectors where they were employed. The second means that where graduates worked mattered too, because their sector mix shifted toward lower-paying industries. This split links job access and pay without treating them as the same statistic. A graduate who does not find work affects an employment measure; among the earnings recorded for the studied population, both within-sector earnings differences and the industries entered contribute to the observed decline. The decomposition therefore points to more than one possible pressure point at entry, while leaving the paper unable to isolate a single mechanism as the cause.

The timing and trajectory are important to the interpretation. The working paper compares outcomes around ChatGPT’s late-2022 introduction and reports that the differences attenuate as graduates move beyond initial labor-market entry. Attenuation means the estimated gap becomes smaller with time since entry; it does not mean that every graduate catches up, that all early costs are repaid, or that the difference disappears for every major. The evidence is about a cohort’s transition into work and how its observed outcomes evolve, rather than a permanent ranking of majors. It makes entry conditions a consequential period to examine, especially for people whose first employment can shape later experience, but the administrative records do not turn that concern into an individual forecast.

The boundary around the finding remains substantial. CES-26-56 is a Census working paper, and its regression-adjusted comparisons are not an identified causal test showing that ChatGPT produced the changes. The observed timing after the tool’s introduction is not, by itself, a counterfactual: other conditions affecting graduate employment and earnings could also differ across cohorts or majors. Exposure is assigned at the major level, so it does not tell us which tasks an individual graduate performed, whether an employer adopted AI, or what happened inside a particular workplace. The paper supports taking early-career access seriously as part of the labor-market picture; it cannot establish that a particular graduate will struggle to find work, that a specific degree has become a poor choice, or that AI explains the whole measured gap. That distinction matters because a major is a broad preparation category, while the work a graduate actually does depends on role, sector, location, and employer decisions.

For the article’s question, this is a meaningful qualification rather than a replacement for every other labor-market measure. Broad wage averages can coexist with a less favorable start for graduates in highly exposed majors, and later attenuation can coexist with a real initial disadvantage. Neither fact cancels the other: the first describes where pressure may be concentrated, while the second describes how the measured difference changes after entry. Read together, they make early access and first-quarter earnings visible outcomes in their own right, not merely smaller versions of economy-wide job loss. The paper consequently raises a focused question about how new graduates enter employment under changing conditions. Answering whether that pattern reflects AI itself, local hiring practices, or other cohort changes requires evidence beyond this working paper’s adjusted associations.

Sources: Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors (CES-26-56)

Is the pressure showing up in hiring before layoffs?

The August 2026 revision of Stanford Digital Economy Lab’s *Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence* adds a broader payroll-record signal centered on younger workers and occupations. Using ADP administrative records covering millions of U.S. workers through June 2026, it reports that employment for workers aged 22–25 in AI-exposed occupations was 19% below a relative peer trajectory. The study says the divergence mainly reflects reduced hiring, rather than a corresponding rise in separations. That distinction changes what the gap suggests: employers may be bringing fewer young workers into certain roles even when the records do not show a wave of current employees being let go. A relative gap is a comparison with the trajectory of peers, not a count of 19% of young workers losing jobs.

This complements the Census graduate finding through a different lens. CES-26-56 groups people by college major and their transition into initial employment; Stanford groups payroll workers by age and occupational exposure. The Census comparison can include whether a graduate enters employment and initial earnings across a quarter. The Stanford signal follows employment patterns among payroll-record workers in occupations, making hiring and separations visible as components of the changing employment level. Their convergence is therefore limited but meaningful: both identify entry access for people early in their careers as a place where unfavorable differences appear. They do not measure the same cohort, define exposure the same way, or supply two replications of one estimate. Read together, they make it harder to treat a stable national headcount as a complete description of labor-market adjustment, because fewer openings for younger entrants can develop without a broad increase in separations.

Stanford’s result is also the strongest counterweight here to a reassuring national headcount headline. The publication reports no widespread economy-wide displacement, yet its subgroup analysis finds a substantial relative divergence concentrated among younger workers in exposed occupations. Both parts matter: the aggregate conclusion constrains claims of broad displacement, while the subgroup pattern shows that aggregate stability can coexist with a weaker route into some jobs. The paper describes the result as an early signal, not a causal estimate. Some divergent trends began before generative AI became widespread; the gap attenuates when education is controlled for, and it is stronger in ADP records than in national survey benchmarks. The study also reports variation between occupations where AI can substitute for tasks and those where use is complementary, which further cautions against treating exposure as one uniform labor-market condition.

The practical inference should stay at that level. These payroll patterns make hiring flows worth watching alongside separations and total employment, especially when assessing prospects for people seeking their first role or moving into a new occupation. They do not establish that generative AI caused the hiring divergence, that every exposed occupation is closing to younger workers, or that the observed relative gap translates into a fixed number of jobs lost. Nor can the age-by-occupation result tell an individual what their own employer will do. It strengthens the case that early-career access deserves attention in the evidence review; the actual balance between fewer hires, ordinary turnover, changing skill requirements, and other labor-market forces remains an empirical question in each setting.

Sources: Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence (revised August 2026)

Why do wage findings resist one headline?

No: the available wage evidence does not show that AI exposure lowers pay uniformly across occupations. The Dallas Fed analysis, “AI is simultaneously aiding and replacing workers, wage data suggest,” examines 205 U.S. occupations and relates occupational AI exposure to wage growth after ChatGPT's release. At the median experience premium, its estimated association is close to zero, and the reported uncertainty includes zero. That estimate is consistent with no detectable average relationship at that point in the model; it is not evidence that every occupation had unchanged wages, nor does it establish that exposure raised or lowered a typical worker's pay.

The analysis also asks whether the association changes with an occupation's experience premium: the wage difference associated with experience, used here as an occupation-level proxy for how strongly pay rewards experience. It is a descriptive measure, not a direct measure of an individual worker's tenure, skill, or ability to use AI. The Dallas Fed estimates vary across this dimension: at lower experience-premium values, greater exposure is associated with weaker wage growth; at higher values, the estimated relationship shifts in a more favorable direction. This interaction is why a single average can mislead. An estimate evaluated at the median summarizes one point in the relationship, while estimates at other points can differ. The variation does not mean that experience itself caused the difference or that accumulating years will protect a particular person.

This pattern can coexist with the entry-role concerns in the graduate and hiring evidence discussed earlier. If occupations differ in how pay relates to experience, an occupation-wide wage average can conceal weaker conditions for people entering those fields alongside different outcomes for experienced workers. That is a plausible interpretation of the distributional pattern, not a result about specific workers: the Dallas Fed analysis groups outcomes by occupation and cannot identify who gained or lost within each one. A relationship observed between two occupation-level measures does not show that the same relationship holds between two workers in one occupation. Nor does the proxy explain what employers value: occupational pay structures can reflect multiple differences that the analysis does not turn into a tested causal pathway. The analysis is useful for testing whether one average hides systematic variation, but it cannot reveal how an individual employer allocates tasks, sets pay, or changes its wage ladder.

It therefore helps explain how apparently conflicting wage readings may arise without resolving every study. Studies can differ in exposure definitions, time periods, samples, outcomes, and controls, so their estimates need not answer the same question. For example, an estimate of average wage growth across occupations and an estimate centered on graduates' earnings at labor-market entry have different populations and outcomes; neither cancels the other. The Dallas Fed relationship is conditional and descriptive; it does not adjudicate all wage research or establish an AI-caused pay effect. The detailed Apollo estimate discussed in the earlier draft is omitted because its linked paper could not be independently verified; no secondary account is substituted. Neither the near-zero median estimate nor the variation around it forecasts an individual's next raise, wage, or job security.

Sources: AI is simultaneously aiding and replacing workers, wage data suggest

When does faster work show up in pay or employment?

Faster task completion does not automatically become higher measured output, higher earnings, or more jobs. The ILO research brief, “The impact of GenAI on jobs, productivity and work organization: a review of the empirical evidence,” synthesizes experiments, firm-level data, platform studies, and worker and firm surveys across seven countries. It finds real but uneven productivity gains, while reported worker time savings have not yet translated consistently into measured output, earnings, or employment. The range of evidence matters: a controlled experiment can show that a task takes less time under specified conditions, while a survey records what respondents say happened in their own setting. Neither result alone demonstrates an economy-wide change in firm productivity or labor demand.

There are several distinct steps between a faster task and a labor-market outcome. Time saved is an input measure. To count as verified output, the work must be usable, meet the relevant quality standard, and remain so after checking, correction, and rework. Firm productivity asks whether output changed relative to the firm's inputs across a workflow, not merely whether one step became quicker. Earnings and employment then depend on how the organization uses any capacity released. As an analytical possibility—not a universal finding—saved time might support more demand and production, be absorbed by verification or coordination, change the mix of tasks assigned to staff, or reduce the hours needed for a fixed volume of work. Bargaining and compensation practices could affect whether any value reaches workers' pay. These are possible links to examine, not mechanisms the ILO review proves operate identically in each workplace.

That chain clarifies what a demonstration or self-reported productivity gain can support: a claim that a task may be completed faster in a particular test or reported context. It cannot by itself establish durable quality-adjusted output, a wage premium, staffing reductions, or new jobs. The ILO review's mixed evidence leaves room for settings where gains do flow into output or employment, while indicating that the conversion has not appeared consistently across the studies it reviews. Absence of consistent translation so far is not evidence that future gains cannot affect output, wages, staffing, or hours. The time horizon also matters: a short trial may detect immediate task speed while missing later workflow changes, the costs of supervision, or whether additional capacity finds buyers. A before-and-after comparison of one task therefore cannot stand in for evidence on longer-run firm or labor-market outcomes. For a worker interpreting a tool trial, the useful distinction is between local time saved and what the employer actually changes in workload, quality checks, production, compensation, or staffing; each is a separate observed outcome.

Sources: The impact of GenAI on jobs, productivity and work organization: a review of the empirical evidence

What counts as employer adoption rather than personal tool use?

Employer adoption is a ladder of organizational change, not a yes-or-no condition inferred from one employee opening a chatbot. At the first layer, an individual uses AI for a task, perhaps to draft text or summarize material. The next question is whether the tool is incorporated into a business function: it becomes part of how a team carries out work, with some repeatability and connection to its workflow. Functional breadth asks how many parts of the business use it. Operational investment is a further signal that the organization is changing how work is arranged around the technology. These layers describe different depths and reach; they do not mean that every function or worker in an adopting firm experiences the same change.

In *The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks (CES-26-25)*, the U.S. Census Bureau analyzes its Business Trends and Outlook Survey AI supplement for November 2025 through January 2026. It reports that 18% of firms used AI in a business function during that period. When weighted by employment, the estimate is 32%: workers are more concentrated in larger firms, so this measure gives larger employers greater weight. The figures answer different questions. The firm-weighted estimate describes the share of businesses reporting functional use; the employment-weighted estimate describes the share of employment associated with firms reporting it. Neither says that 18% or 32% of workers had their jobs automated, or that each reporting business had integrated AI deeply.

The paper also separates functional breadth and operational investment from worker-task integration when examining headcount. Broader functional integration and operational investment are positively associated with reported employment decreases. Worker-task integration by itself has no significant relationship with decreases once the other measures are accounted for. This pattern suggests that a task-level report—someone using AI in their own work—contains less information about headcount change than evidence that the business is reorganizing multiple functions or investing operationally around AI. The measures are not interchangeable: individual use can happen inside a stable workflow, while investment and cross-function deployment can indicate wider organizational change. The study reports associations, however, and does not show that broader deployment caused staffing reductions; other firm conditions may contribute to both adoption and employment decisions.

For a worker, the practical distinction is between access and a changed operating model. A license or a colleague’s experiment establishes availability or local use. Repeated use in a team process, links to shared systems, changed approval or quality-control steps, and deliberate investment in redesign are stronger signs that the organization has moved beyond experimentation. This is an interpretation guide, not a checklist that proves staffing intent: a business may integrate AI to expand output or improve service, and adoption can coexist with hiring. Informal worker use may also precede an official rollout, so not hearing management describe an AI program does not establish that no workflow is changing. The survey’s reference period is a snapshot, and its firm-level responses cannot identify what happened to a particular job. A reader should therefore treat visible workflow integration and staffing decisions as separate observations, rather than translating tool presence into a headcount forecast.

Sources: The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks (CES-26-25)

How much can an exposure score tell you about adoption?

An occupational exposure score can help locate work where AI capabilities may apply, but it cannot tell you whether an employer has deployed a system or plans to reduce staffing. *Generative AI and Jobs: A Refined Global Index of Occupational Exposure* from the International Labour Organization treats exposure as a task-based assessment of potential compatibility with generative AI and likely transformation patterns. Its unit of analysis is the tasks associated with occupations, assessed using task evidence and expert and model-assisted judgment. This makes an index a structured screen: it points to tasks worth examining for possible change. It is not an observation of actual adoption, a measure of a firm's workflow, or a probability that a particular worker will be displaced. Potential describes what could be affected under the index's assumptions; realized change depends on organizational choices and work conditions that an occupation score does not encode.

Census researchers test whether such measures nevertheless carry information about adoption in *AI Exposure and Adoption Among U.S. Firms (CES-26-61)*. They link 14 occupation-derived exposure measures, constructed using literature and BLS occupational employment shares, to firms' reported AI adoption in Census Business Trends and Outlook Survey data. Across the measures, a one-standard-deviation increase in exposure is associated with adoption rates 4 to 11 percentage points higher. That positive range matters: exposure is not an arbitrary label with no relationship to what firms report doing. At the same time, it is a relationship between occupation-based exposure measures and firm-level survey responses. It does not show that an exposed worker's own employer adopted AI, or that the worker's tasks were the reason for adoption.

The estimates become smaller when the analysis includes year and subsector controls: the association ranges from 1 to 8 percentage points. In other words, some of the raw relationship coincides with differences across time and industry subsectors; accounting for those differences leaves a positive but variable association. The gap between ranges also shows why a score cannot be read as a direct conversion rule. Different exposure measures capture task potential differently, and reported adoption varies with firm size and sector. The exposure measure predicts some variation in adoption in this dataset, but much of the variation remains outside that association. The paper does not supply a universal threshold at which a firm will adopt, nor a calibrated mapping from one worker's score to the likelihood of a workflow change.

Use the two studies together as a screening sequence. The ILO index helps identify which tasks in an occupation may be technically amenable to generative AI; CES-26-61 shows that higher occupation-level exposure tends to coincide with greater reported firm adoption, while the strength depends on the measure and controls. A high score therefore raises a reason to inspect actual tools, workflow integration, and local staffing signals. A low score does not establish immunity, since the exposure measures differ and employers may adopt tools for tasks not prominent in an occupational average. Neither score direction supplies a personal redundancy percentage. The inference for a career decision is to move from the broad occupational signal to the actual task bundle and employer evidence before changing course. Exposure is useful for prioritizing that inquiry; adoption and staffing remain separate facts to establish.

Sources: AI Exposure and Adoption Among U.S. Firms (CES-26-61); Generative AI and Jobs: A Refined Global Index of Occupational Exposure

Can an exposed occupation still grow?

Yes. An occupation can be considered susceptible to AI and still have projected employment growth, because exposure describes potential task change while an employment projection estimates the net direction of demand under many interacting forces. In “AI impacts in BLS employment projections,” the U.S. Bureau of Labor Statistics gives software developers and personal financial advisors as examples of occupations it describes as potentially susceptible to AI that are nevertheless projected to grow over 2023–33. These examples refute the simple equation exposure equals decline. They do not show that AI will create those jobs, that every worker will benefit, or that openings will be easy to reach.

A projection is a conditional estimate over a specified horizon, not a count of jobs already observed and not a controlled estimate of AI’s causal effect. BLS’s “AI impacts” factsheet explains that its occupational projections reflect expected influences across the labor market rather than isolating AI’s contribution. Demand for an occupation can move with industry output, replacement needs, technology, and other assumptions at once. If projected employment rises, that means the model’s net balance is positive; it does not identify which force produced the balance or how many positions would exist in a counterfactual world without AI. Nor does an occupation-wide total reveal whether new work is concentrated in particular regions, experience levels, or specialties.

For an individual choice, compare the projection with reachable local evidence. Start with openings within a feasible commute or remote-work market, then check whether their requirements recur across employers or appear in only one posting. Compare those requirements with the reader’s current skills and credentials, and note whether employers accept adjacent experience or demand a new credential, portfolio, or supervised practice. Then inspect which tasks the roles actually assign: a growing occupation may still reorganize routine work, shift entry-level duties, or raise expectations for verification and judgment. Ask what training time, tuition, lost earnings, salary range, schedule, health needs, and family responsibilities the route would require. A national growth figure cannot settle those personal and local conditions; it can only be one piece of the comparison.

Local signals need repetition and context. A single vacancy may be temporary or unusually specialized, while a few similar postings over time can suggest that a skill requirement is becoming common. Talk with people doing the work where possible, but treat an anecdote as a lead to check rather than a labor-market estimate. Compare posted requirements with actual tasks and ask whether the work is changing through tool use, revised review steps, or altered hiring criteria. If a task bundle includes recurring reports, data checking, exception handling, and approval, a worker could record which steps a tool drafts or checks, where a person must correct it, and who remains accountable for approval. This is a hypothetical observation exercise, not evidence about any occupation or employer.

The decision rule is to use projected growth as a reason to investigate a path, not as a promise to enter it. If repeated local openings match skills the reader can build within their time, income, and credential constraints, a small upgrade or adjacent application may be reasonable even in exposed work. If demand appears only in inaccessible locations or requires training the reader cannot presently afford, a positive national projection does not make that route practical. Conversely, flat or falling projections do not establish that every local niche is closed. Compare alternatives on reachable openings, entry routes, task mix, and total transition cost; revisit the choice when several local signals or actual workplace changes point in the same direction.

Sources: AI impacts in BLS employment projections; AI impacts

What is the smallest useful next move?

Choose one recurring task and track it for two to four weeks: how often it occurs, what AI changes, how much checking or rework remains, and whether workload, quality standards, or staffing actually shifts. If the task changes but your role still needs your judgment, try an upgrade: learn the tool and document reliable review. If nearby roles use your existing domain knowledge differently, explore an adjacent move through a small project or conversation. Consider a larger transition only when repeated workplace or local hiring signals justify its training, income, and family costs. If an employer has announced a material change or income is urgently at risk, prioritize immediate support and suitable job-search steps instead of waiting for a measurement period. That may mean checking benefits and severance terms, speaking with a trusted adviser, and targeting roles that use skills you can demonstrate now. The free task checker at /ai-job-risk-checker offers transparent task-level change-pressure signals and first actions; it does not calculate a validated probability of displacement. After the informational answer, the optional paid roadmap can compare scenarios and constraints, but guarantees neither employment nor income.

Questions readers ask