What “Graduating into Disruption” Shows About AI-Exposed Majors

What should a student conclude from “Graduating into Disruption”?

The September 2026 Census working paper “Graduating into Disruption” reports that employment and earnings began diverging after ChatGPT’s late-2022 release for graduates from the most AI-exposed college majors. In adjusted estimates, this top exposure decile saw initial employment likelihood fall by five percentage points and full-quarter initial earnings fall by 13%. Effects lessen farther from labor-market entry but remain substantial for the most exposed majors. These findings concern early outcomes in the study’s administrative records, not every graduate’s prospects. The paper compares groups of majors with differing estimated exposure. It does not show that AI alone caused every difference, that employers adopted AI uniformly, or that an individual with one of these degrees will lose a job. Exposure is not a personal job-loss probability. A major-level result also cannot tell a student whether a particular role nearby is hiring or what tasks an employer expects a new graduate to perform. Take the evidence seriously without treating the label as a verdict. Before paying for another degree or abandoning useful subject knowledge, choose a realistic entry-level role and location, inspect openings and recurring tasks, and identify a complementary skill. Test it with an affordable, bounded project or course. Then compare what the test shows with the cost, time, prerequisites, and constraints of a larger change. The immediate decision is how to apply and demonstrate what you know in actual work; this paper gives a reason to investigate carefully.

Sources: Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors

What did the Census paper actually find?

The Census working paper reports a sharp difference in first labor-market outcomes for graduates from the most AI-exposed majors after ChatGPT became available in late 2022. In its regression-adjusted comparison, the top exposure decile had a five-percentage-point decline in the likelihood of being employed in the quarter after graduation, relative to graduates in the bottom six exposure deciles. Their full-quarter initial earnings were 13% lower relative to less-exposed fields. These are distinct measures: five percentage points describes a change in employment likelihood, while 13% describes relative earnings. Neither number estimates an individual graduate’s chance of unemployment. The paper follows bachelor’s graduates in administrative records and compares outcomes across majors with different estimated exposure around the introduction of large language models. Its event-time analysis asks whether outcomes for graduates from more-exposed fields changed differently after late 2022 than outcomes for less-exposed graduates. That comparison is more informative than simply observing difficulty after 2022, but it does not make AI the sole explanation for every difference. The paper is a working paper, and observational timing is not randomized assignment of AI use. The earnings result reflects where graduates started work as well as what they earned within an industry. The paper attributes roughly half of the decline to lower earnings within employing industries and roughly half to a shift in industry mix toward lower-wage sectors, including restaurants and retail. It also describes movement away from higher-paying areas such as information and professional, scientific, and technical services. A weaker first outcome can therefore arise through job matching and sectoral sorting as well as pay changes within broad industries. This describes early-career outcomes, not a count of positions permanently eliminated. The estimated effects attenuate as graduates move farther from labor-market entry, while remaining substantial for the most exposed majors. The authors connect attenuating earnings differences and increased job switching with possible adjustment into better-paying work. They also state that follow-up is too short to determine when, or whether, exposed graduates catch up with peers who entered before large language models were widely available. An early gap can matter without establishing a permanent penalty; attenuation does not prove that every graduate fully recovers. The sample contains 6,665,500 graduate records from 27 data partners in 24 states, representing about 29% of bachelor’s degrees conferred from 2016 through 2024. Participating institutions are not every U.S. college, and the records have specific coverage limits discussed in the paper. The findings are not a universal ranking of majors or a forecast for every region, institution, or labor-market route. The title signals measured disruption at labor-market entry; it does not mean all tasks associated with an exposed major disappeared or every graduate had the same result. For a student or recent graduate, the practical reading is narrow but serious: first-job access and initial job matching deserve attention, especially in fields near the top of this exposure measure. Check which roles are hiring locally, what recurring tasks openings require, and whether a small additional skill or work sample could improve fit. These results alone do not justify abandoning a major, taking on graduate-school costs, or assuming lasting earnings loss. Such choices require evidence about a specific target role and personal constraints, not an exposure-decile result alone.

Sources: Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors; Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors

What does “AI-exposed” mean in this study?

In “Graduating into Disruption,” “AI-exposed” describes the estimated reach of AI capability across tasks associated with the occupations linked to a college major. It does not mean that employers adopted AI for those tasks, that software completed them reliably without people, or that workers were displaced. The score lets the researchers compare graduate groups by a shared measure; it is not an individual job-loss probability. The Census paper builds its measure in stages. It starts with task-level ratings from GPT-4, based on whether the technology could reduce the time needed for a listed work task by at least half. A task receives a full point when that criterion is met, and a half point when it could be met with access to additional software. Task ratings are aggregated to occupations, then linked to majors using the occupation mix of young workers with a bachelor’s degree. Using survey responses from young workers with a bachelor’s degree, it estimates that mix. Thus the score summarizes associated occupations, not every graduate’s duties. This is a measure of potential task-time change, not a record of use. The working paper itself says the exposure measure is based on AI’s capacity to perform tasks and does not necessarily reflect actual use. It also cautions that technical exposure does not mechanically lead to adoption or labor substitution. Adoption may depend on implementation costs and whether the technology improves on a worker’s contribution. In practice, exposure can accompany automation, assistance that changes how work is done, or little immediate change. The Federal Reserve note “Educational Exposure to Generative Artificial Intelligence,” published in February 2025, offers a useful comparison, but its numbers should not be treated as interchangeable with the Census measure. It uses Felten and colleagues’ occupational exposure scores for language modeling and image generation, based on survey judgments about how related those applications are to occupational abilities. The note maps those scores to majors using the occupations reported by respondents to the National Survey of College Graduates in its 2013–2021 waves. The instrument, population window, and mapping differ. The note says directly that its available data cannot establish whether exposed occupations will be automated away or augmented, and that higher exposure does not automatically imply either potential. Consider a hypothetical graduate whose target role includes drafting summaries from records. AI might shorten a first draft, while deciding which records matter, checking whether the summary is accurate, and taking responsibility for its use remain parts of the work. It shows why capability, adoption, redesign, and displacement must be assessed separately. For a career decision, inspect recurring tasks in the roles and locations you are considering, including the work needed to verify outputs and answer for them. Use the major score as a prompt to investigate those tasks, not as a verdict on the value of your degree.

Sources: Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors; Educational Exposure to Generative Artificial Intelligence

Does the Texas evidence tell the same story?

It points in the same direction, but it is not a numerical replication. In “AI plays a role in weak labor market for college graduates,” the Dallas Fed reports that, after ChatGPT’s 2022 release, each 10-percentage-point higher share of automatable tasks associated with a major corresponded to a 1.7-percentage-point relative decline in the chance of Texas employment within a year of graduation. Among graduates who found work in Texas, first-year earnings for more-exposed majors also fell about 5% from 2021 to 2024 relative to less-exposed majors. These are relative changes across groups, not the share of graduates who lost jobs or a forecast for an individual.

The Texas study links education records from Texas four-year public universities with quarterly employment and earnings records from the Texas Workforce Commission. First-year employment means finding a job in Texas within four quarters after graduation; work elsewhere is not counted as Texas employment. May 2021 graduates were the last cohort whose first labor-market year preceded ChatGPT. The study traces how differences between more- and less-exposed majors changed across graduating cohorts. It reports stable employment-rate differences before the release, followed by a relative decline for more-exposed majors. Earnings are compared only among graduates who found Texas work, not across all graduates.

The exposure mapping differs from the Census working paper. The Dallas Fed uses an Anthropic task measure linked to occupations in O*NET, then maps occupations to majors through Lightcast job ads from 2018 Q1 through 2022 Q3. This pre-release window avoids using post-ChatGPT hiring patterns to define the map. Exposure is the average occupation score among ads mentioning a field of study. By contrast, “Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors” uses GPT-4 task-time-reduction ratings and the occupation mix of young bachelor’s workers. Its multi-state sample and comparison of the top exposure decile with the bottom six deciles differ from the Texas population and exposure gradient.

The studies use different populations, geographies, exposure measures, and comparisons, so their estimates should not be pooled or called exact confirmation. Their similar direction is useful corroboration: separate analyses find weaker early outcomes in groups whose majors map more strongly to automatable work. Texas records capture Texas employment, while the Census sample covers participating institutions and states, not every graduate. Neither study measures employer adoption, completed automation, or personal displacement. Adjusted comparisons strengthen the case for attention, but do not show that AI alone produced every difference.

For a student or recent graduate, the Texas findings are a reason to look closer at the market where work is actually possible. A Texas estimate cannot be transferred unchanged to another state or country. Check local entry-level roles, the tasks they ask graduates to perform, and the evidence employers request. Then decide whether a small project or focused learning step could demonstrate a complementary skill. The study raises a practical question about early-career demand; it does not answer whether a particular major is worth keeping.

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

Why can graduates feel pressure before mass layoffs?

A labor-market warning does not have to begin with dismissals. Employers can reduce new positions, alter a role’s task mix, or pause hiring while current staff remain at work. An entry-level market can tighten before broad layoffs appear. A graduate experiences the missing opening directly, even if no incumbent has been replaced. Hiring, job design, and displacement are related, but not interchangeable. The Dallas Fed’s “Job postings show early signs of AI automation impact” analyzes Lightcast online postings in Texas, assigning occupations an automation-exposure measure based on tasks associated with observed Claude use. Its model compares more- and less-exposed occupations within the same industry after ChatGPT’s November 2022 release. The authors estimate that GenAI automation exposure reduced Texas Lightcast postings by about 1.8% in 2024 and 2.6% in 2025. These are model-based estimates of exposure’s contribution to posting behavior, not counts of positions eliminated or workers laid off. A relative ad decline does not mean every exposed occupation will contract at that rate. The study suggests a plausible route by which AI-related task change could affect new graduates: firms may advertise fewer openings in roles with more automatable work before reducing existing staff. The Dallas Fed notes that online postings often require limited prior experience and argues that a hiring pullback can matter to new entrants. Its companion analysis, “AI plays a role in weak labor market for college graduates,” uses Texas public-university records and reports weaker employment and earnings outcomes for graduates of more-exposed majors. Together, the analyses are consistent with a hiring channel worth watching. They do not show that posting changes caused the graduate outcomes or predict what will happen to one student. Lightcast is a collection of online advertisements, not a complete count of vacancies, hires, separations, or employment. The Dallas Fed notes limited coverage for some work, including farming, construction, building maintenance, and personal services. Its results concern Texas and one exposure model; they do not establish adoption in every workplace. Exposure describes tasks tools may perform or assist with, while actual use depends on organizational choices and workflow fit. A posting decline is evidence about advertised demand, not proof that a task was automated or a worker displaced. For a practical check, choose one or two entry-level roles in a location where you can work. Review recent postings from several employers; note recurring responsibilities, required experience, and task changes, comparing over time when possible. Check official employment information and employer career pages too, since ads do not capture all openings or completed hiring. A repeated decline in suitable opportunities is a reason to investigate adjacent roles or test a complementary skill, not an instruction to abandon a field. Local evidence is more useful than applying a Texas estimate to an individual career.

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

How much of the change can be attributed to AI?

The timing and pattern make generative AI a plausible contributor to weaker entry outcomes, but the studies do not establish it was the sole cause. In Graduating into Disruption, Census researchers compare graduates across majors with differing preassigned exposure in an event-time regression. The divergence follows ChatGPT’s late-2022 release and is concentrated among the most-exposed majors. This is not a randomized experiment: causal interpretation depends on there being no unmeasured, time-varying shock to particular majors that coincided with AI diffusion. The paper identifies alternatives, including geography and industry conditions, changing program composition or quality, seasonal hiring, and remote work. The study adjusts for stable college-major differences, college-wide shocks, and seasonal patterns. It also examines work-from-home trends, graduate supply, and alternative exposure measures, reporting robust results. Its event-time estimates show a break after ChatGPT’s release; pre-release patterns help rule out alternatives such as a different post-pandemic recovery or the unusually tight reference-period labor market. These checks make some explanations less persuasive, not impossible. A college-wide adjustment cannot remove every shock affecting one field at one institution, and observed alternatives do not cover every unmeasured factor. The authors’ conclusion is evidence suggestive of an AI role, not proof each missing first job or lower paycheck was caused by an AI system. The Dallas Fed’s Texas analysis offers a check using a different population and exposure measure. It reports employment differences between more- and less-exposed majors were stable before ChatGPT; afterward, a 10-percentage-point higher share of automatable tasks was associated with a 1.7-percentage-point relative employment decline within a year of graduation. The pre-release pattern supports a change around the technology’s arrival. Still, event-study interpretation assumes no other contemporaneous change selectively affected more-exposed fields. Texas public-university graduates and state employment records do not represent every U.S. graduate, and the Dallas mapping differs from Census’s. Agreement in direction strengthens concern, but estimates should not be pooled or presented as independent proof of one precise causal effect. Graduating-cohort outcomes can also worsen when labor-market conditions are weak, even without AI. The Journal of Labor Economics abstract for “Cashier or Consultant? Entry Labor Market Conditions, Field of Study, and Career Success” reports that, among U.S. graduating classes from 1974 through 2011, entering during a large recession reduced initial earnings by about 10%, with small persistent wage effects. This shows entry conditions shape outcomes; it does not make a recession equivalent to AI exposure or establish that recession dynamics explain the newer pattern. The careful verdict is that AI likely contributed to the divergence, while its exact share and interaction with other changes remain uncertain. That is enough to treat the evidence as an early warning about entry-level demand, but not to assign a cause to an individual graduate’s result. Findings from later cohorts, other regions, and direct measures of employer use could change this conclusion.

Sources: Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors; AI plays a role in weak labor market for college graduates; Cashier or Consultant? Entry Labor Market Conditions, Field of Study, and Career Success

Why do entry conditions matter beyond the first job?

The first job matters because it is where a graduate begins accumulating paid experience, work-specific knowledge, and a record that can help with later moves. A weak start can affect more than the first paycheck. But the evidence here supports care, not a prediction that an exposed graduate’s early earnings gap will last throughout a career. In “Graduating into Disruption,” the Census working paper reports that the earnings effect becomes smaller as graduates move farther from labor-market entry, while remaining substantial for those from the most exposed majors. The paper describes both movement toward lower-paying industries and lower earnings within industries. The study also observes increased job switching and says this pattern is consistent with subsequent adjustment toward higher-paying employment. It cannot yet determine when, or whether, exposed graduates catch up with peers who graduated before large language models became widely available. Its post-2022 observation window is too short to settle that question. Historical research helps explain why the entry point deserves attention, but it cannot tell us that the AI pattern will follow the same path. The Journal of Labor Economics abstract for “Cashier or Consultant? Entry Labor Market Conditions, Field of Study, and Career Success” covers U.S. college graduating classes from 1974 through 2011. It reports that graduating during a large recession reduced initial earnings by about 10%, through both full-time employment and wages, with small persistent effects on wages. The abstract also reports that effects differed by major and that the Great Recession produced larger early losses than past patterns and recession size would have predicted. This is evidence that conditions at graduation can shape early outcomes; it is not a direct comparison of AI-exposed and less-exposed graduates. A recession affects hiring through broad economic demand, while AI exposure concerns tasks and how employers may use technology. The mechanisms are not interchangeable, and the accessible evidence here is the journal abstract rather than the full study. The practical implication is to treat a difficult transition as a period to build options and gather evidence. Relevant experience can come from an entry-level role, a project, or another bounded opportunity to demonstrate the work a target role requires; the feasible route depends on location, income needs, time, and access. Track whether responsibilities and opportunities improve, rather than reading one initial job title or earnings period as a permanent verdict. The Census paper’s evidence of attenuation is limited: the measured gap narrows with time, but full convergence remains unknown. Prioritize credible experience and review progress at a chosen interval, then adjust the search or learning plan if the same barriers persist. This takes the first transition seriously while leaving room for later movement.

Sources: Graduating into Disruption: Labor Market Outcomes for AI-Exposed College Majors; Cashier or Consultant? Entry Labor Market Conditions, Field of Study, and Career Success

Does another degree solve the problem by itself?

No evidence here supports treating another degree as an automatic fix. In the Dallas Fed analysis, Texas public-university graduates from majors with greater estimated exposure were more likely to enroll in graduate school after ChatGPT’s release. For May 2024 graduates, each 10-percentage-point increase in a major’s estimated share of automatable tasks was associated with a 1.4-percentage-point increase in graduate enrollment within a year. Many returned to the same field: two-thirds of 2024 computer science graduates who returned chose a computer science graduate program. The analysis also reports that new master’s graduates in more-exposed fields experienced similar relative earnings declines to four-year graduates. This calls for asking what further study adds; it does not show every master’s degree has low value or that graduate education cannot help. The authors interpret increased enrollment partly through opportunity cost: when employment opportunities weaken, the wages a graduate gives up by studying may be lower. That helps explain why graduates may return to school, but enrollment itself is not evidence that a program improves later outcomes. The Texas study observes patterns in one population and period; it does not compare every program or establish what an individual would earn with versus without a credential. The authors leave open that added expertise could be more complementary to AI than undergraduate knowledge. The practical question is: what capability, access requirement, or hiring signal does this path add for the work I want? A degree may make sense when a target role requires it, or when its curriculum supplies depth, structured feedback, or a recognized qualification that a shorter route cannot provide. A course or certificate can be more focused, but its title alone does not demonstrate workplace ability. A project can show how someone frames a problem, checks outputs, and applies subject knowledge; its value depends on fit with the target work and useful feedback. An apprenticeship offers supervised practice where available. Self-study costs less in tuition but requires sustained time and a way to assess quality. These paths differ in prerequisites, duration, feedback, signaling, and cost; the Dallas findings guarantee no return for any of them. Before enrolling, choose one or two roles and locations you could realistically pursue. Check whether current openings repeatedly require a qualification or capability you lack. Compare the degree’s curriculum and time commitment with a bounded project, focused course, certificate, apprenticeship, or self-study plan that could test the same gap. Include tuition and lost income, along with schedule, health, family responsibilities, and where you can study. If the credential opens a necessary door or teaches a defined capability unavailable through a smaller step, its cost may be justified. If the gap is narrower, test it cheaply first. The study supports scrutiny of added value, not a blanket retreat from education.

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

What can a student check in their own target market?

Start with work you could realistically pursue: choose one or two entry-level roles and a geography, including whether you can relocate or need remote work. Review recent employer postings. Note recurring tasks, required credentials, skills marked required versus preferred, experience thresholds, dates, and locations. This compares advertised requirements with what you can show. It is a snapshot of demand, not a complete vacancy count or proof every listed requirement is fixed. The Dallas Fed’s “Job postings show early signs of AI automation impact” analyzes Texas Lightcast online ads linked to occupation-level task exposure. Its model estimates that GenAI exposure was associated with total Texas postings about 1.8% lower in 2024 and 2.6% lower in 2025. More-exposed positions also had larger relative declines. These are modeled posting changes, not counts of jobs eliminated, hires prevented, or employers that adopted AI. The authors note online-ad coverage varies by occupation. Use the study to motivate a closer look at entry-level demand and task descriptions, not as a forecast for your city or a verdict on your major. Look for a gap recurring across postings rather than reacting to every preferred skill. Compare it with your coursework, experience, and projects. If several openings call for checking or applying information in context, test that capability in a bounded work sample. For instance, prepare a short analysis with an appropriate AI-assisted workflow, then document how you checked the output, corrected errors, and made it useful for a defined task. This is an illustration, not a claim about requirements in a particular field. Ask an instructor or practitioner familiar with the work to critique it. Choose learning to address the observed gap. Problem framing, data literacy, evaluation, subject knowledge, and verification can transfer across tools; a particular interface may change. A course can teach a defined skill, while a project gives practice and something to show. A certificate or degree may fit when target roles repeatedly require it, structured feedback matters, or deeper study serves a clear goal. Before paying, compare prerequisites, time, cost, feedback, and relevance to those openings. The Bureau of Labor Statistics says its ten-year projections describe outcomes under assumptions, not certain forecasts, and AI’s effects remain highly uncertain. Use them for broad context, not as a substitute for local postings. If requirements remain out of reach or openings do not fit your constraints, compare adjacent roles before larger retraining. This check limits sunk cost; it cannot guarantee a job.

Sources: AI Impacts on the Labor Market; Job postings show early signs of AI automation impact

What is the next proportionate move?

Treat “Graduating into Disruption” as a prompt to investigate, not a verdict on your major. The Census working paper and Dallas Fed Texas analysis report entry-level differences by exposure, but neither predicts outcomes for a particular town or occupation. Choose one realistic role and location. Review current entry-level openings; note recurring tasks, qualifications, and preferred skills. Compare those demands with your existing knowledge and experience.

If a gap recurs, test it before making a large commitment. Build a small work sample showing a complementary capability, such as checking AI-assisted output against source material or applying subject knowledge to a workflow. If evidence points to an adjacent role, compare its prerequisites and openings with your current path. Consider a degree or major retraining when it supplies a defined capability or access requirement a focused project, course, certificate, apprenticeship, or self-study cannot reasonably provide. Include time, cost, income needs, location, health, and family responsibilities.

If relevant entry-level openings are contracting sharply, or the role centers on readily automated tasks, accelerate the adjacent-path assessment. Check more than online postings. The Bureau of Labor Statistics says projections depend on assumptions and AI’s labor effects remain highly uncertain; they offer context, not a local hiring forecast. The free task checker organizes change-pressure signals, not job-loss probabilities. The paid career roadmap compares paths against your constraints, but guarantees no job or income. Gather local evidence, run one low-cost test, then revisit your decision as work and hiring signals change.

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

Questions readers ask