What September 2026 Changed in AI Exposure and Graduate Outcomes

What changed in September 2026?

September changed the question from whether AI-exposed work had fewer advertised openings to whether early-career outcomes were weakening. These are separate signals, not one pooled estimate or a forecast for an individual. “Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors” reports lower initial employment and earnings for graduates in the most-exposed major decile in national administrative data. The Dallas Fed’s “AI plays a role in weak labor market for college graduates” finds relative declines in Texas employment and first-year earnings among graduates from more-exposed majors. Its “Job postings show early signs of AI automation impact” measures advertised Texas openings, not realized graduate employment. King’s College London’s “Students not yet dissuaded by AI exposure in choice of degree subject” adds a distinct UK signal: applications rose relatively for more-exposed degrees while junior-position shares fell. These sources differ in geography, unit, outcome, and period: graduate majors, occupations and postings, or degree applications and junior openings. Their convergence makes early-career pressure harder to dismiss, but does not establish that AI alone caused every change. Treat September’s evidence as grounds to inspect entry routes and tasks, not to rank majors or predict anyone’s job loss.

Sources: Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors; AI plays a role in weak labor market for college graduates; Job postings show early signs of AI automation impact

What did September add to the graduate-outcome record?

The national Census working paper reports declines in initial employment and earnings concentrated in the most AI-exposed major decile; the Dallas Fed reports relative declines across Texas majors with higher exposure. These are related warning signals, not interchangeable estimates. Their importance is that the evidence now reaches beyond task applicability or advertised openings: two administrative-record analyses find weaker early outcomes for graduates in more-exposed pathways. In “Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors,” the national analysis compares college majors with different exposure levels after large language models became available. The publication page reports regression-adjusted results for the most-exposed tenth of majors: their likelihood of initial employment fell by five percentage points, and full-quarter initial earnings fell by 13%. It also says the earnings effect eased as graduates moved farther from labor-market entry, while remaining substantial for the most-exposed majors. The earnings result also has two parts. The working paper says roughly half of the decline came from lower earnings within the industry sectors employing these graduates. The remainder was associated with a shift in industry mix toward lower-wage sectors, including restaurants and retail. It cannot be read as a 13% pay cut for every person or employer. The Dallas Fed’s “AI plays a role in weak labor market for college graduates” uses administrative education and earnings records for students and graduates of Texas public universities. It defines first-year outcomes across the graduate’s first four quarters after graduation, using employment and earnings records for jobs held in Texas. These records do not capture work outside Texas. Comparing cohorts from more- and less-exposed majors, the analysis reports that a 10-percentage-point higher exposure share was associated after ChatGPT’s release with a 1.7-percentage-point relative decline in finding employment in Texas within a year. Among graduates who found Texas work, first-year earnings for more-exposed majors fell about 5% relative to less-exposed majors from 2021 to 2024. The study also reports that May 2024 graduates from majors 10 points higher in exposure were 1.4 percentage points more likely to enroll in graduate school within a year; many continued in the same field. The national five-point result is a change in employment likelihood for an exposed decile; the Texas 1.7-point result is a relative difference associated with a 10-point exposure gap. Their earnings outcomes also differ: full-quarter initial earnings in one analysis, first-year earnings among Texas-employed graduates in the other. Geography, exposure construction, cohorts, and outcome definitions differ too. Adding or ranking these percentages as if they measured one common effect would create false precision. Together, however, they make early labor-market entry a credible pressure point to monitor. How directly that finding applies depends on the person’s field, local opportunities, and route from study into work.

Sources: Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors; AI plays a role in weak labor market for college graduates

How is ‘AI exposure’ measured in these studies?

Exposure in these studies is a task-based estimate attached to an occupation or major. It describes how much work in that category appears applicable to generative AI; it does not say that a particular employer has adopted a tool, that the tool completes work reliably on its own, or that a person will lose a job. The Dallas Fed’s September 1 analysis starts with occupations as bundles of recurring activities. It uses an Anthropic metric that maps work listed in O*NET to tasks performed with Claude. The resulting occupation measure is interpreted as the share of tasks that generative AI can automate, based on observed platform activity rather than a purely hypothetical capability list. Recorded use on one platform cannot capture all workers, tools, or employer policies. Exposure therefore remains separate from firm adoption, role redesign, hiring demand, and displacement. The September 22 graduate analysis carries the occupation measure back to college majors through job advertisements. Using Lightcast postings from the first quarter of 2018 through the third quarter of 2022, it connects fields of study named in ads to the occupations those jobs requested. A major’s exposure is the average occupation exposure across postings mentioning that major. The pre-ChatGPT window keeps later demand changes from redefining which occupations feed into a major, but the result still describes jobs historically associated with a field in those ads. It does not grade a curriculum, describe every graduate, or tell whether an individual learned complementary skills. Consider an illustrative records role. A tool might help prepare a first draft from standard information, while a worker checks exceptions, verifies records against source documents, and takes responsibility for a decision. An occupation-level measure compresses a varied task bundle into an average: one worker’s week may contain more drafting, another’s more exception handling. The label cannot show which tasks dominate a particular job, how costly errors are, or what oversight is required. In “Job postings show early signs of AI automation impact,” the Dallas Fed explains occupations as task bundles and uses medical-records work as an example. This illustration clarifies why the bundle matters; it is not an observed result about every records worker. For a student or worker, read exposure as a prompt: which recurring tasks in my actual work are routine and digital, which require checking, context, or accountable judgment, and which build experience for the next role? Compare that task map with current local requirements. A broad major score cannot answer those questions on its own. It can focus attention on a possible change in work mix, while leaving the individual decision grounded in the tasks and experience routes that actually apply.

Sources: AI plays a role in weak labor market for college graduates; Job postings show early signs of AI automation impact

Why are postings, employment, and earnings different signals?

Postings, employment, and earnings answer different questions. Postings show what employers advertise; employment records show who appears in a defined set of work records; earnings show pay for a specified group. A decline in one measure can precede or accompany a change in another without matching its size. Together they locate hiring pressure, not jobs lost. “Job postings show early signs of AI automation impact” analyzes Texas Lightcast ads, linking occupation-level task exposure to quarterly postings. It compares more- and less-exposed occupations within industries. For a 10-percentage-point difference in automatable task share, postings for more-exposed positions were about 5 percent lower by the end of 2023 and about 8 percent lower by the first quarter of 2025. The Dallas Fed also finds incumbent firms reduced exposed-position postings and shifted their advertised role mix away from them. This measures advertised demand and ad composition, not an equal number of occupied jobs disappearing. An online ad is a proxy for an opening, not a complete record of recruitment or work. Employers may recruit through internal mobility, referrals, or other channels; they may revise or consolidate ads, or delay hiring. These are limitations, not established explanations for each missing ad. Lightcast also underrepresents farming, construction, building maintenance, and personal-service openings, the Dallas Fed notes, . The September 22 “AI plays a role in weak labor market for college graduates” examines administrative education and workforce records for graduates of Texas public universities. After ChatGPT’s release, a 10-point higher major exposure measure was associated with a 1.7-percentage-point relative decline in finding Texas employment within a year. Among graduates employed in Texas, first-year earnings for more-exposed majors fell about 5 percent relative to less-exposed majors from 2021 to 2024. These outcomes concern Texas-recorded work and particular cohorts. Earnings cover people who found Texas work, not those outside that group. The national Census working paper “Graduating into Disruption” uses a different graduate sample and exposure comparison. Its abstract reports a five-percentage-point decline in initial-employment likelihood for the most-exposed major decile and a 13 percent drop in full-quarter initial earnings. Roughly half the earnings decline came from lower earnings within industry sectors; the remainder was associated with a shift toward lower-wage industries. Thus an average can fall because pay changes within sectors, graduates enter different sectors, or both. This is not directly comparable to Texas earnings among employed graduates: populations, exposure groups, definitions, and periods differ. Use local postings to check visible openings and requirements, then compare them with your actual work, internal routes, and pay conditions. A thinner set of ads is a signal to investigate, not an estimate of jobs gone or a verdict on your prospects.

Sources: Job postings show early signs of AI automation impact; AI plays a role in weak labor market for college graduates; Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors

Can the timing prove AI caused the changes?

No single September comparison proves AI alone caused weaker outcomes for exposed graduates. Post-ChatGPT changes are consistent with an AI contribution, but evidence that some measures were already worsening before November 2022, alongside other labor-market shifts, prevents an exclusive-cause conclusion. The narrower reading is credible entry pressure, with causes varying by measure and group. The January 2026 preprint “AI-exposed jobs deteriorated before ChatGPT” reports that U.S. unemployment-insurance records showed unemployment risk in AI-exposed occupations rising in early 2022, months before ChatGPT. Its analysis of millions of LinkedIn profiles finds that graduate cohorts from 2021 onward entered exposed jobs at lower rates than earlier cohorts, with gaps opening before late 2022. This complicates a single-release explanation but does not rule out an added shock; the measures cover different populations and outcomes. The preprint also reports that graduates with more AI-exposed curricula had higher first-job pay and shorter job searches after ChatGPT, based on millions of university syllabi. This cautions against avoiding a field because of exposure. But the observational association does not show that a course caused better pay or faster hiring. The preprint’s unemployment, LinkedIn, and syllabus data are separate streams, not one experiment following people from study into work. An April 2026 Census working paper, “You’re (not) Hired: Artificial Intelligence and Early Career Hiring in the Quarterly Workforce Indicators,” adds a distinct post-release signal. Using matched employer-employee administrative data, it reports that employment among workers ages 22–24 in the most AI-exposed industry-state quintile declined 12% over the ten quarters after ChatGPT’s introduction. Fewer hires drove the decline; hiring rates largely recovered by early 2025 against a smaller employment base. Event studies show a discontinuous decline in early-career job gains and backfill hires at release relative to older workers in the same industries. That timing is consistent with an AI-related hiring change, but does not settle causation. The Census paper also finds evidence of earlier shifts in employment, hiring, and separations around COVID’s onset. Its industry-state exposure groups differ from the graduate-major comparisons in September’s studies. Monetary-policy shocks may explain up to one quarter of relative early-career employment declines through 2025’s second quarter, the paper estimates, but do not explain the rapid relative hiring decline at the most exposed firms. These qualifications narrow alternatives without removing them or making the working paper a universal causal estimate. The Dallas Fed’s September graduate analysis belongs in a developing record, not as a verdict alone. Its Texas major-level comparisons add post-ChatGPT employment and earnings evidence; the preprint documents earlier weakness, while the Census study finds a hiring break alongside COVID-era shifts. Their different samples, exposure definitions, and outcomes cannot be pooled into one estimate of AI’s contribution. Readers should take fewer local openings seriously, then check whether the pattern persists and whether tasks or role requirements shifted. Sustained local hiring change or documented workplace redesign could strengthen the AI-contribution interpretation; a reversal would weaken it.

Sources: AI-exposed jobs deteriorated before ChatGPT; You’re (not) Hired: Artificial Intelligence and Early Career Hiring in the Quarterly Workforce Indicators; AI plays a role in weak labor market for college graduates

Where does early-career pressure show up in the work pathway?

The pressure appears most clearly at the doorway into work: fewer openings or hires can make it harder to get a first role and accumulate experience, even without layoffs among established staff. The evidence does not show that every exposed occupation disappears. Their measures should not be treated as one estimate. In “Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors,” national administrative records show that graduates in the most AI-exposed decile of majors experienced a five-percentage-point decline in initial employment likelihood and a 13% decline in full-quarter initial earnings after late 2022. Effects lessen as graduates move further from entry. A separate Census working paper, “You’re (not) Hired,” examines matched employer-employee records by industry-state cells and age, not college major. For workers aged 22–24 in the most exposed quintile of cells, adjusted employment fell 12% over ten quarters after ChatGPT’s introduction, mainly because of fewer hires. Hiring rates largely recovered by early 2025, but on a smaller employment base. This concerns entry hiring, not broad dismissals. The studies use different populations and exposure groupings, so their percentages are not comparable effect sizes. “Job postings show early signs of AI automation impact” supplies another signal. In Texas Lightcast ads, postings for more-exposed positions were 5% lower relative to less-exposed ones by end-2023 and about 8% lower by first quarter 2025, within industries and scaled to a ten-point exposure difference. Ads measure advertised demand, not filled jobs or eliminated positions. The UK study summarized by King’s College London examined 1,090 degrees and applications from 2020–2025: applications rose relatively for more-exposed degrees after ChatGPT, while junior and entry-level positions in target occupations fell 0.25 percentage points per standard deviation of exposure. Applications show interest, not enrollment or eventual employment; junior-position share is not a layoff measure. These findings locate concern at the transition into work, while leaving scale and cause uncertain. One plausible mechanism to investigate is that fewer entry openings narrow opportunities to learn through supervised, routine assignments. Imagine first-pass research, standard documentation, or template code drafted with AI: someone may still need to check sources, handle exceptions, supply context, and take responsibility for sign-off. If employers reduce starter assignments, new workers could have fewer chances to demonstrate judgment and build experience. The studies do not establish that sequence across workplaces; it is an interpretation to test against role design. Ask which tasks are changing and what supervised work demonstrates capability; compare local job descriptions and employer expectations. Fewer hires, ads, or junior-position shares are warning signals about entry access, not proof every exposed task has vanished.

Sources: Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors; You’re (not) Hired: Artificial Intelligence and Early Career Hiring in the Quarterly Workforce Indicators; Job postings show early signs of AI automation impact; Students not yet dissuaded by AI exposure in choice of degree subject

What do education choices and entry-level openings add?

September’s education evidence points in two directions: some students continue to choose exposed fields as early-career openings weaken, while some Texas graduates return to graduate school. Neither finding shows that an exposed major should be abandoned or another degree will protect a graduate. They describe applications, enrollment, and advertised positions, not which education choice produces better employment or pay.

The Dallas Fed’s ‘AI plays a role in weak labor market for college graduates’ links Texas public-university education records with state employment and earnings records. May 2024 graduates from majors with a 10-percentage-point higher exposure measure were 1.4 percentage points more likely to enroll in graduate school within a year. Most returning students stayed in their undergraduate field. Yet new master’s graduates in more-exposed fields experienced relative earnings declines similar to four-year graduates. More formal study alone may not solve a mismatch; the result does not establish that graduate school caused poor outcomes or that every program has the same return. These are Texas records, and enrollment is not completion or proof of career benefit.

A different signal comes from the UK study summarized by King’s College London, ‘Students not yet dissuaded by AI exposure in choice of degree subject.’ It examined domestic applications across 1,090 degrees from 2020 to 2025, linking UCAS data with LinkedIn profiles. After ChatGPT’s release, degrees one standard deviation higher in AI exposure saw six per cent relative growth in applications, while the share of junior and entry-level positions in target occupations fell by 0.25 percentage points per standard deviation. Applications are not enrollments or completions, and a LinkedIn-based share is not a count of graduate hires. The summary cautions that students were not necessarily choosing badly: new tasks may emerge and entry-level hiring may recover. The finding supports information when choosing a field, not rankings of safe or unsafe majors.

The practical question is what a learning step must accomplish. A degree may fit a destination that requires it or a goal needing foundations and supervised depth. A targeted course or certificate may address a bounded gap in current work, such as evaluating AI-assisted outputs; a project or self-study can test a workflow and build applied evidence. They differ in prerequisites, feedback, signaling, time, and cost. Completion does not guarantee job readiness or hiring. The January preprint ‘AI-exposed jobs deteriorated before ChatGPT’ reports a positive association between syllabus-based AI-related education exposure and early outcomes, but its observational design cannot show that a course caused them.

Before committing, name the work outcome you want, check its credential and skill requirements in your location, and identify the gap training would close. Compare tuition and time away from paid work with lower-cost practice and feedback. Include salary needs, caregiving, health, and geography. Choose education for a defined destination and evidenced prerequisite, not because an exposure label makes a field sound unsafe.

Sources: AI plays a role in weak labor market for college graduates; Students not yet dissuaded by AI exposure in choice of degree subject; AI-exposed jobs deteriorated before ChatGPT

What should a current worker infer from graduate evidence?

Graduate evidence is a leading signal about entry routes and task demand, not a direct forecast for an established worker. Its relevance depends on whether your role relies on similar repeatable work and whether the path from routine assignments to more complex responsibility is changing. A current worker may have more tenure, client relationships and context, which can shape task assignment and mobility; these are possible differences, not immunity. The Texas administrative-record analysis, “AI plays a role in weak labor market for college graduates,” links major-level exposure to graduate employment and earnings in Texas. It reports a relative decline in Texas employment within a year of graduation and lower first-year earnings among employed graduates from more-exposed majors after ChatGPT’s release. The Census working paper “You’re (not) Hired” examines a different group: workers aged 22–24 in industry-state cells, not graduates grouped by major. It reports a 12% relative decline in adjusted employment in the most exposed quintile over ten quarters, with fewer hires driving much of the decline. These findings make entry worth watching, but do not forecast an established employee’s outcome; populations and exposure measures differ. Use that distinction to audit your work rather than assign yourself a risk label. List your outputs for a typical week. Mark recurring digital production, such as drafting standard summaries or organizing routine information, separately from work that involves checking exceptions, interpreting context, handling sensitive exchanges or signing off on consequences. Note AI assistance, verification, and which assignments build next-level experience. These are inspection categories, not claims of full automation. Compare this list with recent local postings and internal role descriptions for changed requirements, fewer entry routes, or greater emphasis on review and accountability. “Job postings show early signs of AI automation impact” analyzes advertised Texas openings, a demand signal, not a count of eliminated jobs or a report on your role. If a recurring exposed task is central to your work, first look for a way to demonstrate domain judgment and reliable verification within the field, while checking whether the workflow and employer rules permit the tool. A sustained local decline in openings, direct redesign of your responsibilities, or a blocked route to more complex work would justify a closer adjacent-role review. Until such evidence appears, graduate outcomes support attention to task mix and experience-building, not an automatic career change.

Sources: AI plays a role in weak labor market for college graduates; Job postings show early signs of AI automation impact; You’re (not) Hired: Artificial Intelligence and Early Career Hiring in the Quarterly Workforce Indicators

Which response is proportionate: upgrade, adjacent move, or larger change?

When the signal is broad or indirect, begin with a task-and-entry-route audit and a reversible experiment. Consider an adjacent move or larger change only when local evidence, a defined goal, and your constraints point that way. Exposure alone does not justify an expensive pivot. The Dallas Fed’s Texas graduate analysis reports weaker relative employment and earnings for graduates from more-exposed majors, while its Texas postings analysis finds fewer advertised openings for more-exposed occupations. They are not forecasts for your role or verdicts on degrees. An upgrade in place is the smallest useful test if your field still fits and your work includes a repeatable digital task. Choose one workflow, check employer rules, and keep human review where accuracy, context, or accountability matters. For two to four weeks, record the task, time required, corrections found, and whether the result meets the same standard. This suggested experiment is not a proven intervention or promise of productivity gains. If use is unsafe or correction costs outweigh the time saved, that is useful evidence too. An adjacent move makes sense when your experience transfers but the route to more complex work is changing. Preserve domain knowledge, customer understanding, or familiarity with systems and rules; then inspect local postings and internal role descriptions before targeting roles involving coordination, evaluation, client context, or accountable decisions. Treat these as requirements to verify, not universally protected work. The Dallas Fed’s “Job postings show early signs of AI automation impact” measures online ads and notes that some occupations are underrepresented. King’s College London’s “Students not yet dissuaded by AI exposure in choice of degree subject” reports UK applications rising for exposed degrees while junior and entry-level position shares fell. Neither finding establishes whether a particular opening exists where you live. Check current routes directly. A larger change needs a clearer case because it can cost tuition, time, and foregone income. Name the destination and specific gap first. Compare a degree’s depth and credential value with a targeted course or certificate, supervised project, apprenticeship, or self-study; assess prerequisites, feedback, evidence of skill, duration, and total cost. Include salary floor, location, health, and family obligations. The Dallas Fed graduate analysis found exposed Texas graduates more likely to return to graduate study, often in the same field, while exposed master’s entrants also showed relative earnings declines. This does not prove graduate study is ineffective; it cautions against assuming another credential automatically solves a task-demand problem. Move faster if local hiring keeps weakening, your employer directly redesigns your role, a funded route fits your goal, or your position faces a near-term threat. Otherwise, gather personal evidence and reassess before a costly commitment.

Sources: AI plays a role in weak labor market for college graduates; Job postings show early signs of AI automation impact; Students not yet dissuaded by AI exposure in choice of degree subject

What is the proportionate next move?

Choose one recurring task this week and note what it produces, how often it repeats, and what happens when an exception appears. Mark where AI might help with a first draft, search, summary, or routine transformation; then mark the checks, context, and accountable decisions that remain yours. This is a task audit, not a verdict on your occupation. Next, compare that task with one requirement in a current local posting or internal description for a role you might want. Look for a specific gap you can act on, rather than inferring demand from an exposure label. The Dallas Fed’s “AI plays a role in weak labor market for college graduates” reports relative early-career changes in Texas records, but this regional graduate evidence cannot predict your outcome or establish what your employer will adopt. Choose a small learning or workflow experiment tied to the gap, such as producing and checking one representative work sample, then reassess before committing money or time you cannot spare. The free [AI job risk checker](/ai-job-risk-checker) organizes transparent change-pressure signals and first actions, not validated displacement odds. The paid [career roadmap](/career-roadmap) compares stay-and-redesign, adjacent, and larger-change scenarios against your constraints; it does not guarantee employment or income.

Sources: AI plays a role in weak labor market for college graduates

Questions readers ask