betaFeedback

WorksEconomics, Econometrics and Finance

AI-exposed jobs deteriorated before ChatGPT

Abstract

Public debate links worsening job prospects for AI-exposed occupations to the release of ChatGPT in late 2022. Using monthly U.S. unemployment insurance records, we measure occupation- and location-specific unemployment risk and find that risk rose in AI-exposed occupations beginning in early 2022, months before ChatGPT was released. Analyzing millions of LinkedIn profiles, we show that graduate cohorts from 2021 onward entered AI-exposed jobs at lower rates than earlier cohorts, with gaps opening before late 2022. Finally, from millions of university syllabi, we find that graduates taking more LLM-exposed curricula had shorter job searches after ChatGPT and higher first-job pay when entering highly exposed occupations. Together, these results—pointing to forces pre-dating generative AI and to the ongoing value of LLM-relevant education—suggest that generative AI may not be the root cause of these labor market hardships and that the launch of ChatGPT is a faulty natural experiment for understanding AI's impact on jobs.

Together, these results—pointing to forces pre-dating generative AI and to the ongoing value of LLM-relevant education—suggest that generative AI may not be the root cause of these labor market hardships and that the launch of ChatGPT is a faulty natural experiment for understanding AI’s impact on jobs.

Many observers worry that college-educated workers are disproportionately exposed to task automation driven by generative AI [1, 2] and thus face job displacement as these tools diffuse1. Timing is a key implication of this hypothesis: if deteriorating labor market outcomes in recent years were driven primarily by large language models (LLMs), then the sharpest deterioration should occur after the mass-market emergence of LLM applications. ChatGPT—launched in November 2022 and rapidly adopted by consumers [3]—provides a salient inflection point for the broader diffusion of LLM services (e.g., including Claude, Copilot, and Gemini). And, consistent with employment concerns, early-career workers in LLM-exposed occupations experienced an employment decline [4, 5] and/or decreasing salaries [6] while recent college graduates experience elevated unemployment [7] during this period. If LLMs are the cause of these poor outcomes, then perhaps teaching LLM-exposed skills fails to prepare students for the labor market.

Two features of the past four years complicate this timing logic. First, LLM capability has improved continuously rather than in a single step. Frontier models were already improving before ChatGPT’s release, and the pace of improvement has increased since. The length of task that these models complete with 50% reliability has doubled about every seven months since 2019, and about every three months since 2024 [8]. Benchmark progress accelerated with the introduction of reasoning models [9], and general-purpose systems have since moved from single-turn assistance to longer-horizon agentic work [10]. ChatGPT’s launch was therefore a discontinuity in public awareness and consumer adoption [3]—the fastest consumer-technology diffusion on record—rather than a discontinuity in the underlying rate of capability growth. Second, hiring responds to employers’ expectations about future capability, not only to the tools available on a given date. A randomized information experiment finds that employers systematically underestimate the automatability of white-collar work and revise those beliefs upward when shown evidence of AI capability, yet do not correspondingly change hiring plans, instead anticipating higher productivity, new complementary roles, and greater training investment [11]. Together these facts imply that no single release date cleanly separates “before AI” from “after AI.”

However, more recent public discourse2 casts doubt on AI as the driver of job market woes and instead points to “AI washing.” This occurs when corporate leaders cite artificial intelligence as the primary reason for layoffs, restructuring, or hiring freezes when the actual drivers are financial, such as cost-cutting, pandemic-era overhiring, or poor performance. Other research supports this with studies finding little impact from LLMs on workers’ wages or hours worked [12, 13, 14, 15]. This null result has persisted as adoption has grown: a second wave of Danish survey and administrative data finds near-universal chatbot use in supported workplaces but no measurable effect on earnings, with pre-existing wage differences between users and non-users predating ChatGPT [16]; a U.S. survey linking occupation-level generative AI use to administrative data finds small positive wage effects and no decline in job openings or employment in exposed occupations [17]; and firm- and industry-level analyses of job postings find no reduction in hiring where AI adoption is highest [18].

It’s possible that other macroeconomic and sector-specific forces were already shifting in early 2022 contributing to weakening outcomes in the same AI-exposed occupations. For example, monetary tightening accelerated through 2022–2023 [19] and may have impacted job stability [20]. Also, job postings for software developers declined markedly over 2022–2023 [21] perhaps as a correction to increased hiring during pandemic years [22]. Instead of LLM diffusion, if such forces were central drivers, then deterioration could be detectable before ChatGPT’s launch and worsening labor outcomes may be mis-attributed to AI.

While research finds mixed evidence, public perception has largely been shaped by negative coverage in the news. Even in the months after ChatGPT’s launch, headlines warned of decreased white collar jobs in the near term3. More recently, similar concerns persist with emphasis on programming jobs4 and on the employment prospects of recent college graduates5. While these articles point to AI as a potential cause, they are informed mostly by subjective interviews and perspectives. Efforts to settle the academic debate may additionally ground the public discussion and guide decisions by policymakers, educators, employers, and students moving forward.

To investigate AI’s role in and the timing of these labor market outcomes, we integrate three large-scale datasets spanning administrative unemployment records, online career histories, and university curricula. The unemployment records run monthly from January 2015 through October 2025, the career histories cover graduation cohorts from 2017 through 2024 followed for up to three years after degree completion, with job start dates observed through December 2024, and the syllabi cover courses offered through 2017 (see SI Sections S2, S7, and S12, respectively). First, using monthly unemployment insurance claims linked to occupations and states, we construct a monthly measure of unemployment risk and find that risk rose for LLM exposed occupations beginning in early 2022—months before ChatGPT. Throughout, we define an occupation’s AI or LLM exposure using existing exposure estimates [2] according to the share of its workplace tasks in the U.S. Department of Labor’s O*NET database that could be affected by LLMs with simple interfaces and general training (see SI Section S1 for this measure in detail and SI Section S5 for analyses with alternative exposure scores). Second, analyzing millions of LinkedIn profiles from Revelio Labs, we show that graduate cohorts from 2021 onward entered LLM exposed jobs at lower rates than earlier cohorts with these gaps opening months or even a year before ChatGPT’s launch. Finally, using millions of university course syllabi [23], we measure graduates’ educational exposure to LLM-related content and find that higher exposure predicts higher first-job salaries and shorter job searches after ChatGPT. Together, these results separate pre-existing labor-market weakening from later LLM diffusion and underscore the ongoing value of LLM-relevant skill development during higher education. These findings warn against using ChatGPT’s 2022 launch as a clean natural experiment and cast doubt on LLMs as the cause of the worsening outcomes that recent graduates and AI-exposed workers experienced through 2025.

Unemployment risk for AI-exposed occupations

We first focus on unemployment for the workers of AI exposed occupations that are most likely to experience change because of AI automation. Because unemployment data are reported at the major occupation group level (two-digit codes in the Standard Occupation Classification (SOC) system), we measure each group’s exposure by averaging the LLM exposure scores of its detailed occupations (six-digit SOC codes). That is, a detailed occupation’s exposure is the share of its O*NET workplace tasks that are exposed to LLM automation [2], and a major group’s exposure llmsoc\mathit{llm}_{\mathrm{soc}} is the equal-weighted mean of the exposure scores of the detailed occupations nested within it. By this measure construction and extraction (SOC 47) is the least exposed major group (llmsoc=0.069\mathit{llm}_{\mathrm{soc}} = 0.069) and computer and mathematical occupations (SOC 15) the most exposed (llmsoc=0.741\mathit{llm}_{\mathrm{soc}} = 0.741). SI Section S1 gives this aggregation in full and reports the exposure score of every detailed occupation (SI Table S1) and of every major occupation group (SI Table S2). All other analyses use exposure at the six-digit level directly. Next, we incorporate monthly administrative records from the ETA 203 report from the Employment and Training Administration (ETA) within the U.S. Department of Labor. This report aggregates monthly data from each U.S. state’s unemployment insurance program detailing the most recent major occupation group of each continuing claimant. After combining with annual occupation employment data from the BLS Occupational Employment and Wage Statistics (OEWS) program, we calculate an occupation’s unemployment risk [24] which approximates the probability of claiming unemployment in each state each month given a worker’s occupation (see SI Section S2 for a description of unemployment risk, SI Section S2.1 for a comparison between the ETA 203 report and the US Census Current Population Survey (CPS), and SI Section S2.2 for the calculation of unemployment risk). In cases where data is not available for a state-occupation pair, that pair is not included in any aggregation of unemployment risk scores. Note that unemployment risk is meant to describe one of the worst career outcomes following a job loss (i.e., unemployment by occupation) but other negative outcomes may occur and be missed by unemployment risk (e.g., workers switch to lower paying jobs without claiming unemployment).

For each quarter, we regress log unemployment risk on quarter fixed effects interacted with major-group LLM exposure llmsoc\mathit{llm}_{\mathrm{soc}} and plot the implied percent change in unemployment risk when moving from an unexposed (llmsoc=0\mathit{llm}_{\mathrm{soc}} = 0) to a fully exposed (llmsoc=1\mathit{llm}_{\mathrm{soc}} = 1) major group (see Figure 1A); the observed range runs from construction (0.069) to computer/math (0.741). Typically, unemployment risk is 20-80% lower in LLM-exposed occupations (e.g., computer/mathematical) compared to less exposed occupations (e.g., construction). This gap widens during 2020–2021, consistent with pandemic-era differences in the ability to work remotely. Beginning in early 2022, however, the gap narrows sharply with some quarters of 2023-2025 exhibiting no difference between

Figure 1. Unemployment insurance (UI) risk for AI (LLM) exposed workers increased starting before the launch of ChatGPT. (A) Percent difference in unemployment risk between high- and low-exposure occupations (see SI Section S2.2 for calculation). (B) Unemployment risk for occupations grouped into exposure quintiles. (C) Unemployment risk by major occupation. Lines are colored by major occupation’s AI exposure. We compute unemployment risk monthly across states and report quarterly series as the mean of monthly values across all states within each quarter (see SI Fig. S60-S62 for similar plots with alternative AI exposure scores and SI Section S6 for an analogous monthly time series). Panels B and C use a logarithmic vertical axis; SI Table S4 reports the same quantity for every major occupation group as a percentage. Quarterly aggregation serves two purposes here: it keeps 22 overlapping series legible, and averaging the three monthly values within each quarter smooths month-to-month noise, so the series plotted here are already smoothed relative to the underlying monthly data. The unsmoothed monthly series are reported in SI Section S6, and the formal timing analysis in SI Section S6.1 is performed on those monthly series rather than on these quarterly averages.

high- and low-exposure groups.

Is this deterioration because exposed occupations have more unemployment risk after the launch of ChatGPT? We group occupations into quintiles of LLM exposure and plot average unemployment risk over time (Figure 1B). Across the period, lower-exposure quintiles exhibit higher levels and stronger seasonality (i.e., because of the seasonality in construction and building/grounds maintenance, see Figure 1C), while all quintiles show a pandemic spike in 2020 followed by a trough in early 2022. Importantly, unemployment risk in the most exposed quintiles begins rising after this early-2022 trough—well before ChatGPT’s November 2022 launch—and, rather than accelerating, then stabilizes in the quarters following the launch.

Disaggregating by major occupation group (see Figure 1C and SI Section S3 for each of the 22 major groups plotted separately and without normalization), many high-exposure occupations, including computer and math (SOC 15), office/admin support (43), business/financial operations (13), and architecture/engineering (17), exhibit large increases in unemployment risk during 2022–2025. However, these increases again begin prior to ChatGPT’s launch and flatten or decrease afterward. Overall, the timing is inconsistent with a deterioration that begins only after ChatGPT and suggests that pre-existing macroeconomic and sectoral forces contributed materially to the early-2022 rise in unemployment risk. Rather than impose a treatment date, we also let the data locate structural breaks in each exposure quintile’s unemployment risk series and ask which candidate dates they recover (see SI Section S6.1). We further ask how much of the movement in unemployment risk is accounted for by macroeconomic conditions rather than by AI (see SI Section S6.2). The narrowing of the exposure gap tracks the federal funds rate: the interaction between LLM exposure and the interest rate is positive and significant, and it is large enough that the gap implied at a 5% policy rate is most of the way to the narrowing visible in Figure 1A. Once that interaction is in the model, the terms specific to LLM exposure after the launch of ChatGPT add less than 0.001 to R2R^2 and are not significant. Without the interest rate interaction, the post-launch exposure term is large and highly significant. That is, dating treatment to November 2022 without controlling for the monetary tightening that began eight months earlier produces a large post-launch effect of LLM exposure.

Another potential concern is that an earlier vintage of AI was impacting jobs before the launch of ChatGPT in ways that overlap with exposure to modern LLMs. For example, GitHub’s Copilot became publicly available in June 2022, months before the launch of ChatGPT, and acted as an auto-complete tool for writing computer code. Even if this tool saw widespread adoption among computer programmers (i.e., within computer/math occupation (SOC 15)), its use likely did not diffuse into the work of the other high-exposure occupations, such as office admin/support, architecture/engineering, and legal occupations, that exhibited a similar early rise in unemployment risk. To the best of our knowledge, no other AI technology pre-dates the launch of ChatGPT, would be correlated with occupation exposure to LLMs, and had the potential to impact jobs across so many occupations; we therefore conclude that this early rise in unemployment risk for AI exposed occupations is likely not because of an earlier AI vintage and, instead, is more likely explained by external factors. Beyond the pre-trends issue, we also find that unemployment risk for AI-exposed occupations does not continue to worsen after the launch of ChatGPT (i.e., through 2023, 2024, and 2025)—which is hard to square with AI being the cause of rising unemployment risk if there were anticipatory effects in some corners of the economy, since we would then still expect worsening unemployment risk after the public launch of ChatGPT (and related tools) as AI diffused through the economy.

Our main analysis uses AI exposure scores that are widely used—even by AI companies like Anthropic [25]—to understand how AI is impacting jobs. But these patterns also hold under several alternative AI exposure scores [26, 1] including a measure using actual Claude AI usage data from Anthropic [27] (see SI Section S5). SI Section S6 provides monthly versions of Figure 1B&C and provides a break-point analysis for the monthly exposure quintile time series (SI Section S6.1). While these patterns hold nationally, a small number of states show post-launch increases in computer and math occupation unemployment risk (e.g., AK, CA, and WA; Figure S10, Figure S14, and Figure S56, respectively; see SI Section S4 for all 50 states). In these cases, timing alone cannot rule out a contribution from LLM diffusion. Additionally, California, Connecticut, Delaware, and Nebraska account for most of the state-occupation pairs with no reported claimants, but omitting these states from the analysis does not change the overall result (see Figure S5-Figure S9). However, on aggregate, post-launch unemployment risk remains comparable to pre-pandemic levels, suggesting limited movement in the longer-run equilibrium.

Figure S5. Same as Figure 1C (upper) and Figure 1B (lower), but omitting data from California. Each point is the average unemployment risk across all remaining states and months in each quarter.

Figure S9. Same as Figure 1C (upper) and Figure 1B (lower), but omitting data from California, Connecticut, Delaware, and Nebraska. Each point is the average unemployment risk across all remaining states and months in each quarter.

Figure S10. Quarterly unemployment risk by major occupation group in state of AK. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S14. Quarterly unemployment risk by major occupation group in state of CA. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Image not extracted

Figure S56. Quarterly unemployment risk by major occupation group in state of WA. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Labor market outcomes for recent college graduates

Unemployment risk is built from unemployment insurance claims, so it captures only workers who have lost a job and can claim benefits. Recent college graduates may not be eligible to claim unemployment benefits or may not have a most-recent occupation to include in their unemployment claim and would therefore be absent from our analysis of unemployment risk. Meanwhile, public discourse and other recent studies [4] suggest that AI most impacts early-career workers competing for LLM exposed jobs. Thus, while unemployment risk reflects career outcomes in the workforce generally, it may not reflect the outcomes of workers who experience the bulk of AI’s impact.

We address this shortcoming with 10,584,980 individuals’ LinkedIn profiles provided by Revelio Labs. Each profile contains a worker’s education history, including degrees obtained, field of study, graduation year, and degree-granting university, as well as their career history, including job titles (i.e., mapped to six-digit SOC code), employers, job start dates, and job locations. Additionally, Revelio estimates the salary for each job using a proprietary machine learning model. This salary model uses position-specific information, such as job title, seniority level, company, and location, as well as user-specific information such as the number of years an individual has worked at the company, and is trained on salaries found in publicly-available visa application data, self-reported data, and job postings. We are not able to directly verify the estimated salaries against workers’ actual salaries. Therefore, we additionally consider college graduates’ months to first observed job within three years post-graduation; individuals without an observed job within the window are excluded. This measure requires only that individual’s update their profiles when they start a new job and again corresponds to friction in the job market if graduates spend more time job seeking. Graduates who have not started a job within three years of graduation are excluded from our analysis, so delays are right-censored at three years and our analysis can miss the students who search longest or who leave the labor market entirely. Our estimates of job delays are therefore a lower bound. Because job start dates end in December 2024, the most recent cohorts are also observed for less than three years, so we repeat the job-delay analyses counting only first jobs that start within 12 months (or 6 months) of May of the graduation year and keeping only the cohorts whose data extend through that entire window (SI Section S15).

We examine job seeking for graduation cohorts while controlling for the monthly job opening

rate in the state and month of their first job, fixed effects for the job’s state and sector, and fixed effects for the individuals’ education characteristics including terminal degree type, field of study, and degree-granting university (see SI Section S7). We identify a job’s LLM exposure using Revelio Labs’ mapping of raw job titles to six-digit SOC occupation codes which we reference against [2]. In all years before the launch of ChatGPT, on average, graduates whose first job had high exposure spent less time job seeking than their peers (see Figure 2A). However, graduates from 2023 and 2024 whose first job was exposed instead spent more time job seeking. This reversal does not survive a common observation window. When every cohort is held to first jobs starting within 12 months of May of the graduation year, graduates of 2023 whose first job was exposed still found jobs faster than their peers, by a margin similar to the 2017–2019 cohorts; the advantage is largest for the 2020–2022 cohorts and narrows afterward, and within a 6-month window the 2024 cohort shows no significant difference (SI Section S15). The apparent reversal therefore reflects the shorter observation window available for the most recent cohorts rather than a post-ChatGPT penalty for exposed jobs. We retain the full-sample series in Figure 2A because it describes what recent graduates entering exposed fields actually experienced, and thus the experience that shapes public perception of early-career prospects in technology: recent graduates in computer science and related majors have faced elevated unemployment in these years relative to graduates of many other fields [7].

College graduates from 2022 onward spend more time securing jobs with high LLM exposure, but poor labor market performance starts months before ChatGPT

Figure 2. College graduates from 2022 onward spend more time securing jobs with high LLM exposure, but poor labor market performance starts months before ChatGPT. We consider graduates first job within three years after completing their terminal degree. (A) Graduate’s whose first job had high exposure typically spend less time job seeking compared to peers from the same graduation cohort, but this trend reverses after 2022 in the full sample; on a common observation window the advantage narrows but does not reverse (SI Fig. S122). (B) As a comparison, graduates whose first job was high salary also experienced increasing job delays after 2022, but delays that were still comparable to those in previous years. In (A) & (B), points show regression-adjusted differences; bars are 95% confidence intervals. (C) For each graduation cohort by month from May of their graduation year, the share of jobs with high LLM exposure started compared to the overall share of jobs. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. (D) Similar to (C), comparing the average share of jobs with high exposure (top decile) compared to the overall share of jobs for graduate cohorts around the launch of ChatGPT compared to earlier cohorts. In each time period, the class of 2021 or 2022 was significantly less successful at securing high-exposure jobs compared to earlier graduation cohorts. In (C) & (D), vertical lines indicate 95% confidence intervals and positive values indicate that graduates within a cohort are securing LLM-exposed jobs at a higher rate than less exposed jobs.

LLM exposure and pay are positively but only moderately related across occupations: exposure correlates 0.42 with the log of an occupation’s mean annual wage, and the most exposed occupations are middle- rather than top-earning (SI Section S7.1). Eloundou et al. report the same overall pattern, finding both low-wage occupations with high exposure and high-wage occupations with low exposure alongside a positive trend [2]. If the increase in job search times were specific to LLM-exposed occupations, an alternative salary-based comparison might produce a different result. Thus, we additionally examine the time to first job of graduates whose first job was high salary compared to their peers earning average salaries from the same graduation cohort (see Figure 2B). Consistent with our hypothesis, we observe a different trend for pre-pandemic graduates who spend more time job seeking. But the salary-based analysis moves in the same direction post-2022, suggesting a broader early-career tightening that is not specific to LLM-exposed workers. These descriptive analyses encourage us to look more carefully at the temporality of graduates’ job market performance.

Did recent cohorts’ shortfall in securing LLM-exposed jobs begin with the launch of ChatGPT?

If so, then each cohort’s share of newly started LLM-exposed jobs should fall only after the launch.

For graduation cohort gg, we therefore measure its over- or under-representation among newly started first jobs in LLM exposure decile llmllm as

jobGap⁡g,llm(t)=100⋅p(g∣t,llm)−p(g∣t)p(g∣t),(1)\operatorname{jobGap}_{g,llm}(t) = 100 \cdot\frac{p(g \mid t,llm) - p(g \mid t)}{p(g \mid t)}, \tag*{(1)}

where p(g∣t,llm)p(g \mid t,llm) denotes the share of jobs with LLM exposure decile llmllm started by graduates of class gg in month tt since May of their graduation year, and p(g∣t)p(g \mid t) denotes the corresponding share across all jobs (see SI Section S8).

Across exposure deciles, graduation cohorts 2013 through 2020 each exhibit nearly identical jobGap⁡g,llm\operatorname{jobGap}_{g,llm} trends in their first two years after graduation. For jobs started within each of the first four months after graduation, graduates secure a larger share of LLM-exposed jobs compared to the share of all jobs started. This difference in job finding rates gradually disappears around six months after May of the cohort’s graduation year.

However, only graduate cohorts 2021, 2022, and 2023 would have experienced job markets with generative AI adoption within two years after degree completion. Compared to earlier cohorts, these graduates exhibit lower jobGap⁡g,llm\operatorname{jobGap}_{g,llm} for the most LLM-exposed jobs (i.e., decile llm=10llm = 10. See Figure 2C) in their first few months in the labor market and then again at roughly one year after graduation. And this low performance for 2021 and 2022 graduates starts months before the launch of ChatGPT (note 2023 graduates entered the workforce only after ChatGPT’s launch). Specifically, we calculate the average jobGap⁡g,high(t)\operatorname{jobGap}_{g,high}(t) in the periods before and after the launch of ChatGPT and compare each with the 2013 through 2020 cohorts averaged over the same months since their own graduation; a two-sample tt-test finds significantly lower scores in all four comparisons (see Figure 2D). As a placebo test, we additionally consider graduate cohorts’ jobGap⁡g,llm\operatorname{jobGap}_{g,llm} trends for the least LLM exposed jobs (i.e., decile llm=1llm = 1) and observe no period of significant difference (see SI Fig. S79 and SI Section S8 for statistical results). We provide similar analyses for all LLM-exposure deciles and, alternatively, for job salary deciles in SI Sections S9 and S10, respectively. These analyses show that the shortfall is not specific to LLM-exposed work. Recent cohorts are under-represented in the upper deciles of both exposure and salary, and the gradient across salary deciles is if anything the cleaner of the two: after the launch the 2022 cohort is significantly under-represented in salary deciles four through ten. What neither margin shows is a sharp movement at the launch itself; the change in each cohort’s gap from before to after the launch is small and mostly indistinguishable from zero in both cases (see the summary figures in SI Sections S9 and S10). The deterioration these graduates experienced is broad and it is already underway when ChatGPT arrives. We provide an alternative analysis of the share of jobs by LLM-exposure demonstrating consistent results in SI Section S11.

Whatever AI’s role, recent college graduates faced a difficult job market. And students, educators, employers, and policymakers debate how higher education should change—both in terms of pedagogy [28] and workforce preparedness. For example, should students continue to learn skills that LLMs might automate, such as writing, coding, or information lookup? If LLMs primarily reduced the market value of LLM-adjacent skills, we might expect educational exposure to those skills to produce negative career outcomes after ChatGPT.

To investigate, we again turn to LinkedIn profiles and examine graduates’ first jobs within three years after degree completion. However, using profiles alone, our best proxy for education exposure to LLMs is workers’ self-reported field of study (i.e., college major) according to two-digit codes from the Classification of Instructional Programs (CIP) used by the National Center for Education Statistics which, alone, omits any variation by university or graduation year. Thus, we augment worker profiles with 3 million U.S. higher education course syllabi [23] to assess each worker’s education LLM exposure eduedu, defined from the semantic similarity between a course’s stated learning objectives and each ONET task, weighted by that task’s LLM exposure [2] and averaged over the syllabi associated with their graduation year, degree-granting university, and education field of study (the most recent syllabi are from 2017, so graduates after 2017 are assigned their program’s 2014–2017 syllabi; see SI Section S12). In other words, this measure captures how much of the LLM-exposed ONET task space is covered by the courses associated with a graduate’s inferred education experience. It is a similarity-weighted measure rather than a tally of tasks, so its values are interpreted in standard deviations or as a share of the exposed task space

rather than as a number of tasks (SI § 18.1).

Next, we introduce an indicator variable gptg_{pt} to identify when a first job start date occurred after the launch of ChatGPT. After controlling for local labor-demand conditions (e.g., by state and sector [30]), the share of job postings from 2017-2019 for the same job title as the first job that offered a work-from-home option [29, 31], and detailed education and job fixed effects, educational LLM exposure (eduedu) is not meaningfully associated with first-job salary or time to first job for pre-ChatGPT job starts (Models 1 & 3). In contrast, for post-ChatGPT job starts, higher eduedu predicts higher salaries and shorter job searches (positive edu×gptedu \times g_{pt} for salary; negative edu×gptedu \times g_{pt} for job seeking time). Jobs with higher occupation (i.e., six-digit SOC code) LLM exposure (llmllm) are associated with weaker outcomes after ChatGPT (See Models 2 & 4 containing negative gpt×llmg_{pt} \times llm for salary; positive gpt×llmg_{pt} \times llm for months). However, the three-way interactions (edu×llm×gptedu \times llm \times g_{pt}) indicates that educational exposure may partially offset these post-ChatGPT penalties in more-exposed jobs after the widespread adoption of LLMs, though the implied effect sizes are modest. Because gptg_{pt} is set by the start date of the first job, it is partly determined by the time-to-first-job outcome itself: within a graduation cohort, a graduate whose first job starts after the launch has necessarily searched longer. SI Section S15 therefore re-estimates Models 3 and 4 with post-ChatGPT status assigned by graduation cohort. The negative edu×gptedu \times g_{pt} and positive gpt×llmg_{pt} \times llm associations hold, including when the sample is restricted to the 2019–2023 cohorts and to first jobs starting within 12 months of May of the graduation year, a window that the data cover in full for each of these cohorts, whereas the three-way interaction does not (SI Table S21). SI Figure S105 plots the outcomes implied by Models 2 and 4 as a function of eduedu, separately for jobs before and after the launch and for jobs of high and low LLM exposure, so that the three-way interaction can be read as a change in slope rather than as a coefficient. Because the sample is large enough that significance alone discriminates poorly between findings, SI § 18.3 reports these associations as a percentage of salary and as days of job search, gives the variance explained by each block of terms, and re-estimates both models on random subsamples of 50,000 graduates (SI Table S19). The representativeness of each graduation cohort (e.g., by field of study) in the LinkedIn data is a limitation of this analysis; thus, we conduct additional analyses weighting worker profiles to mimic the graduates by field of study in the U.S. National Science Foundation’s National Survey of College Graduates and find consistent results (see SI § 19). We also repeat the graduation-cohort job gap analysis by decile of educational exposure (eduedu) rather than by decile of first-job LLM exposure (see SI Section S14). See SI Table S18 for a version of the regression analysis reporting the same models with the dependent variables and predictors in their original units.

Image not extracted

Figure S105. What Models 2 and 4 of Table 1 imply for a graduate’s first job. Rows are the two outcomes and columns the two periods, with gpt=0gpt = 0 for first jobs starting before December 2022 and gpt=1gpt = 1 for those starting on or after it. Within a panel the light line holds the first job’s occupational LLM exposure at llm=−1llm = -1 standard deviation and the dark line at llm=+1llm = +1, while education exposure varies along the horizontal axis; all predictors and both dependent variables are zz-scores, so a slope is the change in standard deviations of the outcome per standard deviation of eduedu. Each panel is drawn relative to its own reference graduate, who has average education exposure and enters a job at llm=−1llm = -1 in the same period, which subtracts the level shift shared by every job starting after the launch and retains every term involving eduedu or llmllm. The three-way interaction edu×llm×gptedu \times llm \times gpt is the change between columns in the gap between the two slopes: taking that gap as the high-exposure slope minus the low-exposure slope, it narrows from −0.038-0.038 to −0.016-0.016 for salary and reverses from +0.010+0.010 to −0.009-0.009 for months to first job. See SI Section S12.2.

VariableFull sample: Estimate (S.E.)Subsamples (n=50,000n = 50{,}000, 50 draws): MeanSubsamples (n=50,000n = 50{,}000, 50 draws): S.D.Subsamples (n=50,000n = 50{,}000, 50 draws): Mean S.E.Share of draws: Same signShare of draws: p<0.05p < 0.05
Panel A: log first-job salary (Model 2)
Education Exposure (eduedu)-0.001 (0.001)-0.0010.0040.00458%6%
LLM Exposure (llmllm)0.089 (0.001)0.0900.0050.005100%100%
edu×gptedu \times gpt0.004 (0.002)0.0040.0090.00960%4%
gpt×llmgpt \times llm-0.062 (0.002)-0.0610.0100.009100%100%
edu×llmedu \times llm-0.019 (0.001)-0.0200.0040.004100%100%
edu×llm×gptedu \times llm \times gpt0.011 (0.002)0.0120.0070.00996%28%
Panel B: months to first job (Model 4)
Education Exposure (eduedu)0.005 (0.001)0.0050.0040.00584%16%
LLM Exposure (llmllm)-0.014 (0.001)-0.0160.0050.005100%84%
edu×gptedu \times gpt-0.024 (0.002)-0.0250.0090.010100%76%
gpt×llmgpt \times llm0.032 (0.002)0.0300.0100.010100%84%
edu×llmedu \times llm0.005 (0.001)0.0050.0040.00494%24%
edu×llm×gptedu \times llm \times gpt-0.009 (0.002)-0.0140.0080.01098%30%

Table S19. The education exposure models re-estimated on random subsamples of 50,000 graduates. Models 2 and 4 of Table 1 are refit on 50 independent subsamples of 50,000 graduates drawn without replacement from the estimation sample, using the identical specification and seeded draws. For each term the table reports the full-sample estimate with its standard error, the mean and standard deviation of the estimates across draws, the mean of the standard errors estimated within draws, the share of draws whose estimate carries the same sign as the full-sample estimate, and the share of draws in which the estimate is significant at the 5% level with that sign. All coefficients are in standard deviations of the outcome per standard deviation of the predictor. A sample 25 times smaller implies standard errors 1,258,068/50,000=5.02\sqrt{1{,}258{,}068/50{,}000} = 5.02 times larger; the realized ratio is 5.07 in both panels. See SI Section S12.3.

Dependent Variable:Months to First Job
SampleAll cohorts12-month window
VariableModel 1Model 2Model 3Model 4
Job Opening Rate0.485***0.485***0.185***0.185***
Major SOC Log Annual Wage (2020)-0.018***-0.015***-0.032***-0.030***
Work-From-Home Score-0.013***-0.008***-0.004***-0.001
Education Exposure (eduedu)0.003***0.003***0.005***0.005***
LLM Exposure (llmllm)-0.019***-0.010***
gptcohort×llmgpt_{cohort} \times llm0.043***0.013***
edu×gptcohortedu \times gpt_{cohort}-0.022***-0.021***-0.017***-0.015***
edu×llmedu \times llm0.003***0.006***
edu×llm×gptcohortedu \times llm \times gpt_{cohort}-0.005*0.001
State FEYesYesYesYes
Sector FEYesYesYesYes
Terminal Degree Type FEYesYesYesYes
Graduation Year FEYesYesYesYes
College Major FEYesYesYesYes
University FEYesYesYesYes
R2R^20.1950.1960.0670.067
adj. R2R^20.1940.1950.0660.066
Observations1,258,0681,258,0681,036,2821,036,282

Table S21. Table S21: Time to first job with post-ChatGPT status assigned by graduation cohort. gptcohort=1gpt_{cohort} = 1 for graduation cohorts 2023 and later; it is absorbed by the graduation-year fixed effects, so only its interactions are reported. Models 1–2: all 2019–2024 cohorts in the Table 1 sample. Models 3–4: first jobs within 12 months of May of the graduation year, 2019–2023 cohorts. Continuous predictors and the dependent variable are standardized within each sample, as in Table 1. pval<0.1∗p_{val} < 0.1^{*}, pval<0.01∗∗p_{val} < 0.01^{**}, pval<0.001∗∗∗p_{val} < 0.001^{***}

If LLMs were to blame for graduates’ poor job market performance, then, after ChatGPT’s launch, we would expect to see that education exposure indicates redundant skills that do not add value when job seeking. Graduates who land AI-exposed jobs after the launch of ChatGPT earn less of a salary premium and spend more time job seeking than before, but these penalties are smaller for graduates with greater educational AI exposure. Taken on its own, the exposure of the first job (llmllm) shows the penalty one would expect if LLMs were to blame. The salary premium associated with higher occupational LLM exposure is sharply attenuated after December 2022: a one-SD increase in exposure is associated with 0.089 SD higher log salary before the launch and 0.027 SD after, a reduction of roughly 70%. Time to first job tells a similar story: before the launch, graduates entering more exposed occupations found jobs slightly faster than their peers, by 0.014 SD per SD of exposure, and after the launch they found jobs slightly slower, by 0.018 SD, although both magnitudes are small (Table 1, Models 2 and 4). Educational exposure, by contrast, is associated with less severe losses. Graduates whose coursework covered more AI-exposed skills had shorter job searches after ChatGPT’s launch and higher salaries when entering highly LLM-exposed occupations. These associations are difficult to reconcile with the view that LLM-relevant education became less valuable after ChatGPT. One caveat to our analysis is our inability to control for students’ self-selection into college majors, universities, or occupations upon graduation. Nevertheless, while not causal, our results associate LLM-relevant preparation with better, not worse, early-career outcomes in the post-ChatGPT period.

Dependent Variable:Log Salary (Real 2015 $) | | Months to First Job |
VariableModel 1Model 2Model 3Model 4
Job Opening Rate0.038***0.038***0.621***0.621***
Major SOC Log Annual Wage (2020)0.295***0.278***-0.015***-0.014***
Work-From-Home Score [29]0.074***0.044***-0.009***-0.006***
Post ChatGPT Indicator (gptgpt)0.035***0.031***2.335***2.338***
Education Exposure (eduedu)-0.000-0.0010.005***0.005***
LLM Exposure (llmllm)0.089***-0.014***
gpt×llmgpt \times llm-0.062***0.032***
edu×gptedu \times gpt0.007***0.004*-0.024***-0.024***
edu×llmedu \times llm-0.019***0.005***
edu×llm×gptedu \times llm \times gpt0.011***-0.009***
State FEYesYesYesYes
Sector FEYesYesYesYes
Terminal Degree Type FEYesYesYesYes
Graduation Year FEYesYesYesYes
College Major FEYesYesYesYes
University FEYesYesYesYes
R2R^20.4380.4420.3750.375
adj. R2R^20.4370.4420.3740.374
Observations1,258,0681,258,0681,258,0681,258,068

Table 1. Education exposure to LLM exposed tasks is associated with shorter job searches after ChatGPT and partially offsets salary penalties in highly exposed occupations. gpt=1gpt = 1 if job starts on or after December 2022. eduedu is education exposure, the exposure-weighted similarity between a graduate’s coursework and the O*NET task space, averaged over matched syllabi; it is not a count of tasks (see SI Sections S12 and S12.1). llmllm is first-job occupation six-digit SOC code LLM exposure [2]. Continuous predictors and dependent variables are centered and standardized (i.e., z-scores), so each coefficient is the change in standard deviations of the dependent variable associated with a one standard deviation increase in the predictor, and each interaction term is the product of its standardized components. All models include fixed effects for the state and two-digit NAICS code of the job as well as for the terminal degree type (e.g., bachelors or masters), graduation year, college field of study (i.e., two-digit CIP Code), and degree-granting university. pval<0.1∗p_{val} < 0.1^*, pval<0.01∗∗p_{val} < 0.01^{**}, pval<0.001∗∗∗p_{val} < 0.001^{***}

Discussion

While news media and some academic studies point to AI for the job market woes of early-career workers, our results indicate that worsening labor-market outcomes in 2022–2025 for LLM-exposed workers and graduates were already underway prior to the mass-market emergence of LLM applications. Unemployment risk in highly exposed occupations rose beginning in early 2022—well before ChatGPT—and in most occupations and states we observe no discrete break coincident with its introduction. Early-career workers were affected disproportionately: graduates from the 2021–2023 cohorts entered highly exposed jobs at lower rates and experienced longer observed delays to their first job than earlier cohorts, with gaps opening, again, before ChatGPT’s launch in late 2022. At the same time, LLM-relevant education was associated with better outcomes within this environment, particularly in highly exposed occupations. Graduates whose programs exhibited greater pre-2020 curricular emphasis on LLM-exposed tasks earned higher first-job pay and reached their first jobs more quickly in the post-ChatGPT period, particularly when entering highly exposed occupations. Together, these patterns suggest that LLM diffusion occurred in a labor market already shaped by macroeconomic and sectoral forces, while LLM-complementary skills retained—and possibly increased—their labor-market value. These findings thus cast doubt on the use of ChatGPT’s launch as a natural experiment to study AI’s impact on jobs and suggest that the worsening outcomes for recent graduates and LLM-exposed workers in 2022–2025 may be misattributed to AI.

If not LLM diffusion, then what? Our design identifies timing rather than causes, but the early-2022 onset we document coincides with several forces that plausibly acted on precisely the white-collar occupations that exposure measures rank as most AI-exposed. Monetary tightening began in March 2022 and was the steepest in four decades [19], and contractionary policy is known to fall unevenly across the labor market [20]. Simultaneously, the information sector unwound an extraordinary pandemic-era hiring expansion: software developer postings peaked in early 2022 and then fell sharply [21], consistent with correction of a hiring bubble rather than with automation [22]. A third candidate is fiscal: beginning in tax year 2022, Section 174 of the U.S. Internal Revenue Code required firms to amortize domestic research and experimentation expenditures—including software developer salaries—over five years rather than deduct them immediately, raising the effective cost of technical headcount exactly when our series turn 77. Finally, the reorganization of work itself shifted: early-career hiring is concentrated in roles whose viability changed with remote and hybrid arrangements [31, 29], and firms have adjusted labor demand as much by redesigning job content as by reallocating hiring across occupations [32]. We cannot adjudicate among these channels here, and we do not claim any one of them is decisive. Our claim is narrower: each of these forces was already operating before November 2022, each affected the same occupations that score highest on AI exposure, and a research design that dates treatment to ChatGPT’s release cannot separate any of these forces from LLM diffusion.

Independent evidence has since converged on this timing. Analyzing Lightcast vacancy data with a different exposure measure, Federal Reserve Bank of New York researchers likewise find that the relative decline in postings for AI-exposed occupations predates ChatGPT’s release, and that postings for junior and senior roles within highly exposed occupations declined at similar rates—contrary to the entry-level-first prediction [33]. Our results are nonetheless in tension with the most prominent contrary finding. Using high-frequency payroll records through mid-2026, Brynjolfsson et al. report that employment of 22–25 year-olds in AI-exposed occupations now sits roughly 19% below its counterfactual, operating through reduced hiring rather than separations, and they show this gap survives excluding technology firms and controlling for interest-rate exposure and remotability [4]. Two observations reconcile the accounts more than they oppose them. First, that gap is dated to a widening trend rather than to a level shift at November 2022, which is consistent with our finding of pre-ChatGPT onset. Second, the effect attenuates once education is controlled—which is what one would expect if the operative margin is the market for new college graduates broadly, as our cohort analysis suggests, rather than LLM substitution within occupations. Distinguishing these interpretations requires linked worker–firm data on actual LLM adoption, not exposure proxies tied to a release date. The closest existing design, which compares junior employment at firms that post dedicated generative-AI roles with that at firms that do not, finds a relative decline at adopters beginning in 2023 [5]; yet its aggregate series show junior employment flattening against senior employment from mid-2022, before ChatGPT’s release.

These findings have implications for research and policy. First, they caution against treating ChatGPT’s launch as a clean natural experiment for AI’s labor-market impact: designs that attribute post-2022 labor-market weakness primarily to LLMs risk confounding AI diffusion with concurrent macroeconomic shifts (possible examples include monetary policy, responses to work-from-home [31], sectoral demand, and/or post-pandemic adjustment). This caution is not merely methodological. When firms attribute layoffs to AI, and researchers date treatment to a product launch, the two reinforce a narrative that the underlying measurements do not support: adoption studies covering the period since ChatGPT continue to find little or no effect of LLM use on earnings, hours, or vacancies [16, 17, 18, 13] (but see [5] for junior employment), even as AI is named in a growing share of announced job cuts. Mis-attribution is costly in two ways. It directs policy attention toward technology-displacement remedies while leaving monetary, fiscal, and sectoral drivers unaddressed. And it may discourage students from undertaking the training that our third analysis associates with better outcomes. Second, they suggest that higher education and workforce programs should not abandon skills often labeled “AI-exposed,” such as writing, coding, and information synthesis. Instead, curricula that teach these skills in LLM-complementary ways (e.g., by emphasizing verification, evaluation, and effective human–LLM collaboration) may improve graduates’ resilience in an uncertain labor market. The experimental literature supports this reading of our correlational result. Across randomized and field experiments, generative AI raises output substantially—roughly 20–60% in controlled trials and 15–30% in the field [34]—in writing [35], software development [36], and customer support [37], with the largest gains accruing to less-experienced workers on well-specified tasks. But those gains are not uniform. Dell’Acqua et al. show that AI assistance sits on a “jagged technological frontier”: on tasks inside the frontier consultants improved sharply, while on a task just outside it AI-assisted consultants were markedly less likely to reach the correct answer than unassisted peers [38]. The competence that distinguishes the two cases is the ability to judge which side of that frontier a given task falls on—to evaluate, verify, and redirect model output rather than accept it. That is a curricular skill, and it can be taught in exactly the courses our exposure measure flags as LLM-exposed. Our finding that graduates with more LLM-exposed coursework saw shorter job searches and higher entry pay in highly exposed occupations is consistent with this literature, and it supports teaching these skills more deliberately rather than retreating from them. Our analysis of education exposure is purely descriptive but covers millions of educational experiences and graduate outcomes in the labor market and, to the best of our knowledge, it is the only study that connects LLM exposure in education and labor market outcomes in the wake of LLMs at such a large scale.

Our analyses rely on administrative unemployment insurance records, online career histories, task-based exposure measures, and syllabus-derived proxies for curricular content. Each source involves measurement and selection error. In particular, our educational exposure measure is derived from syllabi that predate 2020 and therefore captures baseline differences across programs rather than curricular responses to ChatGPT. It is also not comprehensive: the syllabi cover a subset of the courses and institutions a graduate may have encountered, so a graduate’s inferred coursework is a sample of what their program taught. Our time-to-first-job analyses exclude individuals whose first observed job begins more than three years after degree completion, which may understate delays among graduates facing the greatest barriers to entry. We do not identify the causal effect of LLMs on labor-market outcomes. Nonetheless, by triangulating across independent data sources, we document timing patterns that are difficult to reconcile with a sudden post-ChatGPT collapse in exposed occupations, shifting attention toward broader labor-market dynamics and toward expanding access to AI-complementary education. Our analysis does not rule out the potential for AI to create job loss in the future and so it is imperative that we continue to improve data, analysis, and our understanding of AI in the workplace. This caveat has grown more pressing since these data end: model capability on long-horizon tasks continues to double on a timescale of months [8, 9], and studies comparing exposure to outcomes over successive windows already report associations that were not detectable earlier [39]. A finding that the 2022–2025 deterioration predates LLM diffusion is therefore a descriptive statement about the past, not a forecast. Future work combining direct measures of LLM adoption with linked worker–firm data will be critical for distinguishing displacement from productivity gains and for identifying which workers, regions, and institutions benefit—or fall behind—in the next phase of diffusion.

Acknowledgments

We are grateful for input from Daniel Rock, Esteban Moro, Nathan Goldshlag, and Yong-Yeol Ahn.

Funding: M. R. F. was funded by a grant from the Alfred P. Sloan Foundation, through a gift from Microsoft’s AI Economy Institute, and this work is supported by the University of Pittsburgh Center for Research Computing.

Author contributions: All authors designed the research and drafted the manuscript. M.R.F. and A.J.S. gathered data and conducted the analysis.

Competing interests: There are no competing interests to declare.

Data and materials availability: Upon publication, all code will be publicly-available for download through a GitHub repository. Underlying worker profile data was provided by Revelio Labs and is not publicly available. However, data can be requested or purchased from Revelio Labs.

Supplementary Materials for AI-exposed jobs deteriorated before ChatGPT

M.R. Frank*; A. Javadian Sabet; L. Simon; S.H. Bana; R. Yu

*Corresponding author. Email: [email protected]

Exploring Eloundou et al. AI Exposure

Throughout the manuscript, an occupation’s exposure to AI is measured using the LLM exposure scores of Eloundou et al. [2]. Those scores are defined at the level of the individual workplace task. For each task τ\tau in the U.S. Department of Labor’s ONET database, Eloundou et al. assign an exposure value eτ∈{0,0.5,1}e_{\tau} \in\{0, 0.5, 1\} indicating whether access to an LLM (or an LLM-powered system) could reduce the time required to complete that task by at least half. An occupation’s exposure is then the average of eτe_{\tau} over the set Tsoc6T_{soc6} of ONET tasks belonging to that detailed (six-digit) SOC occupation,

llmsoc6=1∣Tsoc6∣∑τ∈Tsoc6eτ,(2)\mathit{llm}_{soc6} = \frac{1}{\lvert T_{soc6}\rvert}\sum_{\tau\in T_{soc6}} e_{\tau}, \tag*{(2)}

so that llmsoc6∈[0,1]\mathit{llm}_{soc6} \in[0, 1] can be read as the share of an occupation’s workplace tasks that are exposed to LLM automation. Table S1 reports llmsoc6\mathit{llm}_{soc6} for all 923 detailed occupations, together with ∣Tsoc6∣\lvert T_{soc6}\rvert, the number of O*NET tasks contributing to each average. Exposure ranges from 0.000 for occupations such as pile driver operators, dishwashers, and meat packers to 1.000 for mathematicians, with computer, office/administrative support, and writing-intensive occupations concentrated at the top of the distribution.

Detailed SOC CodeTitleAvg. LLM ExposureNum. Tasks
47-2072.00Pile Driver Operators0.0005
51-3023.00Slaughterers and Meat Packers0.00015
47-2053.00Terrazzo Workers and Finishers0.00026
51-9123.00Painting, Coating, and Decorating Workers0.00013
35-9021.00Dishwashers0.00013
31-1132.00Orderlies0.00024
51-3022.00Meat, Poultry, and Fish Cutters and Trimmers0.00013
47-2071.00Paving, Surfacing, and Tamping Equipment Operators0.00020
53-7031.00Dredge Operators0.0006
53-5022.00Motorboat Operators0.00017
37-2012.00Maids and Housekeeping Cleaners0.00025
47-3016.00Helpers–Roofers0.00018
49-9045.00Refractory Materials Repairers, Except Brickmasons0.00016
51-4052.00Pourers and Casters, Metal0.00016
35-2015.00Cooks, Short Order0.00011
49-9051.00Electrical Power-Line Installers and Repairers0.00023
47-5071.00Roustabouts, Oil and Gas0.00013
49-9096.00Riggers0.00014
51-9111.00Packaging and Filling Machine Operators and Tenders0.00020
47-2051.00Cement Masons and Concrete Finishers0.00026
45-2091.00Agricultural Equipment Operators0.00017
47-5041.00Continuous Mining Machine Operators0.00015
47-5022.00Excavating and Loading Machine and Dragline Operators, Surface Mining0.00016
49-3031.00Bus and Truck Mechanics and Diesel Engine Specialists0.00026
47-3011.00Helpers–Brickmasons, Blockmasons, Stonemasons, and Tile and Marble Setters0.00015
47-5081.00Helpers–Extraction Workers0.00014
27-2021.00Athletes and Sports Competitors0.0009
35-9011.00Dining Room and Cafeteria Attendants and Bartender Helpers0.00024
49-3022.00Automotive Glass Installers and Repairers0.00018
51-9195.00Molders, Shapers, and Casters, Except Metal and Plastic0.00027
35-2011.00Cooks, Fast Food0.00019
Detailed SOC CodeTitleAvg. LLM ExposureNum. Tasks
47-2022.00Stonemasons0.00017
49-3093.00Tire Repairers and Changers0.00026
47-2082.00Tapers0.00016
47-5043.00Roof Bolters, Mining0.00014
37-2011.00Janitors and Cleaners, Except Maids and Housekeeping Cleaners0.00023
47-3015.00Helpers–Pipelayers, Plumbers, Pipefitters, and Steamfitters0.00015
51-9196.00Paper Goods Machine Setters, Operators, and Tenders0.00015
51-4071.00Foundry Mold and Coremakers0.00013
47-3012.00Helpers–Carpenters0.00018
47-3014.00Helpers–Painters, Paperhangers, Plasterers, and Stucco Masons0.00011
47-2042.00Floor Layers, Except Carpet, Wood, and Hard Tiles0.00014
47-5011.00Derrick Operators, Oil and Gas0.00015
49-3052.00Motorcycle Mechanics0.00013
47-4061.00Rail-Track Laying and Maintenance Equipment Operators0.00026
51-9031.00Cutters and Trimmers, Hand0.00018
53-7041.00Hoist and Winch Operators0.00013
53-7073.00Wellhead Pumpers0.00016
49-9063.00Musical Instrument Repairers and Tuners0.01437
51-4031.00Cutting, Punching, and Press Machine Setters, Operators, and Tenders, Metal and Plastic0.01632
51-2011.00Aircraft Structure, Surfaces, Rigging, and Systems Assemblers0.01730
51-6021.00Pressers, Textile, Garment, and Related Materials0.01828
51-9197.00Tire Builders0.02223
51-2041.00Structural Metal Fabricators and Fitters0.02223
49-9044.00Millwrights0.02223
39-9031.00Exercise Trainers and Group Fitness Instructors0.02322
47-2142.00Paperhangers0.02520
47-2221.00Structural Iron and Steel Workers0.02520
45-4021.00Fallers0.02619
47-4051.00Highway Maintenance Workers0.02619
51-4194.00Tool Grinders, Filers, and Sharpeners0.02818
51-9081.00Dental Laboratory Technicians0.02917
51-4032.00Drilling and Boring Machine Tool Setters, Operators, and Tenders, Metal and Plastic0.02917
47-2011.00Boilermakers0.02917
49-9098.00Helpers–Installation, Maintenance, and Repair Workers0.03116
47-2061.00Construction Laborers0.03132
51-4021.00Extruding and Drawing Machine Setters, Operators, and Tenders, Metal and Plastic0.03116
51-2051.00Fiberglass Laminators and Fabricators0.03116
Detailed SOC CodeTitleAvg. LLM ExposureNum. Tasks
51-2023.00Electromechanical Equipment Assemblers0.03614
47-2151.00Pipelayers0.03614
29-9093.00Surgical Assistants0.03628
27-2031.00Dancers0.03614
47-2181.00Roofers0.03628
29-1022.00Oral and Maxillofacial Surgeons0.03614
47-5044.00Loading and Moving Machine Operators, Underground Mining0.03628
47-2021.00Brickmasons and Blockmasons0.03614
37-3011.00Landscaping and Groundskeeping Workers0.03727
39-4011.00Embalmers0.03826
47-2132.00Insulation Workers, Mechanical0.03813
47-2081.00Drywall and Ceiling Tile Installers0.03826
49-9081.00Wind Turbine Service Technicians0.03813
51-4022.00Forging Machine Setters, Operators, and Tenders, Metal and Plastic0.03813
51-6051.00Sewers, Hand0.03813
49-3021.00Automotive Body and Related Repairers0.04025
47-4091.00Segmental Pavers0.04212
47-3013.00Helpers–Electricians0.04224
39-4012.00Crematory Operators0.04212
39-3093.00Locker Room, Coatroom, and Dressing Room Attendants0.04323
51-7021.00Furniture Finishers0.04522
29-1024.00Prosthodontists0.04511
53-7051.00Industrial Truck and Tractor Operators0.04511
47-5012.00Rotary Drill Operators, Oil and Gas0.04522
51-6091.00Extruding and Forming Machine Setters, Operators, and Tenders, Synthetic and Glass Fibers0.04821
45-4011.00Forest and Conservation Workers0.04821
53-7062.00Laborers and Freight, Stock, and Material Movers, Hand0.04821
47-2152.04Solar Thermal Installers and Technicians0.04821
35-2021.00Food Preparation Workers0.04831
47-2131.00Insulation Workers, Floor, Ceiling, and Wall0.05010
37-3012.00Pesticide Handlers, Sprayers, and Applicators, Vegetation0.05010
45-3031.00Fishing and Hunting Workers0.05030
51-3091.00Food and Tobacco Roasting, Baking, and Drying Machine Operators and Tenders0.05020
51-9022.00Grinding and Polishing Workers, Hand0.05020
47-4031.00Fence Erectors0.05020
49-9052.00Telecommunications Line Installers and Repairers0.05020
51-9124.00Coating, Painting, and Spraying Machine Setters, Operators, and Tenders0.05030
51-6042.00Shoe Machine Operators and Tenders0.05319
Detailed SOC CodeTitleAvg. LLM ExposureNum. Tasks
47-2141.00Painters, Construction and Maintenance0.05319
51-9193.00Cooling and Freezing Equipment Operators and Tenders0.05319
53-7121.00Tank Car, Truck, and Ship Loaders0.05319
47-5013.00Service Unit Operators, Oil and Gas0.05319
53-5011.00Sailors and Marine Oilers0.05428
51-9032.00Cutting and Slicing Machine Setters, Operators, and Tenders0.05428
47-2073.00Operating Engineers and Other Construction Equipment Operators0.05428
53-7062.04Recycling and Reclamation Workers0.05618
47-5051.00Rock Splitters, Quarry0.0569
37-3013.00Tree Trimmers and Pruners0.05826
51-9191.00Adhesive Bonding Machine Operators and Tenders0.05917
49-3092.00Recreational Vehicle Service Technicians0.05917
49-9092.00Commercial Divers0.06025
51-6011.00Laundry and Dry-Cleaning Workers0.06232
51-7041.00Sawing Machine Setters, Operators, and Tenders, Wood0.06224
47-2211.00Sheet Metal Workers0.06224
47-2041.00Carpet Installers0.06216
39-5012.00Hairdressers, Hairstylists, and Cosmetologists0.06523
47-2161.00Plasterers and Stucco Masons0.06715
31-1121.00Home Health Aides0.06715
33-9092.00Lifeguards, Ski Patrol, and Other Recreational Protective Service Workers0.06715
33-3011.00Bailiffs0.06715
47-4071.00Septic Tank Servicers and Sewer Pipe Cleaners0.06822
45-2093.00Farmworkers, Farm, Ranch, and Aquacultural Animals0.06822
47-2171.00Reinforcing Iron and Rebar Workers0.0717
53-7061.00Cleaners of Vehicles and Equipment0.07121
47-2043.00Floor Sanders and Finishers0.0717
31-9011.00Massage Therapists0.07114
49-9097.00Signal and Track Switch Repairers0.07114
51-9041.00Extruding, Forming, Pressing, and Compacting Machine Setters, Operators, and Tenders0.07128
49-9094.00Locksmiths and Safe Repairers0.07114
49-9095.00Manufactured Building and Mobile Home Installers0.07114
51-6041.00Shoe and Leather Workers and Repairers0.07427
49-3043.00Rail Car Repairers0.07520
51-4072.00Molding, Coremaking, and Casting Machine Setters, Operators, and Tenders, Metal and Plastic0.07633
51-6031.00Sewing Machine Operators0.07726
51-5113.00Print Binding and Finishing Workers0.07726
49-3051.00Motorboat Mechanics and Service Technicians0.07713
Detailed SOC CodeTitleAvg. LLM ExposureNum. Tasks
51-7042.00Woodworking Machine Setters, Operators, and Tenders, Except Sawing0.07726
53-7063.00Machine Feeders and Offbearers0.07713
51-9195.04Glass Blowers, Molders, Benders, and Finishers0.07919
51-8092.00Gas Plant Operators0.07919
39-2011.00Animal Trainers0.07919
51-4193.00Plating Machine Setters, Operators, and Tenders, Metal and Plastic0.07938
51-9083.00Ophthalmic Laboratory Technicians0.08318
51-2031.00Engine and Other Machine Assemblers0.08312
49-2021.00Radio, Cellular, and Tower Equipment Installers and Repairers0.08330
51-9195.05Potters, Manufacturing0.08324
49-3023.00Automotive Service Technicians and Mechanics0.08330
53-7064.00Packers and Packagers, Hand0.08312
39-5092.00Manicurists and Pedicurists0.08318
51-4033.00Grinding, Lapping, Polishing, and Buffing Machine Tool Setters, Operators, and Tenders, Metal and Plastic0.08318
51-4121.00Welders, Cutters, Solderers, and Brazers0.08330
51-6064.00Textile Winding, Twisting, and Drawing Out Machine Setters, Operators, and Tenders0.08324
51-4122.00Welding, Soldering, and Brazing Machine Setters, Operators, and Tenders0.08629
53-7011.00Conveyor Operators and Tenders0.08723
51-2061.00Timing Device Assemblers and Adjusters0.08817
35-3023.00Fast Food and Counter Workers0.08928
49-2092.00Electric Motor, Power Tool, and Related Repairers0.09039
51-9021.00Crushing, Grinding, and Polishing Machine Setters, Operators, and Tenders0.09122
51-9192.00Cleaning, Washing, and Metal Pickling Equipment Operators and Tenders0.09111
53-3011.00Ambulance Drivers and Attendants, Except Emergency Medical Technicians0.09111
51-6052.00Tailors, Dressmakers, and Custom Sewers0.09122
47-2121.00Glaziers0.09327
49-3091.00Bicycle Repairers0.09416
51-9195.03Stone Cutters and Carvers, Manufacturing0.09416
31-9093.00Medical Equipment Preparers0.09416
29-2055.00Surgical Technologists0.09416
47-4041.00Hazardous Materials Removal Workers0.09521
51-8013.04Hydroelectric Plant Technicians0.09521
51-3092.00Food Batchmakers0.10025
51-4051.00Metal-Refining Furnace Operators and Tenders0.10015
47-2152.00Plumbers, Pipefitters, and Steamfitters0.10030
47-2044.00Tile and Stone Setters0.10025
Detailed SOC CodeTitleAvg. LLM ExposureNum. Tasks
49-9021.00Heating, Air Conditioning, and Refrigeration Mechanics and Installers0.10030
49-3042.00Mobile Heavy Equipment Mechanics, Except Engines0.10020
51-9198.00Helpers–Production Workers0.10334
51-8099.01Biofuels Processing Technicians0.10519
51-9023.00Mixing and Blending Machine Setters, Operators, and Tenders0.10519
31-1131.00Nursing Assistants0.10633
53-6021.00Parking Attendants0.11118
51-9051.00Furnace, Kiln, Oven, Drier, and Kettle Operators and Tenders0.11118
47-2231.00Solar Photovoltaic Installers0.11526
49-9012.00Control and Valve Installers and Repairers, Except Mechanical Door0.11539
51-9194.00Etchers and Engravers0.11730
49-9071.00Maintenance and Repair Workers, General0.11730
51-3093.00Food Cooking Machine Operators and Tenders0.11817
43-5053.00Postal Service Mail Sorters, Processors, and Processing Machine Operators0.11817
51-2022.00Electrical and Electronic Equipment Assemblers0.11817
51-4081.00Multiple Machine Tool Setters, Operators, and Tenders, Metal and Plastic0.11921
51-9012.00Separating, Filtering, Clarifying, Precipitating, and Still Machine Setters, Operators, and Tenders0.11921
49-9011.00Mechanical Door Repairers0.12025
51-8031.00Water and Wastewater Treatment Plant and System Operators0.1258
31-9099.02Endoscopy Technicians0.12512
13-1074.00Farm Labor Contractors0.1258
35-2014.00Cooks, Restaurant0.12520
33-9091.00Crossing Guards and Flaggers0.12512
21-1013.00Marriage and Family Therapists0.12516
29-2053.00Psychiatric Technicians0.12516
29-2012.01Histology Technicians0.1258
49-9031.00Home Appliance Repairers0.12931
35-3023.01Baristas0.13219
49-3011.00Aircraft Mechanics and Service Technicians0.13238
39-9011.01Nannies0.13219
47-5023.00Earth Drillers, Except Oil and Gas0.13330
33-2011.00Firefighters0.13330
51-9082.00Medical Appliance Technicians0.13315
47-2031.00Carpenters0.13829
51-8013.03Biomass Plant Technicians0.13918
53-6032.00Aircraft Service Attendants0.13918
49-9043.00Maintenance Workers, Machinery0.13918
Detailed SOC CodeTitleAvg. LLM ExposureNum. Tasks
51-8021.00Stationary Engineers and Boiler Operators0.14025
33-9032.00Security Guards0.14314
39-3011.00Gambling Dealers0.14321
31-2021.00Physical Therapist Assistants0.14321
51-4192.00Layout Workers, Metal and Plastic0.14314
47-2111.00Electricians0.14321
35-3041.00Food Servers, Nonrestaurant0.14314
51-7031.00Model Makers, Wood0.14314
49-3041.00Farm Equipment Mechanics and Service Technicians0.14314
45-2021.00Animal Breeders0.14624
51-6061.00Textile Bleaching and Dyeing Machine Operators and Tenders0.14624
47-5032.00Explosives Workers, Ordnance Handling Experts, and Blasters0.15030
45-2092.00Farmworkers and Laborers, Crop, Nursery, and Greenhouse0.15030
39-3031.00Ushers, Lobby Attendants, and Ticket Takers0.15223
51-9011.00Chemical Equipment Operators and Tenders0.15223
29-2043.00Paramedics0.15413
51-4191.00Heat Treating Equipment Setters, Operators, and Tenders, Metal and Plastic0.15426
45-4023.00Log Graders and Scalers0.15413
51-4034.00Lathe and Turning Machine Tool Setters, Operators, and Tenders, Metal and Plastic0.15616
27-4015.00Lighting Technicians0.15616
51-4061.00Model Makers, Metal and Plastic0.15616
29-1071.01Anesthesiologist Assistants0.15616
51-4023.00Rolling Machine Setters, Operators, and Tenders, Metal and Plastic0.15819
31-2022.00Physical Therapist Aides0.15819
51-8091.00Chemical Plant and System Operators0.15819
39-4031.00Morticians, Undertakers, and Funeral Arrangers0.15922
39-9011.00Childcare Workers0.16025
29-2056.00Veterinary Technologists and Technicians0.16131
33-9094.00School Bus Monitors0.16718
29-1151.00Nurse Anesthetists0.16724
45-2041.00Graders and Sorters, Agricultural Products0.1676
29-1291.00Acupuncturists0.16718
41-9011.00Demonstrators and Product Promoters0.16724
51-9071.00Jewelers and Precious Stone and Metal Workers0.16730
49-2096.00Electronic Equipment Installers and Repairers, Motor Vehicles0.16712
29-1292.00Dental Hygienists0.16718
51-6062.00Textile Cutting Machine Setters, Operators, and Tenders0.16718
31-9097.00Phlebotomists0.16724
Detailed SOC CodeTitleAvg. LLM ExposureNum. Tasks
49-9099.01Geothermal Technicians0.16724
19-4051.00Nuclear Technicians0.16721
51-3021.00Butchers and Meat Cutters0.16712
39-5011.00Barbers0.17520
51-6063.00Textile Knitting and Weaving Machine Setters, Operators, and Tenders0.17520
47-4021.00Elevator and Escalator Installers and Repairers0.17520
51-7011.00Cabinetmakers and Bench Carpenters0.17520
35-2012.00Cooks, Institution and Cafeteria0.17617
53-5031.00Ship Engineers0.17617
31-1133.00Psychiatric Aides0.17617
49-3053.00Outdoor Power Equipment and Other Small Engine Mechanics0.17914
53-4022.00Railroad Brake, Signal, and Switch Operators and Locomotive Firers0.17928
35-3031.00Waiters and Waitresses0.18025
53-2031.00Flight Attendants0.18025
53-7021.00Crane and Tower Operators0.18211
51-2021.00Coil Winders, Tapers, and Finishers0.18211
51-9071.06Gem and Diamond Workers0.18222
53-7065.00Stockers and Order Fillers0.18330
27-2011.00Actors0.18419
39-3091.00Amusement and Recreation Attendants0.18419
29-1123.00Physical Therapists0.18824
27-1023.00Floral Designers0.18816
25-2051.00Special Education Teachers, Preschool0.19436
39-5094.00Skincare Specialists0.19418
21-1011.00Substance Abuse and Behavioral Disorder Counselors0.19623
51-4062.00Patternmakers, Metal and Plastic0.20015
37-2021.00Pest Control Workers0.20015
47-4099.03Weatherization Installers and Technicians0.20020
49-9064.00Watch and Clock Repairers0.20015
35-3011.00Bartenders0.20020
53-4013.00Rail Yard Engineers, Dinkey Operators, and Hostlers0.20025
53-6031.00Automotive and Watercraft Service Attendants0.20015
29-1126.00Respiratory Therapists0.20522
39-3021.00Motion Picture Projectionists0.20522
51-4111.00Tool and Die Makers0.20617
39-6011.00Baggage Porters and Bellhops0.20617
51-4041.00Machinists0.20729
51-9141.00Semiconductor Processing Technicians0.20824
29-2042.00Emergency Medical Technicians0.20812
53-6061.00Passenger Attendants0.21414
53-7072.00Pump Operators, Except Wellhead Pumpers0.21414
31-9096.00Veterinary Assistants and Laboratory Animal Caretakers0.21428
Detailed SOC CodeTitleAvg. LLM ExposureNum. Tasks
17-3029.08Photonics Technicians0.21730
29-2011.00Medical and Clinical Laboratory Technologists0.21916
53-7081.00Refuse and Recyclable Material Collectors0.21916
49-2098.00Security and Fire Alarm Systems Installers0.21916
33-2022.00Forest Fire Inspectors and Prevention Specialists0.21916
25-2011.00Preschool Teachers, Except Special Education0.22134
53-6011.00Bridge and Lock Tenders0.22218
45-4022.00Logging Equipment Operators0.2229
39-2021.00Animal Caretakers0.22722
29-2092.00Hearing Aid Specialists0.22711
49-2097.00Audiovisual Equipment Installers and Repairers0.22711
51-2092.00Team Assemblers0.22711
51-6093.00Upholsterers0.22722
31-1122.00Personal Care Aides0.22711
51-8093.00Petroleum Pump System Operators, Refinery Operators, and Gaugers0.22924
25-2012.00Kindergarten Teachers, Except Special Education0.23037
35-2013.00Cooks, Private Household0.23113
53-2012.00Commercial Pilots0.23126
39-4021.00Funeral Attendants0.23126
39-9041.00Residential Advisors0.23330
41-2012.00Gambling Change Persons and Booth Cashiers0.23315
53-3052.00Bus Drivers, Transit and Intercity0.23315
31-2012.00Occupational Therapy Aides0.23315
29-2012.00Medical and Clinical Laboratory Technicians0.23315
51-8013.00Power Plant Operators0.23330
51-4035.00Milling and Planing Machine Setters, Operators, and Tenders, Metal and Plastic0.23315
39-7011.00Tour Guides and Escorts0.23719
19-4051.02Nuclear Monitoring Technicians0.23719
39-3092.00Costume Attendants0.23821
21-1021.00Child, Family, and School Social Workers0.23923
29-2061.00Licensed Practical and Licensed Vocational Nurses0.23923
43-9051.00Mail Clerks and Mail Machine Operators, Except Postal Service0.24129
41-2011.00Cashiers0.24129
53-3032.00Heavy and Tractor-Trailer Truck Drivers0.24231
25-2055.00Special Education Teachers, Kindergarten0.25030
53-3053.00Shuttle Drivers and Chauffeurs0.25026
29-1211.00Anesthesiologists0.25018
31-9091.00Dental Assistants0.25016
33-3012.00Correctional Officers and Jailers0.25026
29-1021.00Dentists, General0.25020
53-1042.00First-Line Supervisors of Helpers, Laborers, and Material Movers, Hand0.25024
33-3052.00Transit and Railroad Police0.25014
Detailed SOC CodeTitleAvg. LLM ExposureNum. Tasks
51-7032.00Patternmakers, Wood0.25020
39-3012.00Gambling and Sports Book Writers and Runners0.25022
43-5052.00Postal Service Mail Carriers0.25022
39-9032.00Recreation Workers0.25024
41-9012.00Models0.25014
33-3051.00Police and Sheriff’s Patrol Officers0.25030
41-9091.00Door-to-Door Sales Workers, News and Street Vendors, and Related Workers0.25012
25-4013.00Museum Technicians and Conservators0.25026
35-9031.00Hosts and Hostesses, Restaurant, Lounge, and Coffee Shop0.25022
39-5093.00Shampooers0.2504
51-9161.00Computer Numerically Controlled Tool Operators0.25927
21-1014.00Mental Health Counselors0.25927
29-1128.00Exercise Physiologists0.26025
35-1011.00Chefs and Head Cooks0.26221
47-4011.00Construction and Building Inspectors0.26319
53-5021.00Captains, Mates, and Pilots of Water Vessels0.26730
49-2093.00Electrical and Electronics Installers and Repairers, Transportation Equipment0.26715
27-1012.00Craft Artists0.26715
53-4011.00Locomotive Engineers0.26715
25-9043.00Teaching Assistants, Special Education0.26730
49-9061.00Camera and Photographic Equipment Repairers0.26715
29-1229.04Physical Medicine and Rehabilitation Physicians0.26715
53-7071.00Gas Compressor and Gas Pumping Station Operators0.26913
37-1011.00First-Line Supervisors of Housekeeping and Janitorial Workers0.26926
33-9093.00Transportation Security Screeners0.26926
21-1023.00Mental Health and Substance Abuse Social Workers0.26913
49-2091.00Avionics Technicians0.26913
33-3031.00Fish and Game Wardens0.27124
29-1081.00Podiatrists0.27311
31-2011.00Occupational Therapy Assistants0.27322
53-4031.00Railroad Conductors and Yardmasters0.27322
29-1124.00Radiation Therapists0.27322
51-3011.00Bakers0.27818
17-3027.01Automotive Engineering Technicians0.27818
27-2032.00Choreographers0.27818
27-1026.00Merchandise Displayers and Window Trimmers0.28025
49-2095.00Electrical and Electronics Repairers, Powerhouse, Substation, and Relay0.28116
17-3029.01Non-Destructive Testing Specialists0.28116
49-9041.00Industrial Machinery Mechanics0.28116
29-2011.04Histotechnologists0.28116
53-3054.00Taxi Drivers0.28116
Detailed SOC CodeTitleAvg. LLM ExposureNum. Tasks
33-1011.00First-Line Supervisors of Correctional Officers0.28323
29-1131.00Veterinarians0.28323
45-1011.00First-Line Supervisors of Farming, Fishing, and Forestry Workers0.28330
33-2021.00Fire Inspectors and Investigators0.28330
29-1122.01Low Vision Therapists, Orientation and Mobility Specialists, and Vision Rehabilitation Therapists0.28621
11-3051.03Biofuels Production Managers0.28614
29-2081.00Opticians, Dispensing0.28621
19-4012.00Agricultural Technicians0.28826
49-9062.00Medical Equipment Repairers0.28919
53-1044.00First-Line Supervisors of Passenger Attendants0.28919
53-3031.00Driver/Sales Workers0.29212
17-3021.00Aerospace Engineering and Operations Technologists and Technicians0.29212
29-2033.00Nuclear Medicine Technologists0.29417
29-1122.00Occupational Therapists0.29417
43-5021.00Couriers and Messengers0.29417
45-2011.00Agricultural Inspectors0.29522
51-8011.00Nuclear Power Reactor Operators0.30020
29-1041.00Optometrists0.30010
25-2058.00Special Education Teachers, Secondary School0.30040
29-2032.00Diagnostic Medical Sonographers0.30020
49-2022.00Telecommunications Equipment Installers and Repairers, Except Line Installers0.30040
29-1023.00Orthodontists0.30010
29-2091.00Orthotists and Prosthetists0.30015
25-2057.00Special Education Teachers, Middle School0.30040
29-1129.01Art Therapists0.30025
25-2021.00Elementary School Teachers, Except Special Education0.30338
33-3041.00Parking Enforcement Workers0.30423
29-9091.00Athletic Trainers0.30423
21-1012.00Educational, Guidance, and Career Counselors and Advisors0.30934
21-1094.00Community Health Workers0.31029
29-1141.01Acute Care Nurses0.31527
49-9091.00Coin, Vending, and Amusement Machine Servicers and Repairers0.31619
11-3051.04Biomass Power Plant Managers0.31619
41-2022.00Parts Salespersons0.31619
33-3021.00Detectives and Criminal Investigators0.31730
25-2056.00Special Education Teachers, Elementary School0.31730
29-1125.00Recreational Therapists0.31811
11-3013.00Facilities Managers0.31811
27-2022.00Coaches and Scouts0.32128
Detailed SOC CodeTitleAvg. LLM ExposureNum. Tasks
25-2023.00Career/Technical Education Teachers, Middle School0.32331
53-3051.00Bus Drivers, School0.32417
29-2057.00Ophthalmic Medical Technicians0.32520
33-1012.00First-Line Supervisors of Police and Detectives0.32520
49-2094.00Electrical and Electronics Repairers, Commercial and Industrial Equipment0.32520
29-2035.00Magnetic Resonance Imaging Technologists0.32623
51-5112.00Printing Press Operators0.32623
25-2031.00Secondary School Teachers, Except Special and Career/Technical Education0.32832
53-6051.01Aviation Inspectors0.33315
21-2011.00Clergy0.33321
39-5091.00Makeup Artists, Theatrical and Performance0.33321
19-5011.00Occupational Health and Safety Specialists0.33321
33-3021.02Police Identification and Records Officers0.33315
53-1041.00Aircraft Cargo Handling Supervisors0.3336
43-5051.00Postal Service Clerks0.33321
51-8012.00Power Distributors and Dispatchers0.33315
29-1141.00Registered Nurses0.33327
29-1242.00Orthopedic Surgeons, Except Pediatric0.33315
25-2032.00Career/Technical Education Teachers, Secondary School0.33333
43-5111.00Weighers, Measurers, Checkers, and Samplers, Recordkeeping0.34122
11-3051.06Hydroelectric Production Managers0.34219
27-2023.00Umpires, Referees, and Other Sports Officials0.34416
19-4013.00Food Science Technicians0.34416
29-2099.01Neurodiagnostic Technologists0.34416
33-9011.00Animal Control Workers0.34416
29-1141.03Critical Care Nurses0.34529
29-1212.00Cardiologists0.34626
53-2011.00Airline Pilots, Copilots, and Flight Engineers0.34626
19-1012.00Food Scientists and Technologists0.34613
53-2021.00Air Traffic Controllers0.34823
13-1041.06Coroners0.35020
21-1093.00Social and Human Service Assistants0.35020
11-9072.00Entertainment and Recreation Managers, Except Gambling0.35317
39-7012.00Travel Guides0.35317
33-9021.00Private Detectives and Investigators0.35317
21-1015.00Rehabilitation Counselors0.35317
21-1022.00Healthcare Social Workers0.35317
19-1031.02Range Managers0.35317
19-4071.00Forest and Conservation Technicians0.35721
29-1229.03Urologists0.35714
53-3033.00Light Truck Drivers0.35714
Detailed SOC CodeTitleAvg. LLM ExposureNum. Tasks
37-1012.00First-Line Supervisors of Landscaping, Lawn Service, and Groundskeeping Workers0.35728
25-2022.00Middle School Teachers, Except Special and Career/Technical Education0.35735
49-2011.00Computer, Automated Teller, and Office Machine Repairers0.36025
29-1243.00Pediatric Surgeons0.36118
25-3031.00Substitute Teachers, Short-Term0.36118
17-3024.00Electro-Mechanical and Mechatronics Technologists and Technicians0.36229
53-1043.00First-Line Supervisors of Material-Moving Machine and Vehicle Operators0.36422
25-3021.00Self-Enrichment Teachers0.36730
25-9042.00Teaching Assistants, Preschool, Elementary, Middle, and Secondary School, Except Special Education0.36730
25-9021.00Farm and Home Management Educators0.36715
39-1014.00First-Line Supervisors of Entertainment and Recreation Workers, Except Gambling Services0.36819
27-1027.00Set and Exhibit Designers0.37027
29-2099.05Ophthalmic Medical Technologists0.37131
41-2031.00Retail Salespersons0.37524
19-4031.00Chemical Technicians0.37516
11-9179.01Fitness and Wellness Coordinators0.37524
29-1071.00Physician Assistants0.37512
31-9092.00Medical Assistants0.37520
19-1032.00Foresters0.38025
29-1161.00Nurse Midwives0.38121
33-9099.02Retail Loss Prevention Specialists0.38121
19-4043.00Geological Technicians, Except Hydrologic Technicians0.38330
11-9013.00Farmers, Ranchers, and Other Agricultural Managers0.38330
19-3033.00Clinical and Counseling Psychologists0.38330
29-2034.00Radiologic Technologists and Technicians0.38330
19-2031.00Chemists0.38513
29-1213.00Dermatologists0.38918
19-4021.00Biological Technicians0.38918
29-9099.01Midwives0.38936
43-9071.00Office Machine Operators, Except Computer0.38918
29-1229.06Sports Medicine Physicians0.38927
51-9151.00Photographic Process Workers and Processing Machine Operators0.39132
53-1042.01Recycling Coordinators0.39123
19-1022.00Microbiologists0.39314
19-1023.00Zoologists and Wildlife Biologists0.39314
13-1041.04Government Property Inspectors and Investigators0.39314
19-3034.00School Psychologists0.39519
Detailed SOC CodeTitleAvg. LLM ExposureNum. Tasks
43-5011.00Cargo and Freight Agents0.39624
27-4021.00Photographers0.39729
29-2031.00Cardiovascular Technologists and Technicians0.40025
51-1011.00First-Line Supervisors of Production and Operating Workers0.40020
17-3028.00Calibration Technologists and Technicians0.40015
11-1031.00Legislators0.40030
17-3026.01Nanotechnology Engineering Technologists and Technicians0.40025
27-1022.00Fashion Designers0.40020
29-1218.00Obstetricians and Gynecologists0.40015
47-1011.00First-Line Supervisors of Construction Trades and Extraction Workers0.40015
17-3025.00Environmental Engineering Technologists and Technicians0.40426
19-5012.00Occupational Health and Safety Technicians0.40426
23-1023.00Judges, Magistrate Judges, and Magistrates0.40521
21-1092.00Probation Officers and Correctional Treatment Specialists0.40521
19-4092.00Forensic Science Technicians0.40521
33-1091.00First-Line Supervisors of Security Workers0.40521
29-2052.00Pharmacy Technicians0.40521
11-9151.00Social and Community Service Managers0.40616
51-9061.00Inspectors, Testers, Sorters, Samplers, and Weighers0.40632
27-1013.00Fine Artists, Including Painters, Sculptors, and Illustrators0.40632
13-1111.00Management Analysts0.40911
25-3011.00Adult Basic Education, Adult Secondary Education, and English as a Second Language Instructors0.41039
11-9161.00Emergency Management Directors0.41323
25-1072.00Nursing Instructors and Teachers, Postsecondary0.41429
29-2011.01Cytogenetic Technologists0.41730
11-9111.00Medical and Health Services Managers0.41718
25-1054.00Physics Teachers, Postsecondary0.41724
29-1241.00Ophthalmologists, Except Pediatric0.41718
11-3012.00Administrative Services Managers0.41718
17-2121.00Marine Engineers and Naval Architects0.41730
25-1067.00Sociology Teachers, Postsecondary0.41724
17-3023.00Electrical and Electronic Engineering Technologists and Technicians0.41730
53-4041.00Subway and Streetcar Operators0.41712
17-1012.00Landscape Architects0.42119
21-2021.00Directors, Religious Activities and Education0.42119
11-9032.00Education Administrators, Kindergarten through Secondary0.42232
29-2011.02Cytotechnologists0.42313
Detailed SOC CodeTitleAvg. LLM ExposureNum. Tasks
43-5041.00Meter Readers, Utilities0.42313
29-1215.00Family Medicine Physicians0.42313
53-2022.00Airfield Operations Specialists0.42627
11-9141.00Property, Real Estate, and Community Association Managers0.42627
27-4014.00Sound Engineering Technicians0.42914
47-4011.01Energy Auditors0.42921
25-1052.00Chemistry Teachers, Postsecondary0.42928
29-1051.00Pharmacists0.42921
17-2021.00Agricultural Engineers0.42914
11-9051.00Food Service Managers0.42928
11-9171.00Funeral Home Managers0.43222
23-1011.00Lawyers0.43222
49-1011.00First-Line Supervisors of Mechanics, Installers, and Repairers0.43222
29-1129.02Music Therapists0.43330
27-2012.05Media Technical Directors/Managers0.43315
33-1021.00First-Line Supervisors of Firefighting and Prevention Workers0.43330
19-1031.00Conservation Scientists0.43330
27-2042.00Musicians and Singers0.43330
19-1041.00Epidemiologists0.43315
47-1011.03Solar Energy Installation Managers0.43315
17-3024.01Robotics Technicians0.43523
25-1122.00Communications Teachers, Postsecondary0.43523
27-4031.00Camera Operators, Television, Video, and Film0.43824
33-9031.00Gambling Surveillance Officers and Gambling Investigators0.4388
29-1217.00Neurologists0.43824
11-9199.09Wind Energy Operations Managers0.43816
39-6012.00Concierges0.43816
19-1021.00Biochemists and Biophysicists0.43824
25-1113.00Social Work Teachers, Postsecondary0.43824
41-9022.00Real Estate Sales Agents0.43933
11-9021.00Construction Managers0.44025
51-5111.00Prepress Technicians and Workers0.44025
29-1214.00Emergency Medicine Physicians0.44117
31-9095.00Pharmacy Aides0.44117
43-3041.00Gambling Cage Workers0.44117
19-2011.00Astronomers0.44117
39-1022.00First-Line Supervisors of Personal Service Workers0.44117
11-2022.00Sales Managers0.44117
17-2051.00Civil Engineers0.44117
19-4042.00Environmental Science and Protection Technicians, Including Health0.44226
11-9199.08Loss Prevention Managers0.44427
Detailed SOC CodeTitleAvg. LLM ExposureNum. Tasks
19-3039.03Clinical Neuropsychologists0.44418
17-2111.00Health and Safety Engineers, Except Mining Safety Engineers and Inspectors0.44427
25-1121.00Art, Drama, and Music Teachers, Postsecondary0.44628
17-2141.00Mechanical Engineers0.44628
29-1222.00Physicians, Pathologists0.44719
11-9041.01Biofuels/Biodiesel Technology and Product Development Managers0.44719
29-1216.00General Internal Medicine Physicians0.44719
29-1299.01Naturopathic Physicians0.45020
17-3026.00Industrial Engineering Technologists and Technicians0.45030
29-2036.00Medical Dosimetrists0.45020
29-1141.04Clinical Nurse Specialists0.45030
39-1013.00First-Line Supervisors of Gambling Services Workers0.45030
19-3091.00Anthropologists and Archeologists0.45030
11-1011.00Chief Executives0.45231
43-4181.00Reservation and Transportation Ticket Agents and Travel Clerks0.45221
53-6051.00Transportation Inspectors0.45221
29-1181.00Audiologists0.45522
29-1011.00Chiropractors0.45511
33-3051.04Customs and Border Protection Officers0.45511
25-1063.00Economics Teachers, Postsecondary0.45522
11-9199.11Brownfield Redevelopment Specialists and Site Managers0.45723
13-2099.04Fraud Examiners, Investigators and Analysts0.45723
25-1112.00Law Teachers, Postsecondary0.45723
27-2012.03Media Programming Directors0.45723
29-1223.00Psychiatrists0.45812
29-1127.00Speech-Language Pathologists0.45824
29-1141.02Advanced Practice Psychiatric Nurses0.45824
11-9081.00Lodging Managers0.45824
19-2099.01Remote Sensing Scientists and Technologists0.45824
17-2041.00Chemical Engineers0.46213
17-2199.11Solar Energy Systems Engineers0.46213
13-2051.00Financial and Investment Analysts0.46226
13-1041.01Environmental Compliance Inspectors0.46226
41-4011.07Solar Sales Representatives and Assessors0.46213
25-1042.00Biological Science Teachers, Postsecondary0.46327
19-3039.02Neuropsychologists0.46414
25-1066.00Psychology Teachers, Postsecondary0.46629
11-9033.00Education Administrators, Postsecondary0.46629
27-2012.04Talent Directors0.46715
17-3027.00Mechanical Engineering Technologists and Technicians0.46730
11-2011.00Advertising and Promotions Managers0.46730
Detailed SOC CodeTitleAvg. LLM ExposureNum. Tasks
29-1229.05Preventive Medicine Physicians0.46715
29-1229.01Allergists and Immunologists0.46916
51-6092.00Fabric and Apparel Patternmakers0.46916
19-2032.00Materials Scientists0.46916
19-2042.00Geoscientists, Except Hydrologists and Geographers0.46932
17-2199.10Wind Energy Engineers0.46916
11-3051.02Geothermal Production Managers0.47117
11-9031.00Education and Childcare Administrators, Preschool and Daycare0.47117
15-1211.01Health Informatics Specialists0.47117
11-1021.00General and Operations Managers0.47117
27-1021.00Commercial and Industrial Designers0.47117
29-1221.00Pediatricians, General0.47117
17-2151.00Mining and Geological Engineers, Including Mining Safety Engineers0.47218
13-1023.00Purchasing Agents, Except Wholesale, Retail, and Farm Products0.47419
25-2059.01Adapted Physical Education Specialists0.47520
13-1022.00Wholesale and Retail Buyers, Except Farm Products0.47520
11-9041.00Architectural and Engineering Managers0.47621
17-2131.00Materials Engineers0.47621
11-3051.00Industrial Production Managers0.47621
13-1121.00Meeting, Convention, and Event Planners0.47621
11-3031.00Financial Managers0.47621
53-6051.07Transportation Vehicle, Equipment and Systems Inspectors, Except Aviation0.47621
11-3031.01Treasurers and Controllers0.47722
25-1031.00Architecture Teachers, Postsecondary0.47722
25-1065.00Political Science Teachers, Postsecondary0.47722
25-1071.00Health Specialties Teachers, Postsecondary0.47823
13-1141.00Compensation, Benefits, and Job Analysis Specialists0.47924
25-1081.00Education Teachers, Postsecondary0.47924
53-6041.00Traffic Technicians0.47924
17-2112.03Manufacturing Engineers0.47924
25-1011.00Business Teachers, Postsecondary0.48025
25-1043.00Forestry and Conservation Science Teachers, Postsecondary0.48025
19-3032.00Industrial-Organizational Psychologists0.48025
25-1125.00History Teachers, Postsecondary0.48025
25-1064.00Geography Teachers, Postsecondary0.48025
19-3051.00Urban and Regional Planners0.48025
11-3121.00Human Resources Managers0.48126
25-1051.00Atmospheric, Earth, Marine, and Space Sciences Teachers, Postsecondary0.48126
17-2141.01Fuel Cell Engineers0.48126
19-1013.00Soil and Plant Scientists0.48127
Detailed SOC CodeTitleAvg. LLM ExposureNum. Tasks
29-1171.00Nurse Practitioners0.48127
11-3013.01Security Managers0.48228
43-1011.00First-Line Supervisors of Office and Administrative Support Workers0.48329
27-4011.00Audio and Video Technicians0.48329
11-3071.04Supply Chain Managers0.48330
11-3071.00Transportation, Storage, and Distribution Managers0.48330
29-1224.00Radiologists0.48330
17-2199.06Microsystems Engineers0.48431
17-3031.00Surveying and Mapping Technicians0.50030
11-2021.00Marketing Managers0.50020
17-1022.00Surveyors0.50024
41-1011.00First-Line Supervisors of Retail Sales Workers0.50021
13-2052.00Personal Financial Advisors0.50021
19-2041.00Environmental Scientists and Specialists, Including Health0.50022
25-4011.00Archivists0.50013
35-1012.00First-Line Supervisors of Food Preparation and Serving Workers0.50026
11-3111.00Compensation and Benefits Managers0.50024
27-4032.00Film and Video Editors0.50023
25-4012.00Curators0.50015
19-2043.00Hydrologists0.50025
23-2011.00Paralegals and Legal Assistants0.50012
25-1082.00Library Science Teachers, Postsecondary0.50025
27-2012.00Producers and Directors0.50030
13-2061.00Financial Examiners0.50017
13-1161.00Market Research Analysts and Marketing Specialists0.50013
25-1041.00Agricultural Sciences Teachers, Postsecondary0.50023
13-1051.00Cost Estimators0.50014
19-3092.00Geographers0.50012
11-9131.00Postmasters and Mail Superintendents0.50013
43-5071.00Shipping, Receiving, and Inventory Clerks0.50011
13-1131.00Fundraisers0.50028
19-1029.02Molecular and Cellular Biologists0.50022
19-3094.00Political Scientists0.50014
19-4044.00Hydrologic Technicians0.50016
15-2011.00Actuaries0.50015
19-1031.03Park Naturalists0.50018
29-2099.08Patient Representatives0.50013
29-9092.00Genetic Counselors0.50019
17-2171.00Petroleum Engineers0.50023
23-1012.00Judicial Law Clerks0.50019
17-3013.00Mechanical Drafters0.50015
25-1021.00Computer Science Teachers, Postsecondary0.50026
25-1032.00Engineering Teachers, Postsecondary0.50024
Detailed SOC CodeTitleAvg. LLM ExposureNum. Tasks
13-1081.00Logisticians0.50022
13-2023.00Appraisers and Assessors of Real Estate0.50030
19-1029.04Biologists0.50022
27-4012.00Broadcast Technicians0.50029
23-1021.00Administrative Law Judges, Adjudicators, and Hearing Officers0.50014
41-9021.00Real Estate Brokers0.50019
25-1053.00Environmental Science Teachers, Postsecondary0.50025
43-4081.00Hotel, Motel, and Resort Desk Clerks0.50020
19-1042.00Medical Scientists, Except Epidemiologists0.50014
19-4012.01Precision Agriculture Technicians0.50022
25-1062.00Area, Ethnic, and Cultural Studies Teachers, Postsecondary0.50023
27-1011.00Art Directors0.50016
17-2199.07Photonics Engineers0.50026
11-3031.03Investment Fund Managers0.50020
11-3131.00Training and Development Managers0.50012
19-1011.00Animal Scientists0.5009
27-1025.00Interior Designers0.50015
11-9121.02Water Resource Specialists0.50021
17-2051.01Transportation Engineers0.50026
27-1024.00Graphic Designers0.50016
13-2031.00Budget Analysts0.50013
17-1011.00Architects, Except Landscape and Naval0.50024
19-3093.00Historians0.50021
13-1032.00Insurance Appraisers, Auto Damage0.5007
25-1061.00Anthropology and Archeology Teachers, Postsecondary0.50026
25-1194.00Career/Technical Education Teachers, Postsecondary0.50020
11-3051.01Quality Control Systems Managers0.50027
19-3011.00Economists0.50013
25-9044.00Teaching Assistants, Postsecondary0.50020
25-1111.00Criminal Justice and Law Enforcement Teachers, Postsecondary0.50023
17-2051.02Water/Wastewater Engineers0.50028
23-1022.00Arbitrators, Mediators, and Conciliators0.50020
13-1081.01Logistics Engineers0.51730
11-9199.02Compliance Managers0.51730
29-1031.00Dietitians and Nutritionists0.51828
17-2112.01Human Factors Engineers and Ergonomists0.51926
41-9031.00Sales Engineers0.52025
25-1193.00Recreation and Fitness Studies Teachers, Postsecondary0.52124
19-1029.03Geneticists0.52124
19-2041.02Environmental Restoration Planners0.52223
25-1126.00Philosophy and Religion Teachers, Postsecondary0.52223
Detailed SOC CodeTitleAvg. LLM ExposureNum. Tasks
13-1199.07Security Management Specialists0.52223
19-3099.01Transportation Planners0.52322
11-9179.02Spa Managers0.52421
17-2112.00Industrial Engineers0.52520
33-3021.06Intelligence Analysts0.52520
19-2041.03Industrial Ecologists0.52638
25-3041.00Tutors0.52619
27-2091.00Disc Jockeys, Except Radio0.52619
43-4171.00Receptionists and Information Clerks0.52619
41-1012.00First-Line Supervisors of Non-Retail Sales Workers0.52818
41-4011.00Sales Representatives, Wholesale and Manufacturing, Technical and Scientific Products0.52836
11-3021.00Computer and Information Systems Managers0.52917
11-9121.01Clinical Research Coordinators0.53033
43-4121.00Library Assistants, Clerical0.53033
21-1091.00Health Education Specialists0.53116
29-1299.02Orthoptists0.53116
17-1022.01Geodetic Surveyors0.53116
41-2021.00Counter and Rental Clerks0.53116
13-1041.07Regulatory Affairs Specialists0.53132
11-9199.10Wind Energy Development Managers0.53315
25-9031.00Instructional Coordinators0.53330
13-2054.00Financial Risk Specialists0.53330
25-4022.00Librarians and Media Collections Specialists0.53330
13-1031.00Claims Adjusters, Examiners, and Investigators0.53429
19-2041.01Climate Change Policy Analysts0.53614
17-3011.00Architectural and Civil Drafters0.53628
13-1011.00Agents and Business Managers of Artists, Performers, and Athletes0.53614
17-1021.00Cartographers and Photogrammetrists0.53614
27-1014.00Special Effects Artists and Animators0.53614
19-2021.00Atmospheric and Space Scientists0.53727
19-3041.00Sociologists0.53813
19-4099.01Quality Control Analysts0.53826
29-2051.00Dietetic Technicians0.53813
17-2111.02Fire-Prevention and Protection Engineers0.53813
25-1124.00Foreign Language and Literature Teachers, Postsecondary0.54025
25-1022.00Mathematical Science Teachers, Postsecondary0.54224
13-1021.00Buyers and Purchasing Agents, Farm Products0.54212
25-1192.00Family and Consumer Sciences Teachers, Postsecondary0.54323
17-2199.05Mechatronics Engineers0.54323
17-2071.00Electrical Engineers0.54522
31-9099.01Speech-Language Pathology Assistants0.54511
13-2041.00Credit Analysts0.54511
Detailed SOC CodeTitleAvg. LLM ExposureNum. Tasks
25-1123.00English Language and Literature Teachers, Postsecondary0.54533
17-2072.00Electronics Engineers, Except Computer0.54821
17-2199.03Energy Engineers, Except Wind and Solar0.54821
11-2032.00Public Relations Managers0.55020
13-1151.00Training and Development Specialists0.55020
17-2161.00Nuclear Engineers0.55020
13-2011.00Accountants and Auditors0.55030
41-3011.00Advertising Sales Agents0.55020
13-1082.00Project Management Specialists0.55020
11-9071.00Gambling Managers0.55319
13-1075.00Labor Relations Specialists0.55428
17-2081.00Environmental Engineers0.55428
11-9199.01Regulatory Affairs Managers0.55627
11-1011.03Chief Sustainability Officers0.55618
17-2011.00Aerospace Engineers0.55618
13-1161.01Search Marketing Strategists0.55636
23-2093.00Title Examiners, Abstractors, and Searchers0.55917
13-2081.00Tax Examiners and Collectors, and Revenue Agents0.56025
11-2033.00Fundraising Managers0.56216
11-9121.00Natural Sciences Managers0.56216
41-3041.00Travel Agents0.5628
13-1081.02Logistics Analysts0.56531
43-4051.00Customer Service Representatives0.56715
43-4141.00New Accounts Clerks0.56715
17-3022.00Civil Engineering Technologists and Technicians0.56715
19-4099.03Remote Sensing Technicians0.56822
15-1299.09Information Technology Project Managers0.57121
13-1199.04Business Continuity Planners0.57121
29-1229.02Hospitalists0.57114
13-2022.00Appraisers of Personal and Business Property0.57114
25-4031.00Library Technicians0.57135
13-2053.00Insurance Underwriters0.5717
43-9061.00Office Clerks, General0.57121
41-3021.00Insurance Sales Agents0.57919
43-5011.01Freight Forwarders0.58131
13-2072.00Loan Officers0.58330
17-2199.08Robotics Engineers0.58324
17-3012.00Electrical and Electronics Drafters0.58330
27-3031.00Public Relations Specialists0.58318
43-4131.00Loan Interviewers and Clerks0.58318
11-3061.00Purchasing Managers0.58318
15-1299.02Geographic Information Systems Technologists and Technicians0.58629
43-3071.00Tellers0.58629
15-1232.00Computer User Support Specialists0.59416
Detailed SOC CodeTitleAvg. LLM ExposureNum. Tasks
15-1221.00Computer and Information Research Scientists0.60015
41-3031.00Securities, Commodities, and Financial Services Sales Agents0.60030
17-2199.09Nanosystems Engineers0.60025
19-3011.01Environmental Economists0.60020
17-2031.00Bioengineers and Biomedical Engineers0.60030
43-3061.00Procurement Clerks0.60519
13-1041.03Equal Opportunity Representatives and Officers0.60519
41-4012.00Sales Representatives, Wholesale and Manufacturing, Except Technical and Scientific Products0.60519
43-4161.00Human Resources Assistants, Except Payroll and Timekeeping0.60519
13-1199.05Sustainability Specialists0.60714
15-1299.03Document Management Specialists0.60923
13-2071.00Credit Counselors0.60923
15-1241.01Telecommunications Engineering Specialists0.61526
29-2072.00Medical Records Specialists0.61817
43-5061.00Production, Planning, and Expediting Clerks0.61817
15-1212.00Information Security Analysts0.62512
19-3022.00Survey Researchers0.62516
13-1041.00Compliance Officers0.62516
43-6014.00Secretaries and Administrative Assistants, Except Legal, Medical, and Executive0.62532
15-2031.00Operations Research Analysts0.62516
13-2082.00Tax Preparers0.62512
19-4061.00Social Science Research Assistants0.63023
13-1041.08Customs Brokers0.63023
43-2011.00Switchboard Operators, Including Answering Service0.63219
41-3091.00Sales Representatives of Services, Except Advertising, Insurance, Financial Services, and Travel0.63315
13-1071.00Human Resources Specialists0.63526
15-1231.00Computer Network Support Specialists0.63526
43-6011.00Executive Secretaries and Executive Administrative Assistants0.63622
43-5031.00Public Safety Telecommunicators0.63918
43-9031.00Desktop Publishers0.63918
27-3041.00Editors0.64321
15-2051.02Clinical Data Managers0.64321
43-4111.00Interviewers, Except Eligibility and Loan0.64717
13-1199.06Online Merchants0.64734
27-3023.00News Analysts, Reporters, and Journalists0.65030
43-5032.00Dispatchers, Except Police, Fire, and Ambulance0.65413
29-9021.00Health Information Technologists and Medical Registrars0.65616
43-4041.00Credit Authorizers, Checkers, and Clerks0.65616
15-2041.01Biostatisticians0.66025
Detailed SOC CodeTitleAvg. LLM ExposureNum. Tasks
17-2141.02Automotive Engineers0.66025
15-1255.01Video Game Designers0.66724
15-1241.00Computer Network Architects0.66733
17-2072.01Radio Frequency Identification Device Specialists0.66721
13-2099.01Financial Quantitative Analysts0.66721
27-3011.00Broadcast Announcers and Radio Disc Jockeys0.66724
15-1299.05Information Security Engineers0.67520
43-4061.00Eligibility Interviewers, Government Programs0.67617
43-6012.00Legal Secretaries and Administrative Assistants0.67914
43-4031.00Court, Municipal, and License Clerks0.68330
43-6013.00Medical Secretaries and Administrative Assistants0.68816
17-2061.00Computer Hardware Engineers0.69418
27-3042.00Technical Writers0.70015
43-3011.00Bill and Account Collectors0.70015
43-4011.00Brokerage Clerks0.70010
19-2012.00Physicists0.70015
27-2041.00Music Directors and Composers0.70030
15-1244.00Network and Computer Systems Administrators0.70020
17-2112.02Validation Engineers0.70522
15-2051.01Business Intelligence Analysts0.70617
43-4151.00Order Clerks0.71119
19-1029.01Bioinformatics Scientists0.72520
15-1299.06Digital Forensics Analysts0.72520
43-9022.00Word Processors and Typists0.75020
15-1211.00Computer Systems Analysts0.75022
15-2051.00Data Scientists0.75016
43-4071.00File Clerks0.76319
15-1299.08Computer Systems Engineers/Architects0.76828
51-9162.00Computer Numerically Controlled Tool Programmers0.78116
43-3021.00Billing and Posting Clerks0.78330
43-3031.00Bookkeeping, Accounting, and Auditing Clerks0.78330
15-2041.00Statisticians0.78919
41-9041.00Telemarketers0.79212
15-1299.04Penetration Testers0.79522
43-9111.00Statistical Assistants0.82114
43-9041.00Insurance Claims and Policy Processing Clerks0.83327
43-9021.00Data Entry Keyers0.8339
15-1255.00Web and Digital Interface Designers0.83330
43-3051.00Payroll and Timekeeping Clerks0.84122
15-1243.00Database Architects0.86025
15-1253.00Software Quality Assurance Analysts and Testers0.86730
31-9094.00Medical Transcriptionists0.86715
15-2099.01Bioinformatics Technicians0.86819
15-1252.00Software Developers0.86819
27-3043.00Writers and Authors0.87520
Detailed SOC CodeTitleAvg. LLM ExposureNum. Tasks
27-3043.05Poets, Lyricists and Creative Writers0.87516
27-3091.00Interpreters and Translators0.88217
43-2021.00Telephone Operators0.90616
15-1299.01Web Administrators0.91435
15-1254.00Web Developers0.93129
15-1251.00Computer Programmers0.94117
15-1243.01Data Warehousing Specialists0.94418
15-1242.00Database Administrators0.94418
43-4021.00Correspondence Clerks0.95221
43-9081.00Proofreaders and Copy Markers0.95511
27-3092.00Court Reporters and Simultaneous Captioners0.96414
15-1299.07Blockchain Engineers0.97117
15-2021.00Mathematicians1.00011

Table S1. Average LLM exposure by detailed occupation. Average LLM exposure [2] of each detailed (six-digit) ONET-SOC occupation, sorted from least to most exposed. The final column reports the number of ONET workplace tasks contributing to each average. Table S2 aggregates these scores to the major occupation group level.

Most analyses in this study use llmsoc6\mathit{llm}_{soc6} directly. The exception is the unemployment risk analysis, because the ETA 203 report records only the major occupation group (two-digit SOC code) of each continuing claimant and therefore cannot be linked to detailed occupations. For that analysis we aggregate exposure to the major occupation group level by taking the equal-weighted mean of the detailed exposure scores of the occupations nested within each group,

llmsoc=1∣Ssoc∣∑soc6∈Ssocllmsoc6,(3)\mathit{llm}_{soc} = \frac{1}{\lvert S_{soc}\rvert}\sum_{soc6 \in S_{soc}} \mathit{llm}_{soc6}, \tag*{(3)}

where SsocS_{soc} is the set of detailed SOC codes belonging to major occupation group socsoc. The average is equal-weighted across detailed occupations rather than weighted by employment, so that llmsoc\mathit{llm}_{soc} describes the exposure of the kinds of work performed within a major group rather than the exposure of the particular mix of workers currently employed in it; this keeps the exposure measure fixed over time and independent of the employment counts that enter unemployment risk in Eq. (4). Table S2 reports the resulting llmsoc\mathit{llm}_{soc} for all 22 civilian major occupation groups, together with ∣Ssoc∣\lvert S_{soc}\rvert. These are the values that determine the ordering and coloring of occupations in Figure 1C, the assignment of occupations to exposure quintiles in Figure 1B, and the variable llmsocllm_{\mathrm{soc}} entering Eq. (6). Major group exposure ranges from 0.069 for construction and extraction to 0.741 for computer and mathematical occupations. Analyses repeating our results under alternative AI exposure measures appear in SI § 11.

Major SOC CodeTitleAvg. LLM ExposureNum. Detailed SOC Codes
47Construction and Extraction0.06961
37Building and Grounds Cleaning and Maintenance0.1218
51Production0.128107
49Installation, Maintenance, and Repair0.13350
45Farming, Fishing, and Forestry0.13412
35Food Preparation and Serving Related0.14616
39Personal Care and Service0.20631
53Transportation and Material Moving0.21752
31Healthcare Support0.23919
33Protective Service0.28426
21Community and Social Service0.31814
29Healthcare Practitioners and Technical0.34189
25Educational Instruction and Library0.43362
19Life, Physical, and Social Science0.46160
41Sales and Related0.46222
11Management0.46356
27Arts, Design, Entertainment, Sports, and Media0.47140
23Legal0.4857
17Architecture and Engineering0.49156
13Business and Financial Operations0.52448
43Office and Administrative Support0.60251
15Computer and Mathematical0.74136

Table S2. Average LLM exposure by major occupation group. Average LLM exposure [2] by two-digit SOC major group, computed from Eq. (3) as the equal-weighted mean across the detailed (six-digit) SOC codes within each group, sorted from least to most exposed. The final column reports the number of detailed SOC codes contributing to each average.

Unemployment Risk

The ETA 203 (officially titled “Distribution of Characteristics of the Insured Unemployed”) is a monthly report submitted by state Unemployment Insurance (UI) agencies to the Employment & Training Administration (ETA) of the United States Department of Labor. The ETA 203 is a demographic & characteristics snapshot of continued UI claimants (specifically, the “insured unemployed”) across each state and for the nation as a whole. It is reported on a monthly basis, with data capturing the week including the 12th of the month which corresponds to the similarly-timed survey week for the U.S. Bureau of Labor Statistics (BLS) and U.S. Census’s Current Population Survey (CPS). The data are submitted by states to the national office of ETA via the Office of Workforce Security. Our unemployment risk panel covers January 2015 through October 2025, the most recent month available at the time of this revision, at monthly frequency for each major occupation group in each of the 51 reporting jurisdictions (the 50 states and the District of Columbia). Because the ETA 203 is an administrative tabulation of the insured unemployed rather than a sample survey, its coverage is properly described by how completely states map claimants onto the Standard Occupation Classification System (SOC), not by a sampling error. Across the major occupation group ×\times state ×\times month cells of the January 2020 through October 2025 window tabulated in Table S3, 41 of the 51 reporting jurisdictions record at least one continuing claimant in every major occupation group in every month. All 2,901 zero-claimant cells fall in the ten jurisdictions listed in that table, and two of them—Delaware and California—account for 2,231 of those cells (77%). Zero-claimant cells are fewer than 5% of the cells in that window; excluding Delaware and California, they number 670 in total.

Major Occupation GroupStateDECTCANEKYNHVTAKKSSD
life/physical/social science (19)7025523000000
management (11)7005430000000
transportation (53)7015362000050
healthcare support (31)7022815200000
education (25)7012527500000
food prep/serving (35)003629300000
personal care (39)7045522100000
sales/retail (41)00410300000
production (51)00530400000
computer/math (15)034810100030
construction (47)7005430300000
community/social services (21)70132225200000
legal (23)703746293021700
architecture/engineering (17)704540300000
installation/maintenance/repair (49)70054133580000
healthcare practitioner (29)7003630300000
business/financial operations (13)7004310300000
protective services (33)70175528500000
building/grounds cleaning/maintenance (37)7075418400000
farming/fishing/forestry (45)1755301300000
arts/entertainment (27)7023910100000
office/admin support (43)690540000000

Table S3. Missing data from the ETA 203 report. For each major occupation group and state, the number of months with zero continuing unemployment claimants between January 2020 and October 2025 (i.e., 70 is the maximum). Each cell counts the individual months in which that state reported no continuing claimants for that major occupation group. States with no missing data are omitted.

Two distinguishable patterns generate these zeros. A jurisdiction reporting zero claimants for a given major occupation group in nearly every month of the panel indicates a failure to align that state’s internal occupation classification system with the SOC rather than a genuine absence of claimants; Delaware exhibits this pattern for most occupation groups, and California exhibits it for business/financial operations, management, and education occupations while still reporting counts for computer/math occupations (see Table S3). Isolated zero months in a small state for a small occupation group—in the worst case, four months in which Legal occupations (Major SOC 23) report zero claimants in five states (see Figure S1)—are instead consistent with genuinely small claimant counts. In either case the affected state-occupation pairs are excluded from aggregation rather than imputed. You can download the complete data set here: https://oui.doleta.gov/unemploy/csv/ar203.csv

Figure S1. Number of states with non-zero unemployment claims by major occupation group and month. A state reporting zero claimants may be accurate or may reflect a failure to connect a state’s internal occupation classification system with the Standard Occupation Classification (SOC) used by the U.S. Department of Labor.

ETA 203 Report compared to Census Current Populatiom Survey (CPS) Unemployment

The ETA 203 report (“Characteristics of the Insured Unemployed”) is an administrative unemployment insurance (UI) tabulation compiled from state UI systems that describes the demographic, industry, and occupation characteristics of the insured unemployed. In the ETA 203 framework, the insured unemployed are defined as regular state UI continued weeks claimed (intrastate plus interstate agent), excluding extended and other compensation programs [40]. Importantly for comparisons to survey-based unemployment statistics, ETA guidance aligns the ETA 203 “survey week” to the Current Population Survey (CPS) reference week (typically the week containing the 12th of the month), operationalized through continued weeks claimed filed during the week containing the 19th, and explicitly notes that ETA 203 characteristics can be compared to characteristic figures for the CPS total unemployed [40, 41, 42]. By contrast, CPS unemployment is a population-representative household survey construct rather than a program-based count, and cannot be inferred from UI rolls because many unemployed individuals are ineligible for UI, do not apply, or exhaust benefits [43].

Empirical benchmarks underscore both the value and the limitations of ETA 203–CPS comparisons. Cortes and Forsythe compare CPS-predicted UI recipients to actual UI recipients reported to the U.S. Department of Labor (DOL) via the ETA 203 system during April 2020–February 2021, emphasizing that the DOL administrative compilation aggregates state reports of varying quality and exhibits substantial missingness for some variables (e.g., race and occupation) [44]. Their comparisons show close agreement for some characteristics (e.g., female and Hispanic shares) alongside larger discrepancies for others (e.g., Black and older-age shares), which they interpret as potentially reflecting CPS classification issues in certain contexts and/or reporting differences across sources [44]. Taken together, these resources suggest that ETA 203–CPS comparisons are most defensible when framed explicitly as contrasts between insured unemployment (receipt/claims-based) and total unemployment (labor-force-status-based), with careful attention to timing alignment, program scope (regular UI only), and item-specific data quality in administrative tabulations [40, 44, 43].

This comparison also supplies the standard against which the occupational coverage reported in Section S2 should be judged. First, no survey-based series provides a benchmark at the same resolution: the CPS monthly sample is far too small to support estimates by occupation within state within month, so at the granularity our analysis requires the ETA 203 is the more complete of the two sources, and the relevant contrast remains one between insured and total unemployment rather than between two measurements of the same quantity [43]. Second, the item-specific pattern we observe is the pattern the published assessment of this series reports: Cortes and Forsythe identify occupation and race as precisely the ETA 203 characteristics exhibiting substantial missingness, while other characteristics track their CPS analogues closely [44]. The occupational coverage of our panel is therefore typical of this administrative series rather than anomalous, and—unlike a survey nonresponse problem, which is diffuse—it is concentrated in a small number of identifiable jurisdictions.

Calculating Unemployment Risk

Combined with employment statistics, we calculate an occupation’s unemployment risk according to

p(unemp∣soc,s,t)=4⋅n(unemp∣soc,s,t)4⋅n(unemp∣soc,s,t)+n(emp∣soc,s,t)(4)p(\mathrm{unemp}\mid\mathit{soc},s,t)=\frac{4\cdot n(\mathrm{unemp}\mid\mathit{soc},s,t)}{4\cdot n(\mathrm{unemp}\mid\mathit{soc},s,t)+n(\mathrm{emp}\mid\mathit{soc},s,t)} \tag*{(4)}

where emp indicates a worker is employed, unemp indicates a worker is unemployed, soc\mathit{soc} denotes an occupation’s two-digit Standard Occupation Classification (SOC) code, ss denotes a state, tt denotes a time period (i.e., year and month), and nn captures the number of employed or unemployed workers by occupation, state, and time period. n(unemp∣soc,s,t)n(\mathrm{unemp}\mid\mathit{soc},s,t) comes from the detailed unemployment data while n(emp∣soc,s,t)n(\mathrm{emp}\mid\mathit{soc},s,t) comes from annual BLS OEWS. We multiply n(unemp∣soc,s,t)n(\mathrm{unemp}\mid\mathit{soc},s,t) by 4 to roughly scale from one week of data to one month under the assumption that the week including the 12th of the month is representative of the whole month. Employment by occupation within a state n(emp∣soc,s,t)n(\mathrm{emp}\mid\mathit{soc},s,t) varies annually and so we use it here missing monthly variation. Given a state and time, p(unemp∣soc,s,t)p(\mathrm{unemp}\mid\mathit{soc},s,t) is the probability that a worker with occupation soc\mathit{soc} claims continuing unemployment benefits. Unemployment risk varies by occupation (e.g., construction workers versus transportation workers) while controlling for states’ total labor market size and occupations’ local employment share (e.g., more retail workers may receive unemployment benefits in a state where retail workers are a larger share of employment).

Many studies focus on an occupation’s change in employment following a technological disruption, so one might assume that an occupation’s rising employment would directly correspond to falling unemployment risk. Similarly, a more typical analysis of a state’s total unemployment rate may sufficiently explain variation in unemployment risk as being largely driven by systemic economic factors rather than factors varying by occupation. Using Bayes’ Theorem, we can relate an occupation’s unemployment risk to its local employment share and the state’s total unemployment rate according to

p(unemp∣soc,s,t)=p(soc∣unemp,s,t)⋅p(unemp∣s,t)p(soc∣emp∪unemp,s,t)(5)p(\mathrm{unemp}\mid soc,s,t)=\frac{p(soc\mid\mathrm{unemp},s,t)\cdot p(\mathrm{unemp}\mid s,t)}{p(soc\mid\mathrm{emp}\cup\mathrm{unemp},s,t)} \tag*{(5)}

where p(unemp∣s,t)p(\mathrm{unemp}\mid s,t) is the total unemployment rate from BLS LAUS and p(soc∣emp∪unemp,s,t)p(soc\mid\mathrm{emp}\cup\mathrm{unemp},s,t) is the share of the local labor force associated with socsoc. However, as Figure 1 in [24] shows, an occupation’s employment share over time within a state has no consistent empirical relationship with the occupation’s unemployment risk. And, an occupation-state pair’s unemployment risk co-varies with the state’s total unemployment rate over time but not always.

In Figure 1A, we estimate differences in unemployment risk for low- and high-exposure occupations using the interaction terms βt\beta_t from the following ordinary least squares regression model:

log⁡(p(unemp∣soc,s,t))∼∑t∈Tαt⋅It+βt⋅It⋅llmsoc(6)\log\left(p(\mathrm{unemp}\mid soc,s,t)\right)\sim\sum_{t\in T}\alpha_t\cdot I_t+\beta_t\cdot I_t\cdot llm_{soc} \tag*{(6)}

where TT indexes over quarters each year, llm(soc)∈[0,1]llm(soc)\in[0,1] is the average AI-exposure [2] of the major occupation group soc∈SOCsoc\in\mathrm{SOC}, and ItI_t is an indicator variable for time tt. As a result, 100⋅(eβt−1)100\cdot(e^{\beta_t}-1) describes the percentage change in unemployment risk in a given quarter as llmsocllm_{soc} varies from llmsoc=0llm_{soc}=0 to llmsoc=1llm_{soc}=1. Because unemployment claims are seasonal, we also report this estimate after adjusting for quarter-of-year effects (see Figure S2). The adjustment expresses the estimated high- versus low-exposure contrast in a given quarter as a deviation from the average contrast for that same quarter across all years in the panel. This removes the recurring seasonal component, but it also removes the average level of the contrast, so the vertical scale of Figure S2 is a deviation rather than a level. A value of zero in that figure therefore indicates that the exposure gap in that quarter equals the typical gap for that quarter, not that unemployment risk is equal across high- and low-exposure occupations. Questions about the level of the gap—including whether unemployment risk is higher or lower in LLM-exposed occupations—should be read from Figure 1A, which reports the unadjusted contrast, or from Figure 1B, which reports unemployment risk in levels by exposure quintile. What Figure S2 adds is that the narrowing of the exposure gap beginning in 2022 is not an artifact of seasonality: it persists once quarter-of-year effects are removed. Note also that this adjustment does not control for longer-term annual trends such as the COVID-19 pandemic in 2020 and 2021; the structural break analysis in SI Section S6.1 addresses the pandemic explicitly.

Figure S2. The high- versus low-exposure contrast of Figure 1A after adjusting for quarter-of-year effects. For each quarter, the estimated contrast is expressed as a deviation from the average contrast for that same quarter across all years in the panel. The vertical axis is therefore a deviation and not a level: zero indicates that the exposure gap in that quarter equals the typical gap for that quarter, not that unemployment risk is equal across high- and low-exposure occupations. The level of the gap should be read from Figure 1A or, in levels, from Figure 1B. Note also that long-term annual trends, such as the COVID-19 pandemic in 2020 and 2021, are not controlled for here; see SI § 12.1.

While unemployment risk could in principle also be calculated by sector rather than occupation, we focus on occupation because labor economics research consistently shows that technology reshapes work at the task level, and workplace tasks are organized around occupations rather than industries. The same occupation—for example, secretary—appears across many disparate industries while performing a largely similar set of workplace tasks regardless of industry context; this is reflected in databases such as O*NET (U.S. Bureau of Labor Statistics), which represent each occupation once rather than separately by industry. As a result, a technology that automates the tasks performed by secretaries may register as only a minor factor within any single industry’s aggregate exposure, while still carrying substantial implications for secretaries as an occupational group across the economy. For this reason, the majority of AI exposure studies measure exposure at the occupation and/or task level rather than at the sector level, and we adopt the same approach here.

Unemployment Risk by Major Occupation Group

Figure 1C in the main text shows quarterly unemployment risk for selected major occupation groups. Here we show the same series for all 22 civilian major occupation groups individually (see Figure S3 and Figure S4), so that the trend for any single group can be inspected without the normalization used in the main text. The panels are drawn three to a row across two pages rather than five to a row on one page, so that each series can be read at the size at which it is printed.

Figure S3. Quarterly unemployment risk by major occupation group. Each point is the average unemployment risk across all states and months in each quarter. Major occupation groups 11 through 31; the remaining groups are shown in Figure S4, and the same series are reported as numbers in Table S4.

Figure S4. Quarterly unemployment risk by major occupation group, continued. As in Figure S3, for major occupation groups 33 through 53.

Table S4 reports the same series as numbers. For each major occupation group it gives the group’s average LLM exposure and its average unemployment risk, in percent, over four windows: the pre-pandemic years, the pandemic years, the months of 2022 before the launch of ChatGPT, and the period after the launch. The final column is the change from the last pre-launch window to the post-launch period. Unemployment risk rises after the launch in 20 of the 22 groups, and the increase is largest in the most LLM-exposed groups: +87% for computer and mathematical occupations, +51% for arts and entertainment, and +47% for architecture and engineering, against +5% for construction and −3% for farming, fishing and forestry. Because these are averages over windows rather than a timing analysis, they say nothing about when within a window the increase began; Figure 1B and SI § 12.1 address that question, and both place the turn in early 2022, before the launch.

Major occupation group (SOC)llmllm2015–192020–2120222023–25Change
computer/math (15)0.743.8436.4252.4474.571+87%
office/admin support (43)0.603.3199.1962.6173.205+22%
business/financial operations (13)0.523.1295.7321.9212.666+39%
architecture/engineering (17)0.493.3647.6261.9762.907+47%
legal (23)0.492.6755.6112.1732.857+31%
arts/entertainment (27)0.474.96514.6523.5195.329+51%
management (11)0.467.01014.7094.2495.747+35%
sales/retail (41)0.463.00710.8962.2682.791+23%
life/physical/social science (19)0.463.6187.3062.4363.271+34%
education (25)0.431.4605.3401.0301.285+25%
healthcare practitioner (29)0.341.2615.0151.0011.077+8%
community/social services (21)0.323.0016.9582.2222.524+14%
protective services (33)0.282.2465.9931.7022.119+24%
healthcare support (31)0.243.4938.8982.0822.258+8%
transportation (53)0.224.6319.5352.7973.582+28%
personal care (39)0.212.54917.6562.8012.503-11%
food prep/serving (35)0.152.70515.4952.2472.457+9%
farming/fishing/forestry (45)0.1320.05226.01315.91615.481-3%
installation/maintenance/repair (49)0.133.9909.0502.4312.947+21%
production (51)0.136.85214.4964.2025.340+27%
building/grounds cleaning/maintenance (37)0.124.82812.7933.6514.030+10%
construction (47)0.0711.75116.9088.3618.798+5%

Table S4. Unemployment risk by major occupation group, in percent. The series plotted in Figure S3 and Figure S4, reported as numbers. Each entry is the average of p(unemp∣soc,s,t)p(\mathrm{unemp}\mid\mathrm{soc},s,t) over the state-months of the window, expressed as a percentage, and llmllm is the group’s average LLM exposure [2] as defined in SI Section S1. The pre-launch window ends with 2022Q4 and the post-launch window begins with 2023Q1, following the convention used throughout the paper; the final column is the change between them. Groups are ordered by LLM exposure. The military group (55) is omitted, as elsewhere.

Unemployment risk by State and Major Occupation Group

Two kinds of state can distort the national aggregates in Figure 1B&C. California and Connecticut show quarterly trends visibly unlike most other states, and California is also one of the two jurisdictions responsible for most of the zero-claimant cells documented in Table S3; Delaware is the other, and Nebraska is the next largest contributor of zero-claimant cells. As a robustness check we therefore recompute the national series while omitting each of these states in turn, and then all four together (see Figure S5–Figure S9). In every case the increase in unemployment risk for the most LLM-exposed occupations still begins before the launch of ChatGPT, so the timing reported in the main text is not produced by any one state.

Following the same layout in each figure, the upper panel corresponds to Figure 1C (unemployment risk by major occupation group) and the lower panel to Figure 1B (unemployment risk by LLM exposure quintile).

Figure S6. Same as Figure 1C (upper) and Figure 1B (lower), but omitting data from Connecticut. Each point is the average unemployment risk across all remaining states and months in each quarter.

Figure S7. Same as Figure 1C (upper) and Figure 1B (lower), but omitting data from Delaware. Each point is the average unemployment risk across all remaining states and months in each quarter.

Figure S8. Same as Figure 1C (upper) and Figure 1B (lower), but omitting data from Nebraska. Each point is the average unemployment risk across all remaining states and months in each quarter.

Figure S11. Quarterly unemployment risk by major occupation group in state of AL. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S12. Quarterly unemployment risk by major occupation group in state of AR. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S13. Quarterly unemployment risk by major occupation group in state of AZ. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S15. Quarterly unemployment risk by major occupation group in state of CO. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S16. Quarterly unemployment risk by major occupation group in state of CT. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S17. Quarterly unemployment risk by major occupation group in state of DE. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S18. Quarterly unemployment risk by major occupation group in state of FL. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S19. Quarterly unemployment risk by major occupation group in state of GA. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S20. Quarterly unemployment risk by major occupation group in state of HI. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S21. Quarterly unemployment risk by major occupation group in state of IA. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S22. Quarterly unemployment risk by major occupation group in state of ID. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S23. Quarterly unemployment risk by major occupation group in state of IL. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S24. Quarterly unemployment risk by major occupation group in state of IN. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S25. Quarterly unemployment risk by major occupation group in state of KS. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S26. Quarterly unemployment risk by major occupation group in state of KY. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S27. Quarterly unemployment risk by major occupation group in state of LA. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S28. Quarterly unemployment risk by major occupation group in state of MA. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S29. Quarterly unemployment risk by major occupation group in state of MD. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S30. Quarterly unemployment risk by major occupation group in state of ME. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S31. Quarterly unemployment risk by major occupation group in state of MI. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S32. Quarterly unemployment risk by major occupation group in state of MN. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S33. Quarterly unemployment risk by major occupation group in state of MO. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S34. Quarterly unemployment risk by major occupation group in state of MS. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S35. Quarterly unemployment risk by major occupation group in state of MT. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S36. Quarterly unemployment risk by major occupation group in state of NC. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S37. Quarterly unemployment risk by major occupation group in state of ND. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S38. Quarterly unemployment risk by major occupation group in state of NE. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S39. Quarterly unemployment risk by major occupation group in state of NH. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S40. Quarterly unemployment risk by major occupation group in state of NJ. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S41. Quarterly unemployment risk by major occupation group in state of NM. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S42. Quarterly unemployment risk by major occupation group in state of NV. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S43. Quarterly unemployment risk by major occupation group in state of NY. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S44. Quarterly unemployment risk by major occupation group in state of OH. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S45. Quarterly unemployment risk by major occupation group in state of OK. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S46. Quarterly unemployment risk by major occupation group in state of OR. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S47. Quarterly unemployment risk by major occupation group in state of PA. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Figure S48. Quarterly unemployment risk by major occupation group in state of RI. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Image not extracted

Figure S49. Quarterly unemployment risk by major occupation group in state of SC. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Image not extracted

Figure S50. Quarterly unemployment risk by major occupation group in state of SD. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Image not extracted

Figure S51. Quarterly unemployment risk by major occupation group in state of TN. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Image not extracted

Figure S52. Quarterly unemployment risk by major occupation group in state of TX. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Image not extracted

Figure S53. Quarterly unemployment risk by major occupation group in state of UT. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Image not extracted

Figure S54. Quarterly unemployment risk by major occupation group in state of VA. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Image not extracted

Figure S55. Quarterly unemployment risk by major occupation group in state of VT. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Image not extracted

Figure S57. Quarterly unemployment risk by major occupation group in state of WI. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Image not extracted

Figure S58. Quarterly unemployment risk by major occupation group in state of WV. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Image not extracted

Figure S59. Quarterly unemployment risk by major occupation group in state of WY. Each point is the average unemployment risk across months in each quarter. Lines are shifted to be relative to the first quarter of 2022

Trends in Unemployment Risk using Alternative Exposure Scores

Here we consider some alternative AI exposure scores to the one used in the main text [2]. The first alternative score we use is from Felten et al [26, 1]. That the main results about the early rise of unemployment risk in the main figure remain unchanged (see Figure S60).

Image not extracted

Figure S60. Unemployment risk for AI (LLM) exposed workers increased starting before the launch of ChatGPT using an alternative AI exposure score from Felten et al [26, 1]. (A) Unemployment risk for occupations with high exposure compared to occupations with low exposure. (B) Unemployment risk for occupations grouped into exposure quintiles. (C) Unemployment risk by major occupation. Lines are colored by major occupation’s AI-exposure. We compute unemployment risk monthly across states and report quarterly series as the mean of monthly values across all states within each quarter.

Next, we incorporate actual AI usage data from Claude according to Anthropic’s Economic Index Report [27]. This report includes O*NET tasks inferred from the conversations of actual Claude AI users. However, additional work is needed to turn these statistics into something like a general AI exposure score that can be applied to occupations across the U.S. workforce because Claude users may not represent all workers and Claude users may not represent all LLM users (e.g., AI users may tend toward Claude for certain tasks but other AI systems for other tasks). Additionally, the report characterizes conversations containing each task as either automating or augmenting.

Let T\mathcal{T} denote the set of O*NET work activities (tasks) and let AI∈{0,1}\mathrm{AI} \in\{0,1\} indicate whether an occupation’s task is engaged via an AI system (specifically, Claude). We seek to estimate p(automation∣occ)p(\mathrm{automation} \mid\mathrm{occ}), the probability that a given occupation occ\mathrm{occ} is subject to automative AI use, and p(augmenting∣occ)p(\mathrm{augmenting} \mid\mathrm{occ}), the probability that a given occupation occ\mathrm{occ} is subject to augmenting AI use, drawing on the following empirical distributions:

  • p(task∣occ)p(\mathrm{task} \mid\mathrm{occ}): the distribution over O*NET tasks within an occupation, sourced from the Anthropic Economic Index, which inherits task–occupation mappings from O*NET;

  • p(task∣AI)p(\mathrm{task} \mid\mathrm{AI}): the probability that a given O*NET task is detected in Claude user conversations;

  • p(automation∣task,AI)p(\mathrm{automation} \mid\mathrm{task}, \mathrm{AI}) and p(augmentation∣task,AI)p(\mathrm{augmentation} \mid\mathrm{task}, \mathrm{AI}): conditional on AI use and task detection, the probability that the interaction is classified as automative or augmentative, respectively;

  • p(occ)p(\mathrm{occ}): the marginal occupation distribution, estimated from Bureau of Labor Statistics Occupational Employment and Wage Statistics (BLS OEWS);

  • p(AI∣occ)p(\mathrm{AI} \mid\mathrm{occ}): the probability that workers in a given occupation use AI, approximated via [2]’s AI exposure measure in the absence of ground truth. Note that this is likely an overestimation since exposure to AI only indicates that AI could do something and not that AI will actually be used to do it.

We begin by marginalizing over tasks and AI use:

p(automation∣occ)=∑AI∑taskp(automation∣task,AI)⋅p(task∣AI,occ)⋅p(AI∣occ).(7)p(\mathrm{automation} \mid\mathrm{occ}) = \sum_{\mathrm{AI}} \sum_{\mathrm{task}} p(\mathrm{automation} \mid\mathrm{task}, \mathrm{AI}) \cdot p(\mathrm{task} \mid\mathrm{AI}, \mathrm{occ}) \cdot p(\mathrm{AI} \mid\mathrm{occ}). \tag*{(7)}

Since p(automation∣task,AI=0)=0p(\mathrm{automation} \mid\mathrm{task}, \mathrm{AI}=0)=0 by definition—automation risk is only realized through AI-mediated task engagement—the sum over AI collapses to the AI = 1 term, and we write:

p(automation∣occ)=∑taskp(automation∣task,AI)⋅p(task∣AI,occ)⋅p(AI∣occ).(8)p(\mathrm{automation} \mid\mathrm{occ}) = \sum_{\mathrm{task}} p(\mathrm{automation} \mid\mathrm{task}, \mathrm{AI}) \cdot p(\mathrm{task} \mid\mathrm{AI}, \mathrm{occ}) \cdot p(\mathrm{AI} \mid\mathrm{occ}). \tag*{(8)}

The term p(task∣AI,occ)p(\mathrm{task} \mid\mathrm{AI}, \mathrm{occ}) is not directly observed. Applying Bayes’ rule yields:

p(task∣AI,occ)=p(task∣occ)⋅p(AI∣task,occ)p(AI∣occ).(9)p(\mathrm{task} \mid\mathrm{AI}, \mathrm{occ}) = \frac{p(\mathrm{task} \mid\mathrm{occ}) \cdot p(\mathrm{AI} \mid\mathrm{task}, \mathrm{occ})}{p(\mathrm{AI} \mid\mathrm{occ})}. \tag*{(9)}

We invoke the simplifying assumption that AI adoption and occupation are conditionally independent given task—that is, AI use is mediated entirely through the nature of the task rather than the broader occupational context. That is, p(AI∣task,occ)≈p(AI∣task)p(\mathrm{AI} \mid\mathrm{task}, \mathrm{occ}) \approx p(\mathrm{AI} \mid\mathrm{task}).

Under this assumption:

p(task∣AI,occ)≈p(task∣occ)⋅p(AI∣task)p(AI∣occ).(10)p(\mathrm{task} \mid\mathrm{AI}, \mathrm{occ}) \approx\frac{p(\mathrm{task} \mid\mathrm{occ}) \cdot p(\mathrm{AI} \mid\mathrm{task})}{p(\mathrm{AI} \mid\mathrm{occ})}. \tag*{(10)}

Substituting (S9) into (S6), the p(AI∣occ)p(\mathrm{AI} \mid\mathrm{occ}) terms cancel:

p(automation∣occ)≈∑taskp(automation∣task,AI)⋅p(task∣occ)⋅p(AI∣task).(11)p(\mathrm{automation} \mid\mathrm{occ}) \approx\sum_{\mathrm{task}} p(\mathrm{automation} \mid\mathrm{task}, \mathrm{AI}) \cdot p(\mathrm{task} \mid\mathrm{occ}) \cdot p(\mathrm{AI} \mid\mathrm{task}). \tag*{(11)}

The available data supply p(task∣AI)p(\mathrm{task} \mid\mathrm{AI}) rather than p(AI∣task)p(\mathrm{AI} \mid\mathrm{task}). Applying Bayes’ rule:

p(AI∣task)=p(task∣AI)⋅p(AI)p(task),(12)p(\mathrm{AI} \mid\mathrm{task}) = \frac{p(\mathrm{task} \mid\mathrm{AI}) \cdot p(\mathrm{AI})}{p(\mathrm{task})}, \tag*{(12)}

where the marginals are estimated from the joint distribution over tasks and occupations:

p(task)=∑occp(task∣occ) p(occ),p(AI)=∑occp(AI∣occ) p(occ).(13)p(\mathrm{task}) = \sum_{\mathrm{occ}} p(\mathrm{task} \mid\mathrm{occ})\,p(\mathrm{occ}), \qquad p(\mathrm{AI}) = \sum_{\mathrm{occ}} p(\mathrm{AI} \mid\mathrm{occ})\,p(\mathrm{occ}). \tag*{(13)}

Substituting (S11) into (S10) yields the complete estimator:

p(automation∣occ)≈∑taskp(automation∣task,AI)⋅p(task∣occ)⋅p(task∣AI)⋅p(AI)p(task).(14)p(\mathrm{automation} \mid\mathrm{occ}) \approx\sum_{\mathrm{task}} p(\mathrm{automation} \mid\mathrm{task}, \mathrm{AI}) \cdot p(\mathrm{task} \mid\mathrm{occ}) \cdot\frac{p(\mathrm{task} \mid\mathrm{AI}) \cdot p(\mathrm{AI})}{p(\mathrm{task})}. \tag*{(14)}

(14) expresses p(automation∣occ)p(\mathrm{automation} \mid\mathrm{occ}) entirely in terms of the empirically available distributions enumerated above, under the conditional independence assumption that AI adoption is fully mediated by task content rather than occupational identity. Applying a similar approach yields p(augmenting∣occ)p(\text{augmenting} \mid\text{occ}). Using either p(automation∣occ)p(\text{automation} \mid\text{occ}) or p(augmenting∣occ)p(\text{augmenting} \mid\text{occ}) produces similar results to those presented in Figure 1 using Eloundou et al. (see Figure S61 and Figure S62).

Image not extracted

Figure S61. Unemployment risk for AI (LLM) exposed workers increased starting before the launch of ChatGPT using an alternative AI exposure score for task automation from Anthropic’s Economic Index Report [27]. (A) Unemployment risk for occupations with high exposure compared to occupations with low exposure. (B) Unemployment risk for occupations grouped into exposure quintiles. (C) Unemployment risk by major occupation. Lines are colored by major occupation’s AI-exposure. We compute unemployment risk monthly across states and report quarterly series as the mean of monthly values across all states within each quarter.

Image not extracted

Figure S62. Unemployment risk for AI (LLM) exposed workers increased starting before the launch of ChatGPT using an alternative AI exposure score for task augmentation from Anthropic’s Economic Index Report [27]. (A) Unemployment risk for occupations with high exposure compared to occupations with low exposure. (B) Unemployment risk for occupations grouped into exposure quintiles. (C) Unemployment risk by major occupation. Lines are colored by major occupation’s AI-exposure. We compute unemployment risk monthly across states and report quarterly series as the mean of monthly values across all states within each quarter.

Unemployment Risk By Month

This section recreates unemployment risk time series from the main text with monthly time points instead of quarterly. Figure S63 presents the monthly series using the LLM exposure scores of Eloundou et al. [2] used throughout the main text, and Figure S64–Figure S66 repeat the exercise for each alternative exposure measure considered in SI Section S5: the AI occupational impact scores of Felten et al. [26, 1], and task augmentation and task automation from Anthropic’s Economic Index Report [27]. Across all four exposure measures, the monthly series reproduce the quarterly patterns reported in the main text, indicating that the quarterly aggregation used in Figure 1 is not responsible for those patterns.

Image not extracted

Figure S63. Monthly unemployment risk time series grouping major occupations into AI exposure [2] quintiles (A) or as individual time series (B).

Image not extracted

(a)

Image not extracted

(b)

Figure S64. Monthly unemployment risk time series grouping major occupations into AI exposure [1] quintiles (A) or as individual time series (B).

Image not extracted

Figure S66. Monthly unemployment risk time series grouping major occupations into AI exposure quintiles (A) or as individual time series (B) according to their task automation from Anthropic’s Economic Index Report [27]. See SI Section S5 for exposure calculation.

Image not extracted

Figure S65. Monthly unemployment risk time series grouping major occupations into AI exposure quintiles (A) or as individual time series (B) according to their task augmentation from Anthropic’s Economic Index Report [27]. See SI Section S5 for exposure calculation.

Finding breaks in the unemployment risk time series

To identify structural changes in unemployment insurance (UI) risk over time, we applied a Bai and Perron [45]-style multiple structural change procedure independently to the log⁡10\log_{10}-transformed monthly UI risk series for each quintile of occupational LLM exposure (see Figure S67 and Table S5). For each quintile, we located, for every candidate number of breaks k∈{0,…,10}k \in\{0,\ldots,10\}, the exact partition minimizing the within-regime sum of squared residuals via dynamic programming (ruptures, Dynp, L2L_2 cost), enforcing a minimum regime length of 12 months, and selected kk by minimizing the Bayesian Information Criterion (BIC). All five exposure quintiles independently identify a break in April 2020, consistent with the onset of the COVID-19 pandemic and the associated surge in unemployment insurance claims, a disruption that affected occupations broadly regardless of LLM exposure. All five quintiles additionally identify breaks in April 2021 and April 2022, indicating that labor-market disruption from this period persisted across the entire exposure distribution through at least the early stages of pandemic recovery. By contrast, only the three highest-exposure quintiles exhibit a further, distinct break in 2023 (April–June), following the public launch of ChatGPT in November 2022—a pattern absent from the two lowest-exposure quintiles. However, this late shift is not associated with increasing unemployment risk, but rather flattening or decreasing unemployment risk.

Exposure QuintileDetected Breakpoints (Year-Month)NN BreaksBIC
1 (low)2017-05, 2020-04, 2021-04, 2022-044−518.02-518.02
22016-04, 2017-09, 2020-04, 2021-04, 2022-045−565.64-565.64
32017-09, 2020-04, 2021-04, 2022-04, 2023-065−572.72-572.72
42018-03, 2020-04, 2021-04, 2022-04, 2023-045−600.36-600.36
5 (high)2018-04, 2020-04, 2021-04, 2022-04, 2023-045−626.34-626.34

Table S5. Structural breaks in log⁡10\log_{10} monthly UI risk by LLM exposure quintile, identified via Bai and Perron (2003)-style multiple structural change detection (minimum 12-month regime length, break count selected by BIC).

Image not extracted

Figure S67. Log10_{10} monthly unemployment insurance (UI) risk over time by quintile of occupational LLM exposure. Colored points show, for each exposure quintile, the average log10_{10} UI risk across constituent occupations in each month; darker orange denotes higher LLM exposure. Dotted vertical lines mark structural breaks identified independently within each quintile’s series by a Bai and Perron (2003)-style multiple structural change procedure.

Macroeconomic conditions and unemployment risk

The break analysis above establishes when unemployment risk changed but not what changed it. This section asks how much of the variation in unemployment risk can be accounted for by ordinary macroeconomic conditions, and how much remains for AI exposure and for the launch of ChatGPT to explain. We estimate a sequence of nested ordinary least squares models on the monthly panel of major occupation group ×\times state ×\times month cells described in SI Section S2, covering January 2015 through October 2025. The dependent variable is log⁡10p(unemp∣soc,s,t)\log_{10} p(\mathrm{unemp}\mid soc,s,t) from Eq. (4). The macroeconomic condition we use is the effective federal funds rate fedtfed_t in percentage points [19], which the monetary tightening cycle beginning in March 2022 makes the most salient candidate for the period in question. As above, llmsoc∈[0,1]llm_{soc} \in[0,1] is the major group’s average LLM exposure [2] and gpttgpt_t is an indicator for months from December 2022 onward. All models include state and month-of-year fixed effects, the latter absorbing the seasonality of unemployment claims; the final model adds major occupation group fixed effects. Standard errors are two-way clustered by state and by month. Results appear in Table S6.

Table S6. Macroeconomic conditions, LLM exposure, and unemployment risk. Nested ordinary least squares models of log⁡10\log_{10} monthly unemployment risk across major occupation group ×\times state ×\times month cells, January 2015 through October 2025. fedfed is the effective federal funds rate in percentage points [19], llm∈[0,1]llm \in[0, 1] is the major occupation group’s average LLM exposure [2], and gptgpt indicates months from December 2022 onward. Standard errors are two-way clustered by state and by month. Model 4 omits llm×fedllm \times fed to show the coefficient on llm×gptllm \times gpt that results when the rate cycle is excluded from the interaction structure; every other model is nested in the one that follows it.

VariableModel 1Model 2Model 3Model 4Model 5Model 6Model 7
Federal Interest Rate (fedfed)-0.042***-0.042***-0.119***-0.129***-0.129***-0.129***
LLM Exposure (llmllm)-0.467***-0.537***-0.467***-0.413***
Post ChatGPT Indicator (gptgpt)0.366***0.068***0.068***0.068***
llm×fedllm \times fed0.075***0.076***0.119***0.120***
llm×gptllm \times gpt0.261***-0.110-0.110
gpt×fedgpt \times fed0.117***0.117***0.117***
gpt×fed×llmgpt \times fed \times llm-0.033*-0.034*
State F.E.YesYesYesYesYesYesYes
Month F.E.YesYesYesYesYesYesYes
Major SOC F.E.NoNoNoNoNoNoYes
R2R^20.3250.3500.3790.3950.3990.3990.560
adj. R2R^20.3250.3490.3780.3950.3980.3990.560
pval<0.1∗p_{val} < 0.1^*, pval<0.01∗∗p_{val} < 0.01^{**}, pval<0.001∗∗∗p_{val} < 0.001^{***}

Three findings bear on the timing question. First, the interest rate accounts for a substantial share of the movement in unemployment risk on its own: adding fedfed to the fixed effects alone raises R2R^2 from 0.325 to 0.350 (Models 1–2). Second, the interaction llm×fedllm \times fed is positive and significant (0.075, Model 3), which is to say that the exposure gap in unemployment risk narrows as interest rates rise. The magnitude is large enough to account for the narrowing seen in Figure 1A without reference to AI. Evaluated at a zero federal funds rate, Model 3 places unemployment risk 66% lower in a fully exposed occupation than in an unexposed one (10−0.467=0.3410^{-0.467} = 0.34), consistent with the range reported in the main text; evaluated at a rate of 5%, the same model places it only 19% lower (10−0.092=0.8110^{-0.092} = 0.81). Third, and most directly relevant to whether the launch of ChatGPT marks a change, the two terms that carry an AI-specific post-launch effect—llm×gptllm \times gpt and gpt×fed×llmgpt \times fed \times llm—raise R2R^2 from 0.399 to 0.399, a change of less than 0.001 (Models 5–6). Neither term is individually significant at conventional levels, and both carry negative signs, the opposite of what displacement of exposed workers would predict.

Model 4 is included because it shows what happens when the rate cycle is left out of the interaction structure. It is the specification a researcher would estimate if they suspected ChatGPT and did not suspect monetary policy: it omits llm×fedllm \times fed and retains llm×gptllm \times gpt. In that model llm×gptllm \times gpt is positive and highly significant (0.261), which would ordinarily be read as exposed occupations deteriorating after the launch. Restoring llm×fedllm \times fed in Model 6 drives the same coefficient to −0.110 and to insignificance, and drives the gptgpt indicator from 0.366 to 0.068, a reduction of 81%. The apparent post-ChatGPT penalty for exposed occupations is therefore not robust to controlling for the interest rate environment; it is what the tightening cycle looks like when the tightening cycle is not in the model. Model 7 adds major occupation group fixed effects and leaves every interaction essentially unchanged (llm×fedllm \times fed from 0.119 to 0.120, llm×gptllm \times gpt unchanged at −0.110, gpt×fedgpt \times fed unchanged at 0.117), which indicates that this structure is not an artifact of exposed occupations carrying lower unemployment risk on average. The llmllm main effect is absorbed by those fixed effects and is therefore not reported in Model 7.

We are deliberate about what these models do and do not establish. They are an accounting of variance rather than an identification of causes, and we do not interpret the federal funds coefficient causally: the Federal Reserve raises rates in response to labor market conditions, so the negative pre-launch rate coefficient reflects policy tightening into a strong labor market rather than tightening reducing unemployment. The near-collinearity of gptgpt and fedfed over this window is itself a limitation and, we would argue, the central point. The federal funds rate exceeded 3% in only two months before the launch of ChatGPT (October and November 2022), so the two regressors separate only weakly in these data, and the gpt×fedgpt \times fed terms are identified from a thin region of the sample. This is precisely the identification problem the paper documents: a treatment date placed at the launch of ChatGPT is nearly indistinguishable from the monetary regime that began eight months earlier, and any design that dates treatment to November 2022 will attribute the one to the other. What we can say is that once the rate cycle is in the model, the terms specific to AI exposure after the launch explain no additional variance.

Estimating College-Graduate Job Seeking Times

We first filter to individuals in the Revelio Labs worker profile data who “graduated” (i.e., completed their terminal degree) in 2017 through 2024. Graduation cohorts 2017 through 2024 therefore define the observation period for the worker profile data, and each graduate is followed for up to three years after degree completion. We then filter each individual’s work history to their first job after degree completion restricting to within three years of degree completion. When an exact graduation is not provided, we assume a May graduation in the year of the individual’s terminal degree completion. Each job has an salary estimated by Revelio Labs, which we convert to real 2015 dollars, and an AI-exposure score [2]. Individual ii’s jobDelay⁡i\operatorname{jobDelay}_{i} is the number of months between terminal degree completion and start date of each graduate’s first post-graduation job.

We study early-career outcomes for college graduates by graduation cohort using an event study focused around the launch of ChatGPT in November 2022. We estimate an ordinary least-squares regression model of the form

jobDelay⁡i∼∑g∈G(βg⋅Ig+βg,llm⋅Ig⋅llmi)+βopenings⁡⋅jobOpeningRate⁡i+∑s∈State⁡βs⋅Is+∑n∈NAICS2⁡βn⋅In+∑fos∈FieldOfStudy⁡βfos⋅Ifos+∑d∈Degrees⁡βd⋅Id+∑u∈Universities⁡βU⋅Iu(15)\operatorname{jobDelay}_{i} \sim\sum_{g \in G} \left(\beta_g \cdot I_g + \beta_{g,llm} \cdot I_g \cdot llm_i\right) + \beta_{\operatorname{openings}} \cdot\operatorname{jobOpeningRate}_{i} + \sum_{s \in\operatorname{State}} \beta_s \cdot I_s + \sum_{n \in\operatorname{NAICS2}} \beta_n \cdot I_n + \sum_{fos \in\operatorname{FieldOfStudy}} \beta_{fos} \cdot I_{fos} + \sum_{d \in\operatorname{Degrees}} \beta_d \cdot I_d + \sum_{u \in\operatorname{Universities}} \beta_U \cdot I_u \tag*{(15)}

where β\beta represent regression coefficients, IgI_g is an indicator variable for the year of terminal degree completion, llmillm_i is the LLM exposure of the first job after graduation (i.e., according to six-digit SOC code), jobOpeningRate⁡i\operatorname{jobOpeningRate}_{i} is the job opening rate in the state and month of the job according to BLS JOLTS, IsI_s is an indicator variable for the state of the job, InI_n is an indicator variable for the sector of the job (i.e., two-digit NAICS code), IfosI_{fos} is the field of study of the individual’s terminal degree (i.e., two-digit codes from the Classification of Instructional Programs (CIP) used by the National Center for Education Statistics), IdI_d is an indicator variable for the type of terminals degree (i.e., GED, associates, bachelors, masters, MBA, or doctoral), and IuI_u is an indicator variable for the university or college from which the individual earned their terminal degree (e.g., University of Pittsburgh). Each regression contains 2,665,306 individuals (i.e., rows of data). We conduct an additional but similar regression swapping LLM exposure (llmillm_i) for the estimated salary of the first job (salaryisalary_i). Estimates of both specifications are reported in Table S11.

VariableLLM exposureFirst-job salary
Dependent variable: jobDelay⁡i\operatorname{jobDelay}_{i} (months)
Job opening rate (jobOpeningRate)2.298∗∗∗2.298^{***} (0.005)2.289∗∗∗2.289^{***} (0.005)
I2017×llmI_{2017} \times llm−0.585∗∗∗-0.585^{***} (0.048)
I2018×llmI_{2018} \times llm−0.632∗∗∗-0.632^{***} (0.048)
I2019×llmI_{2019} \times llm−0.982∗∗∗-0.982^{***} (0.048)
I2020×llmI_{2020} \times llm−0.436∗∗∗-0.436^{***} (0.049)
I2021×llmI_{2021} \times llm−0.637∗∗∗-0.637^{***} (0.051)
I2022×llmI_{2022} \times llm−0.911∗∗∗-0.911^{***} (0.057)
I2023×llmI_{2023} \times llm0.486∗∗∗0.486^{***} (0.074)
I2024×llmI_{2024} \times llm0.582∗∗∗0.582^{***} (0.146)
I2017×log⁡(salary)I_{2017} \times\log(salary)1.112∗∗∗1.112^{***} (0.019)
I2018×log⁡(salary)I_{2018} \times\log(salary)1.361∗∗∗1.361^{***} (0.019)
I2019×log⁡(salary)I_{2019} \times\log(salary)1.095∗∗∗1.095^{***} (0.020)
I2020×log⁡(salary)I_{2020} \times\log(salary)0.568∗∗∗0.568^{***} (0.020)
I2021×log⁡(salary)I_{2021} \times\log(salary)0.409∗∗∗0.409^{***} (0.022)
I2022×log⁡(salary)I_{2022} \times\log(salary)0.332∗∗∗0.332^{***} (0.025)
I2023×log⁡(salary)I_{2023} \times\log(salary)0.898∗∗∗0.898^{***} (0.034)
I2024×log⁡(salary)I_{2024} \times\log(salary)0.872∗∗∗0.872^{***} (0.070)
Graduation-year indicators (IgI_g)YesYes
State F.E.YesYes
Sector (NAICS) F.E.YesYes
Field of study (CIP) F.E.YesYes
Degree F.E.YesYes
University F.E.YesYes
R2R^20.1480.151
adj. R2R^20.1470.150
Observations2,665,3062,665,306

Table S11. Table S11: Estimates of the graduate job seeking time models. Ordinary least squares estimates of Eq. S14 (LLM exposure) and of the parallel specification that replaces LLM exposure with the log of the estimated first-job salary, which are the models plotted in Fig. 2A and Fig. 2B. The dependent variable is jobDelay⁡i\operatorname{jobDelay}_{i} in months. Classical standard errors appear in parentheses. LLM exposure enters on its original [0, 1] scale and log⁡(salary)\log(salary) is centered at its sample mean, as described in the text. The graduation-year indicators and the five fixed-effect blocks are entered as complete sets of indicators without an intercept and are reported as included rather than individually; they are dominated by the roughly fifteen hundred university indicators. pval<0.1∗p_{val}<0.1^{*}, pval<0.01∗∗p_{val}<0.01^{**}, pval<0.001∗∗∗p_{val}<0.001^{***}

Because the size of this sample and its coverage bear on how the regression results should be read, we document here how the analytic sample is constructed and what it contains. The processed Revelio extract contains 40,106,792 position records held by 10,584,980 workers with a degree record, each record carrying a Revelio salary estimate, a six-digit SOC code, a NAICS sector, a state, a start date, and the degree record of the worker who holds it. The restrictions described above reduce this to one position per graduate; Table S7 reports the number of position records and the number of distinct graduates surviving each restriction in turn. The binding restrictions are the graduation-cohort window and the requirement that the position begin on or after May of the graduation year, which together remove workers whose terminal degree predates 2017 and positions held before degree completion. The final two rows are worth separating: 2,665,306 graduates hold 3,380,832 positions that begin on their earliest post-graduation start date, because a graduate who begins two positions in the same month contributes more than one record, and we retain one position per graduate. This is the source of the sample size of 2,665,306 that each regression reports: it is the number of graduates, not a rounded or sampled figure.

Table S7: Construction of the analytic sample of recent graduates. Each row applies the restriction named in addition to all restrictions above it, and reports the number of position records and the number of distinct graduates that survive. Position counts exceed graduate counts until the final row because a graduate may hold several qualifying positions.

RestrictionPosition recordsGraduates
Processed Revelio position records40,106,79210,584,980
with a Revelio salary estimate40,106,79210,584,980
with a state-month JOLTS job opening rate40,106,72610,584,967
with a 2020 OES occupation mean wage40,097,57910,583,651
terminal degree completed 2017–202423,093,1005,289,250
position begins within three years of degree completion21,646,0155,249,311
position begins on or after May of the graduation year5,882,2592,665,306
earliest such position3,380,8322,665,306
Analytic sample: one position per graduate2,665,3062,665,306

Table S7.

Table S8 reports the distribution of every continuous variable entering the regression. The median graduate begins their first post-graduation position 3.1 months after May of the graduation year, the mean is 6.0 months, and the distribution is right-skewed, with a 90th percentile of 15.3 months; 13.7% of graduates begin a position in the month of graduation itself. First-job salaries have a median of $51,102 in real 2015 dollars, and first-job LLM exposure [2] has a mean of 0.460 on the unit interval, close to the midpoint of the scale. Table S9 reports the composition of the sample by terminal degree and by graduation cohort. Bachelor’s degree recipients are 69.8% of the sample and master’s recipients a further 21.0%. Cohorts are of similar size from 2017 through 2020 and fall off thereafter, from 469,100 graduates in 2017 to 38,690 in 2024, which is the mechanical consequence of the three-year follow-up window: a graduate of 2024 enters the sample only if they begin a position before the end of the data. The 2024 cohort is therefore selected toward fast job finders, which is visible in its mean job delay of 0.77 months, and we include graduation-year fixed effects in every specification for this reason.

Table S8: Continuous variables in the analytic sample of 2,665,306 recent graduates. jobDelay is measured in months from May of the graduation year. Salaries are the Revelio Labs estimate for the first post-graduation position, converted to real 2015 dollars. LLM exposure is the exposure of the first job’s six-digit SOC code [2]. The job opening rate is the BLS JOLTS rate in the state and month of the position.

VariableMeanS.D.p25p_{25}Medianp75p_{75}p90p_{90}
Months to first job (jobDelay)6.046.661.033.078.1715.27
First-job salary (real 2015 $) | 58,266 | 31,010 | 36,347 | 51,102 | 72,211 | 95,721
Log first-job salary (real 2015 $) | 10.85 | 0.48 | 10.50 | 10.84 | 11.19 | 11.47
First-job LLM exposure (llmllm)0.4600.1980.3500.4670.5560.706
State-month job opening rate (%)5.171.244.204.806.107.00

Table S8.

Graduates% of sampleMean jobDelayMedian salaryMean ℓℓm\ell\ell m
Panel A: terminal degree
Bachelor1,860,68669.85.7847,1860.453
Master560,73321.06.3061,3990.488
Doctor92,7913.55.6974,5580.444
MBA77,1612.98.3365,3770.491
Associate62,3262.39.2137,5360.396
High School11,6090.46.4845,8940.376
Panel B: graduation cohort
2017469,10017.67.0752,3730.473
2018463,46617.46.5252,3860.470
2019454,66817.16.5352,5790.467
2020421,29315.87.2652,6280.459
2021373,34714.05.4550,8900.448
2022285,05210.74.2448,5370.443
2023159,6906.02.9544,4110.432
202438,6901.50.7741,6450.457

Table S9. Composition of the analytic sample by terminal degree and graduation cohort.

Table S10 reports the composition of graduates’ first jobs by field of study, sector, and major occupation group. The sample covers all 50 states, 25 two-digit NAICS sectors, 22 major occupation groups, 335 distinct six-digit SOC codes, 38 two-digit CIP fields of study, and 1,535 degree-granting universities and colleges, with a median of 567 graduates per institution. Coverage is broad but not uniform. The ten largest institutions account for 10.1% of graduates, the largest hundred for 46.5%, and the largest five hundred for 86.2%; the ten largest states account for 58.4%. Business, management, and marketing is the largest field of study at 22.5% of graduates. Mean first-job LLM exposure varies substantially across these groupings, from 0.281 in healthcare practitioner occupations to 0.781 in computer and mathematical occupations, and this variation is what identifies the coefficients on llmllm.

CodeCategoryGraduates% of sampleMean llmllm
Panel A: field of study of terminal degree (CIP)
52Business, management, and marketing599,42022.50.487
51Health professions267,28510.00.342
14Engineering213,0928.00.502
30Multi/interdisciplinary studies207,5907.80.620
45Social sciences155,9505.90.468
09Communication and journalism155,2865.80.516
50Visual and performing arts150,3505.60.446
42Psychology119,9744.50.385
26Biological and biomedical sciences114,4084.30.375
13Education99,8903.70.379
Panel B: sector of first job (NAICS)
54Professional, scientific, and technical services369,49013.90.532
62Health care and social assistance326,87012.30.346
61Educational services315,55911.80.425
51Information234,1028.80.587
92Public administration223,1628.40.411
52Finance and insurance205,2877.70.555
33Manufacturing170,0126.40.511
45Retail trade118,9474.50.429
81Other services96,9203.60.426
71Arts, entertainment, and recreation86,2683.20.387
Panel C: major occupation group of first job (SOC)
11Management300,84711.30.468
13Business and financial operations278,23110.40.535
27Arts, design, entertainment, sports, and media271,77210.20.550
15Computer and mathematical252,6399.50.781
25Educational instruction and library214,7198.10.370
43Office and administrative support206,1857.70.616
29Healthcare practitioners and technical183,0476.90.281
41Sales and related171,3776.40.407
17Architecture and engineering166,1726.20.493
19Life, physical, and social science135,2565.10.491

Table S10. Table S10: Composition of the analytic sample by field of study, sector of first job, and major occupation group of first job. The ten largest categories are shown in each panel. Codes are two-digit CIP codes for field of study, two-digit NAICS codes for sector, and two-digit SOC codes for major occupation group. Mean llmllm is the mean LLM exposure of the first job [2].

On representativeness, we compare the field-of-study distribution of Revelio graduates against the National Survey of College Graduates in SI Section S13, and re-estimate our regressions on a sample reweighted to match the survey. Compared to the NSCG, the Revelio data over-represent business and communications majors and under-represent engineering and education majors, and the reweighted estimates leave our conclusions unchanged (Table S20). That comparison is limited to bachelor’s degree recipients in the biennial survey years and to the field-of-study margin, so it should be read as evidence that our results are not driven by the composition of majors rather than as a general claim of representativeness. Three limitations remain and we state them plainly. Worker profiles are self-reported and are a non-probability sample of the workforce, so graduates who do not maintain a profile are absent and workers in occupations where profiles are less common are under-represented. Salaries are estimated by Revelio Labs rather than observed, which is why we also report time to first job, a quantity that depends only on the dates a graduate records. And the profiles carry no demographic information beyond education and employment history, so we can neither weight on nor report descriptive statistics for characteristics such as sex, race, or parental education.

Dependent Variable:Log Salary (Real 2015 $) | | Months to First Job |
VariableModel 1Model 2Model 3Model 4
Job Opening Rate0.041{sup}***`0.047{sup}***`0.513{sup}***`0.512{sup}***`
Major SOC Log Annual Wage (2020)0.393{sup}***`0.323{sup}***`-0.038{sup}***`-0.022{sup}***`
Post ChatGPT Indicator (gptg_{pt})-0.0120.020{sup}*`3.276{sup}***`3.267{sup}***`
Education Exposure (eduedu)-0.013{sup}***`-0.004{sup}***`0.008{sup}***`0.006{sup}***`
edu×gptedu \times g_{pt}0.021{sup}***`0.006{sup}*`-0.021{sup}***`-0.017{sup}***`
LLM Exposure (llmllm)0.175{sup}***`-0.047{sup}***`
gpt×llmg_{pt} \times llm-0.078{sup}***`0.052{sup}***`
edu×llmedu \times llm-0.037{sup}***`0.006{sup}***`
edu×llm×gptedu \times llm \times g_{pt}0.027{sup}***`-0.015{sup}***`
State FEYesYesYesYes
Sector FEYesYesYesYes
Graduation Year FEYesYesYesYes
College Major FEYesYesYesYes
R2R^20.3600.3690.1650.165
adj. R2R^20.3600.3690.1650.165
Number of Rows: 1,005,481
pval<0.1p_{val} < 0.1, pval<0.01p_{val} < 0.01** , pval<0.001p_{val} < 0.001**

Table S20. Similar to Table 1 where we restrict the user sample to Bachelors degree recipients from graduation cohorts from 2015, 2019, 2021, and 2023 to match the NSCG survey and run OLS regression weighting individuals’ by their major to match graduation distributions over field of study according to NSCG survey. gpt=1g_{pt} = 1 if job starts on or after December 2022. eduedu is education exposure averaged over matched syllabi, a similarity-weighted measure rather than a count of tasks (see SI Sections S12 and S12.1). llmllm is first-job occupation six-digit SOC code LLM exposure [2]. Continuous predictors and dependent variables are centered and standardized (i.e., z-scores), as in Table 1. All models include fixed effects for the state and two-digit NAICS code of the job as well as for graduation year and college field of study (i.e., two-digit CIP Code). We do not include a terminal degree type fixed effect here because the sample is restricted to workers whose terminal degree is a bachelors degree. We also omit the university fixed effect used in Table 1, because the reweighting is defined over graduation year and field of study.

Table S11 reports the estimates of Eq. (15) alongside those of the parallel specification that replaces LLM exposure with the log of the estimated first-job salary. The two columns are the models plotted in Figure 2A and Figure 2B respectively. Both are estimated by ordinary least squares on all 2,665,306 graduates, and standard errors, shown in parentheses, are the classical ordinary least squares standard errors. Two conventions are needed to read the interaction coefficients. First, llmillm_i enters the interaction on its original scale, so lg×llml_g \times llm is the difference in months to first job between a graduate of cohort gg whose first job is fully LLM-exposed (llm=1llm = 1) and one whose first job is unexposed (llm=0\mathit{llm}=0); this is the quantity plotted in Figure 2A. Second, log⁡(salaryi)\log(\mathit{salary}_{i}) is centered at its sample mean before it is interacted, so Ig×log⁡(salary)I_g \times\log(\mathit{salary}) is the additional months to first job associated with a first-job salary one log point above the sample mean for cohort gg. The graduation-year indicators and the five fixed-effect blocks are each entered as complete sets of indicators with no intercept, so their levels are identified only up to constants that offset across blocks; we therefore report them as included rather than as individual coefficients. The interaction terms and the job opening rate are not affected by this and are reported in full. The five fixed-effect blocks contribute the bulk of the estimated parameters, the large majority of them university indicators.

The estimates reproduce the pattern shown in Figure 2. Within every cohort that graduated before the launch of ChatGPT, graduates whose first job was LLM-exposed reached that job sooner than their classmates, by between 0.44 months (2020) and 0.98 months (2019). The sign reverses for the 2023 and 2024 cohorts, for whom a fully exposed first job is associated with 0.49 and 0.58 additional months of job seeking. The salary specification shows no comparable reversal: the coefficient on Ig×log⁡(salary)I_g \times\log(\mathit{salary}) is positive in every cohort and its 2023 and 2024 values (0.90 and 0.87) are close to those of 2017 through 2019 (1.11, 1.36, and 1.09). The job opening rate enters with a coefficient of about 2.3 months per percentage point in both specifications, so the state-month labor market conditions a graduate faces account for more variation in job seeking time than exposure does. Both models explain a modest share of the variance in individual job seeking times, with R2R^2 of 0.148 and 0.151; job delay is dominated by individual circumstances that administrative and profile data do not record, and we make no claim to predict it.

The number of observations is identical across specifications because it is the count of graduates in the analytic sample rather than a draw from it, as Table S7 traces. A sample of this size does raise the question of whether ordinary least squares remains an appropriate model. Consistency and unbiasedness do not deteriorate as NN grows; what changes is that standard errors shrink until statistical significance stops discriminating between findings, so the interpretive burden shifts from significance to magnitude. We therefore read these coefficients in months rather than in stars. Against a mean job delay of 6.04 months with a standard deviation of 6.66 months (Table S8), a pre-launch exposure advantage of roughly half a month to one month, and a post-launch penalty of roughly half a month, are small relative to the dispersion of individual outcomes and are best read as shifts in a cohort-level average. The quantity that carries the paper’s argument is not the size of any single coefficient but the reversal of sign between cohorts, and the year-to-year movement before 2023 (for example from −0.98 in 2019 to −0.44 in 2020) is itself a useful gauge of how much these estimates move for reasons unrelated to LLMs.

To check directly that the estimates are not an artifact of the sample size, we re-estimate both specifications on 50 independent random subsamples of 50,000 graduates each, drawn without replacement from the analytic sample. Each draw is roughly 1.9% of the data and refits the identical specification, including the full set of university indicators, which is thinly populated at this sample size: the median institution contributes about eleven graduates to a draw, and the number of estimated parameters falls by about a fifth, to a mean of 1,403, because a draw contains fewer distinct institutions than the full sample. Draws are seeded so that the results are reproducible. Table S12 reports, for each graduation cohort, the full-sample estimate, the mean and standard deviation of the estimates across draws, the mean of the standard errors estimated within draws, the share of draws whose estimate carries the same sign as the full-sample estimate, and the share significant at the 5% level.

CohortEstimate (S.E.)MeanS.D.Mean S.E.Same signp<0.05p < 0.05
Panel A: graduation cohort × LLM exposure
2017-0.585 (0.048)-0.6240.3610.35594%40%
2018-0.632 (0.048)-0.7410.4180.35596%48%
2019-0.982 (0.048)-1.0220.3010.358100%90%
2020-0.436 (0.049)-0.4350.2900.36492%20%
2021-0.637 (0.051)-0.6420.3220.37698%40%
2022-0.911 (0.057)-0.8980.3570.420100%64%
20230.486 (0.074)0.4010.3800.54990%4%
20240.582 (0.146)0.5410.3671.07692%0%
Panel B: graduation cohort × log salary
20171.112 (0.019)1.1500.1550.144100%100%
20181.361 (0.019)1.3730.1830.144100%100%
20191.095 (0.020)1.0880.1510.145100%100%
20200.568 (0.020)0.5730.1010.150100%100%
20210.409 (0.022)0.4060.1380.160100%68%
20220.332 (0.025)0.3300.1760.185100%38%
20230.898 (0.034)0.9030.1720.252100%100%
20240.872 (0.070)0.8640.1870.512100%26%

Table S12. The job seeking time models re-estimated on random subsamples of 50,000 graduates. Both specifications of Table S11 are refit on 50 independent subsamples of 50,000 graduates drawn without replacement from the analytic sample, using the identical specification and seeded draws. For each graduation cohort the table reports the full-sample estimate with its standard error, the mean and standard deviation of the estimates across draws, the mean of the standard errors estimated within draws, the share of draws whose estimate carries the same sign as the full-sample estimate, and the share of draws in which the estimate is significant at the 5% level with that sign. A sample 53 times smaller implies standard errors 2,665,306/50,000=7.30\sqrt{2{,}665{,}306/50{,}000} = 7.30 times larger; the realized ratios are 7.38 (Panel A) and 7.37 (Panel B). One of the 900 subsample estimates had no computable standard error and is excluded from the standard error column of its row.

Three features of that table answer the question. First, the subsample estimates are centered on the full-sample estimates: across all sixteen interaction terms, the mean across draws differs from the full-sample estimate by at most 0.11 months, and the mean of the 2023 LLM-exposure interaction is 0.40 months against a full-sample value of 0.49. Second, the standard errors behave as sampling theory requires. A sample 53 times smaller should inflate standard errors by 2,665,306/50,000\sqrt{2,665,306/50,000} = 7.30, and the mean within-draw standard error is 7.38 times the full-sample standard error in the LLM-exposure specification and 7.37 times it in the salary specification; the dispersion of estimates across draws is of the same order. The standard errors reported in Table S11 are therefore small because the sample is large, not because the model understates uncertainty. Third, the pattern the paper reports survives resampling even though individual coefficients often do not reach significance in a draw: every pre-2023 LLM-exposure interaction is negative in 82% of draws, the 2023 interaction is positive in 90% and the 2024 interaction in 92%, the 2023 estimate exceeds every pre-2023 estimate in 94% of draws, and the average of the 2023 and 2024 estimates exceeds the average of the 2017 through 2022 estimates in every one of the 50 draws. The salary specification shows no such reversal in any draw.

Two caveats belong with this exercise. Individual significance falls sharply in the subsamples, as it must: the 2023 LLM-exposure interaction is significant at the 5% level in only 4% of draws and the 2024 interaction in none, because an effect of half a month cannot be resolved against a standard error of roughly half a month. This is the expected consequence of estimating the same effect on 1.9% of the data and is the reason we do not rest the argument on any single coefficient. And fitting roughly 1,400 parameters to 50,000 observations makes the design nearly singular in places: R2R^2 rises from 0.148 in the full sample to a mean of 0.174 across draws, which is overfitting rather than improved explanation, and one of the 900 subsample coefficient estimates had no computable standard error. Neither affects the full-sample estimates, which are fit on all 2,665,306 observations.

LLM exposure and occupational wages

The comparison in Figure 2B places first-job salary beside LLM exposure, so it is worth stating how the two are related across occupations rather than leaving a reader to infer it. We take the ONET-occupation exposure scores of SI Section S1 and aggregate them one step further, to six-digit SOC occupations, by an equal-weighted mean across the ONET occupations sharing a SOC code; the wage of an occupation is the mean annual wage published for it in the May 2021 BLS Occupational Employment and Wage Statistics, the same wage source used by Eloundou et al. [2], converted to real 2015 dollars. Of the 923 O*NET occupations carrying an exposure score, 767 match an OES detailed occupation, together covering 130.6 million employed workers; 321 of those occupations are entered by graduates in our sample, accounting for 95.2% of graduates. Table S13 reports the distribution of both variables and Table S14 reports how they move together.

MeanS.D.Minp25p_{25}Medianp75p_{75}Max
Panel A: all 767 occupations
LLM exposure (llmllm)0.3260.2200.0000.1270.3330.5001.000
Mean annual wage (real 2015 $) | 60,619 | 38,735 | 21,993 | 37,207 | 50,070 | 71,604 | 309,414
Panel B: the 321 occupations entered by graduates
LLM exposure (llmllm)0.4180.2140.0000.2590.4620.5500.955
Mean annual wage (real 2015 $) | 62,350 | 32,525 | 22,727 | 38,383 | 54,764 | 83,444 | 258,925
Median first-job salary (real 2015 $) | 52,726 | 17,354 | 22,264 | 40,018 | 48,775 | 64,248 | 110,452

Table S13. Distributions of LLM exposure and of occupational wages. Panel A covers the 767 detailed occupations that carry both an LLM exposure score [2] and a mean annual wage in the May 2021 BLS Occupational Employment and Wage Statistics, together employing 130.6 million workers. Panel B restricts to the 321 of those occupations entered by graduates in our sample, which account for 95.2% of graduates, and adds the median Revelio first-job salary of the graduates entering each occupation. Wages and salaries are in real 2015 dollars. Each row describes the distribution across occupations, not across workers.

GroupOccupationsEmployment (millions)Median wageMean llm\mathit{llm}
Panel A: by LLM exposure band
0.0–0.116520.038,5750.041
0.1–0.29420.239,4490.147
0.2–0.39514.144,0030.247
0.3–0.48718.055,4110.350
0.4–0.513218.375,3020.445
0.5–0.612124.170,2710.528
0.6–0.7389.951,5030.637
0.7–1.0356.154,5100.824
Panel B: by decile of the occupation’s mean annual wage
17730.327,0540.182
27719.632,3340.219
3769.637,2070.237
47712.241,3640.270
5777.147,0980.253
6767.553,5360.315
77713.460,3760.389
87610.071,6040.449
9778.986,3290.501
107711.9120,6290.446

Table S14. Panel A groups occupations by their LLM exposure score [2] and Panel B by decile of their mean annual wage. Employment and wages are from the May 2021 BLS Occupational Employment and Wage Statistics; wages are converted to real 2015 dollars. Median wage is the median across the occupations in the group, so each occupation counts once regardless of size; the employment column reports how many workers those occupations employ. Median wage rises with exposure through the 0.4–0.5 band and falls thereafter, and mean exposure rises across wage deciles through the ninth and falls in the tenth.

Exposure and pay are positively but only moderately related. Across the 767 occupations, the correlation between exposure and the log of the mean annual wage is 0.417 (Spearman 0.484), and weighting occupations by employment raises it to 0.485. Within the occupations our graduates enter the correlation is 0.376, or 0.486 when occupations are weighted by the number of graduates entering them, and at the level of individual graduates the correlation between first-job exposure and log first-job salary is 0.339. For the same 321 occupations, the OES mean annual wage and the median Revelio first-job salary correlate 0.751, which is a useful external check on the salary estimates discussed above: the two measures order occupations similarly although one is an administrative average over all incumbents and the other an estimate for new graduates.

The relationship is also not monotonic at the top of either variable. Mean exposure rises across wage deciles from 0.182 in the lowest decile to 0.501 in the ninth and then falls to 0.446 in the highest (Table S14, Panel B). Read the other way, median wage rises with exposure up to the 0.4–0.5 band, where it peaks at $75,302, and then falls back to $54,510 among the most exposed occupations, close to the $50,070 median across all occupations (Panel A). The most exposed occupations are therefore middle-earning rather than top-earning, which is the pattern Eloundou et al. describe when they note that their data contain both low-wage occupations with high exposure and high-wage occupations with low exposure alongside an overall positive trend. Our aggregation to the occupation level is an equal-weighted mean across tasks, whereas theirs gives core tasks twice the weight of supplemental ones, so small differences from their figure are expected.

This moderate correlation is what makes the salary comparison in Figure 2B informative rather than redundant. If exposure and pay were close to interchangeable, the salary panel would restate the exposure panel and its different behavior after 2022 could not be interpreted. Because the two orderings of occupations agree only in part, the two panels are genuinely distinct comparisons, and the fact that the exposure comparison reverses after 2022 while the salary comparison returns to its pre-pandemic level is informative about which margin moved.

Table S9: Composition of the analytic sample by terminal degree and graduation cohort.

Salary is the median first-job salary in real 2015 dollars; LLM exposure is the mean exposure of the first job [2]. Cohort sizes decline after 2020 because each graduate is followed for at most three years, so later cohorts enter the sample only if they begin a position before the end of the data.

Tracking Significant Difference in Graduation Cohort Job Gaps

To determine whether the relative change in job-start shares for LLM-exposed jobs differs significantly across graduation cohorts, we construct a month-by-month test based on a Delta-method approximation of the sampling variability of the statistic of interest. For a given graduation cohort gg in month tt, let p(g∣t,llm)p(g \mid t, llm) denote the share of LLM-exposed jobs started by graduates of class gg, and let p(g∣t)p(g \mid t) denote the corresponding share across all jobs. The quantity of interest is the proportional difference

jobGap⁡g,llm(t)=p(g∣t,llm)−p(g∣t)p(g∣t)\operatorname{jobGap}_{g,llm}(t) = \frac{p(g \mid t,llm)-p(g \mid t)}{p(g \mid t)}

which measures how over- or under-represented graduates of class gg are in newly started LLM-exposed jobs relative to the overall job-start distribution. Because p(g∣t,llm)p(g \mid t,llm) and p(g∣t)p(g \mid t) are estimated from finite samples—specifically, binomial proportions based on the counts of LLM-exposed jobs and all jobs, respectively—we approximate their joint sampling distribution using a multivariate normal form and apply the Delta method to obtain an analytic expression for Var⁡(jobGap⁡g,llm(t))\operatorname{Var}(\operatorname{jobGap}_{g,llm}(t)). Under an independence assumption between the two estimates, this variance reduces to a closed-form expression involving the binomial variances of p^(g∣t,llm)\hat{p}(g \mid t,llm) and p^(g∣t)\hat{p}(g \mid t).

Having obtained Var⁡^(jobGap⁡g,llm(t))\widehat{\operatorname{Var}}(\operatorname{jobGap}_{g,llm}(t)) for each graduation cohort and month, we test whether two cohorts g1g_1 and g2g_2 differ significantly by evaluating the null hypothesis H0:jobGap⁡g1,llm(t)=jobGap⁡g2,llm(t)H_0 : \operatorname{jobGap}_{g_1,llm}(t) = \operatorname{jobGap}_{g_2,llm}(t) for each month tt. The test statistic is a standard zz-score,

zg1,g2,llm(t)=jobGap⁡g1,llm(t)−jobGap⁡g2,llm(t)Var⁡^(jobGap⁡g1,llm(t))+Var⁡^(jobGap⁡g2,llm(t))z_{g_1,g_2,llm}(t) = \frac{\operatorname{jobGap}_{g_1,llm}(t)-\operatorname{jobGap}_{g_2,llm}(t)}{\sqrt{\widehat{\operatorname{Var}}(\operatorname{jobGap}_{g_1,llm}(t))+\widehat{\operatorname{Var}}(\operatorname{jobGap}_{g_2,llm}(t))}}

which compares the estimated difference in representation against its sampling uncertainty. The resulting two-sided pp-value identifies the months in which the two graduation cohorts differ significantly in how strongly they are represented among new LLM-exposed job starts. This approach provides a transparent, nonparametric month-by-month comparison that directly targets the statistic used in descriptive figures while maintaining principled statistical inference through variance estimation and standard error propagation. We report the resulting pp-values separately for each decile of first-job LLM exposure, from the least exposed decile to the most exposed (see Figure S68 through Figure S77). Significant differences between the 2021 and 2022 graduation cohorts and earlier reference cohorts concentrate in the most exposed deciles.

Image not extracted

Figure S68. Two-proportion z-tests per month (with multiple-testing correction) comparing graduate cohort’s jobGap by month after May of their graduation year for jobs with LLM exposure decile 1. Lines are colored by reference graduation cohort which are selected to be ones that did not experience the launch of ChatGPT during their early years in the workforce post-graduation.

Image not extracted

Figure S77. Two-proportion z-tests per month (with multiple-testing correction) comparing graduate cohort’s jobGap by month after May of their graduation year for jobs with LLM exposure decile 10. Lines are colored by reference graduation cohort which are selected to be ones that did not experience the launch of ChatGPT during their early years in the workforce post-graduation.

Image not extracted

Figure S69. Two-proportion z-tests per month (with multiple-testing correction) comparing graduate cohort’s jobGap by month after May of their graduation year for jobs with LLM exposure decile 2. Lines are colored by reference graduation cohort which are selected to be ones that did not experience the launch of ChatGPT during their early years in the workforce post-graduation.

Image not extracted

Figure S70. Two-proportion z-tests per month (with multiple-testing correction) comparing graduate cohort’s jobGap by month after May of their graduation year for jobs with LLM exposure decile 3. Lines are colored by reference graduation cohort which are selected to be ones that did not experience the launch of ChatGPT during their early years in the workforce post-graduation.

Image not extracted

Figure S71. Two-proportion z-tests per month (with multiple-testing correction) comparing graduate cohort’s jobGap by month after May of their graduation year for jobs with LLM exposure decile 4. Lines are colored by reference graduation cohort which are selected to be ones that did not experience the launch of ChatGPT during their early years in the workforce post-graduation.

Image not extracted

Figure S72. Two-proportion z-tests per month (with multiple-testing correction) comparing graduate cohort’s jobGap by month after May of their graduation year for jobs with LLM exposure decile 5. Lines are colored by reference graduation cohort which are selected to be ones that did not experience the launch of ChatGPT during their early years in the workforce post-graduation.

Image not extracted

Figure S73. Two-proportion z-tests per month (with multiple-testing correction) comparing graduate cohort’s jobGap by month after May of their graduation year for jobs with LLM exposure decile 6. Lines are colored by reference graduation cohort which are selected to be ones that did not experience the launch of ChatGPT during their early years in the workforce post-graduation.

Image not extracted

Figure S74. Two-proportion z-tests per month (with multiple-testing correction) comparing graduate cohort’s jobGap by month after May of their graduation year for jobs with LLM exposure decile 7. Lines are colored by reference graduation cohort which are selected to be ones that did not experience the launch of ChatGPT during their early years in the workforce post-graduation.

Image not extracted

Figure S75. Two-proportion z-tests per month (with multiple-testing correction) comparing graduate cohort’s jobGap by month after May of their graduation year for jobs with LLM exposure decile 8. Lines are colored by reference graduation cohort which are selected to be ones that did not experience the launch of ChatGPT during their early years in the workforce post-graduation.

Image not extracted

Figure S76. Two-proportion z-tests per month (with multiple-testing correction) comparing graduate cohort’s jobGap by month after May of their graduation year for jobs with LLM exposure decile 9. Lines are colored by reference graduation cohort which are selected to be ones that did not experience the launch of ChatGPT during their early years in the workforce post-graduation.

Graduation Cohort Job Gaps by Occupations’ AI-Exposure Decile

This section disaggregates the job gap statistic jobGapg,llm(t)jobGap_{g,llm}(t) defined in SI Section S8 by decile of first-job LLM exposure [2] rather than by quintile. For each decile, we plot the share of jobs started by each graduation cohort in each month since May of the graduation year against the overall share of jobs started in that month (see Figure S79 through Figure S88). The under-representation of the 2021 and 2022 cohorts grows across deciles, and is largest in the most LLM-exposed occupations. Each figure pairs the month-by-month series in panel A with a summary in panel B that averages the same statistic over the months before and after the launch of ChatGPT, so that the timing of a change can be read directly rather than traced along the series. Figure S78 collects those summaries for all ten deciles in one place.

Image not extracted

Figure S78. Summary of the decile figures in this section. Each cell reports the difference in mean jobGapg,llm(t)jobGap_{g,\mathrm{llm}}(t) between a graduation cohort and the 2013–2020 cohorts averaged over the same months since their own graduation, in percentage points, for the decile named on the horizontal axis. The first two rows of each block are the two windows shown in panel B of the corresponding decile figure, and the third is their difference, that is, how much the cohort’s gap moves at the launch of ChatGPT. Stars mark two-sample tt-tests of the cell against zero at the 5%, 1%, and 0.1% levels. Each column carries a fixed occupational composition, so a column’s level reflects which occupations fall in that decile as much as its exposure; the comparison to read is within a column, before against after, and the change rows report exactly that.

Image not extracted

Figure S79. (A) For each graduation cohort by month from May of their graduation year, the share of jobs with AI-exposure decile 1 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S88. (A) For each graduation cohort by month from May of their graduation year, the share of jobs with AI-exposure decile 10 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S80. (A) For each graduation cohort by month from May of their graduation year, the share of jobs with AI-exposure decile 2 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S81. (A) For each graduation cohort by month from May of their graduation year, the share of jobs with AI-exposure decile 3 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S82. (A) For each graduation cohort by month from May of their graduation year, the share of jobs with AI-exposure decile 4 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S83. (A) For each graduation cohort by month from May of their graduation year, the share of jobs with AI-exposure decile 5 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S84. (A) For each graduation cohort by month from May of their graduation year, the share of jobs with AI-exposure decile 6 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S85. (A) For each graduation cohort by month from May of their graduation year, the share of jobs with AI-exposure decile 7 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S86. (A) For each graduation cohort by month from May of their graduation year, the share of jobs with AI-exposure decile 8 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S87. (A) For each graduation cohort by month from May of their graduation year, the share of jobs with AI-exposure decile 9 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Graduation Cohort Job Gaps by Occupations’ Salary Decile

As a placebo comparison for SI § 15, we repeat the same disaggregation using deciles of first-job salary (in real 2015 dollars) in place of deciles of LLM exposure (see Figure S90 through Figure S99). If recent graduation cohorts were simply losing ground in better-paying jobs, the pattern would track the salary deciles as clearly as it tracks the exposure deciles. As in SI § 15, panel A of each figure gives the month-by-month series and panel B averages the same statistic over the months before and after the launch of ChatGPT, and Figure S89 collects those summaries for all ten deciles. It does track the salary deciles: in the months after the launch the 2022 cohort is significantly under-represented in salary deciles four through ten, by between 9.5 and 16.0 percentage points, a cleaner gradient than the exposure deciles produce. The cohort shortfall is therefore not specific to LLM-exposed work, a point we return to in the main text.

Image not extracted

Figure S89. Summary of the decile figures in this section. As in Figure S78, but for deciles of first-job salary. Each cell reports the difference in mean jobGapjobGap between a graduation cohort and the 2013–2020 cohorts over the same months since their own graduation, in percentage points; the third row of each block is the change in that difference at the launch of ChatGPT. Stars mark two-sample tt-tests at the 5%, 1%, and 0.1% levels.

Image not extracted

Figure S90. (A) For each graduation cohort by month from May of their graduation year, the share of jobs with salary decile 1 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. All salaries are converted to real 2015 dollars to account for inflation. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S99. (A) For each graduation cohort by month from May of their graduation year, the share of jobs with salary decile 10 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. All salaries are converted to real 2015 dollars to account for inflation. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S91. (A) For each graduation cohort by month from May of their graduation year, the share of jobs with salary decile 2 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. All salaries are converted to real 2015 dollars to account for inflation. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S92. (A) For each graduation cohort by month from May of their graduation year, the share of jobs with salary decile 3 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. All salaries are converted to real 2015 dollars to account for inflation. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S93. (A) For each graduation cohort by month from May of their graduation year, the share of jobs with salary decile 4 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. All salaries are converted to real 2015 dollars to account for inflation. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S94. (A) For each graduation cohort by month from May of their graduation year, the share of jobs with salary decile 5 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. All salaries are converted to real 2015 dollars to account for inflation. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S95. (A) For each graduation cohort by month from May of their graduation year, the share of jobs with salary decile 6 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. All salaries are converted to real 2015 dollars to account for inflation. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S96. (A) For each graduation cohort by month from May of their graduation year, the share of jobs with salary decile 7 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. All salaries are converted to real 2015 dollars to account for inflation. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S97. (A) For each graduation cohort by month from May of their graduation year, the share of jobs with salary decile 8 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. All salaries are converted to real 2015 dollars to account for inflation. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S98. (A) For each graduation cohort by month from May of their graduation year, the share of jobs with salary decile 9 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. All salaries are converted to real 2015 dollars to account for inflation. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Monthly Graduation Cohort Job Shares

The pointwise difference test is a simple and widely used method for determining whether two time series of probabilities differ at specific points in time. We use it here as a robustness check to compare the job market performance by graduation cohort based on jobs started each month since May of their graduation year (i.e., similar to Figure 2). For a given graduation cohort gg, we create a time series p(g∣t,llm)p(g|t,\mathit{llm}) representing the share of jobs started in a given month tt (i.e., number of months since May graduation year) with LLM exposure [2] quintile llm\mathit{llm}. That is, p(g∣t,llm)=xg,t,llm/ntg,llmp(g|t,\mathit{llm}) = x_{g,t,\mathit{llm}}/n_{t_g,\mathit{llm}} where xg,t,llmx_{g,t,\mathit{llm}} is the number of jobs started in month tt within LLM exposure quintile llm\mathit{llm} by members of graduation cohort gg and ntg,llmn_{t_g,\mathit{llm}} is the total number of jobs within LLM exposure quintile llm\mathit{llm} started in month tgt_g by anybody. Here, we index the month tgt_g with graduation cohort gg because the month tgt_g is relative to the graduation cohort gg; that is, t2016=2t_{2016} = 2 is June 2016, two months from May of the graduation year 2016, while t2022=2t_{2022} = 2 is June 2022. We construct a time series in this way for each graduation cohort from 2013 through 2023.

At each month from May graduation year, we compare the time series for each pre-ChatGPT graduation cohort (i.e., 2013 through 2020) to graduation cohorts that experienced ChatGPT’s launch while they were new to the labor market (i.e., 2021, 2022, 2023). For each pair of time series p(g1∣t,llm)p(g_1|t,\mathit{llm}) and p(g2∣t,llm)p(g_2|t,\mathit{llm}), we calculate the pointwise difference statistic according to

zg1,g2,t,llm=p(g1∣t,llm)−p(g2∣t,llm)pt,llm(1−pt,llm)(1ntg1,llm+1ntg2,llm)(16)z_{g_1,g_2,t,\mathit{llm}} = \frac{p(g_1|t,\mathit{llm}) - p(g_2|t,\mathit{llm})}{\sqrt{p_{t,\mathit{llm}}(1-p_{t,\mathit{llm}})\left(\frac{1}{n_{t_{g_1},\mathit{llm}}}+\frac{1}{n_{t_{g_2},\mathit{llm}}}\right)}} \tag*{(16)}

where

pt,llm=xg1,t,llm+xg2,t,llmntg1,llm+ntg2,llm(17)p_{t,\mathit{llm}} = \frac{x_{g_1,t,\mathit{llm}} + x_{g_2,t,\mathit{llm}}}{n_{t_{g_1},\mathit{llm}} + n_{t_{g_2},\mathit{llm}}} \tag*{(17)}

is the pooled proportion. zg1,g2,t,llmz_{g_1,g_2,t,\mathit{llm}} measures how many standard errors apart the two observed probabilities are. A value of zg1,g2,t,llmz_{g_1,g_2,t,\mathit{llm}} near zero indicates that the difference between the two estimates is small relative to sampling noise, while large positive or negative values indicate stronger evidence against the null.

Interpreting the statistic proceeds in the same way as any standard normal zz-test: one computes a one-sided pp-value to test if p(g2∣t,llm)p(g_2|t,\mathit{llm}) (here, g2∈{2021,2022,2023}g_2 \in\{2021, 2022, 2023\}) is statistically significantly lower than p(g1∣t,llm)p(g_1|t,\mathit{llm}) according to pvalue=1−Φ(∣zt∣)p_{\mathit{value}} = 1 - \Phi(|z_t|), where Φ(⋅)\Phi(\cdot) is the standard normal cumulative distribution function. If this pp-value falls below a predetermined significance level (e.g., 0.05), then the difference at time tt is deemed statistically significant. Because the test is conducted independently at each time point, we plot the sequence of pvaluep_{value} values to visualize when the series diverge (see Figure S100 through Figure S104). In Figure S104, we examine graduate cohort performance in the occupations with the greatest AI-exposure and find that graduation cohorts 2021 and 2022 are often significantly under-performing compared to earlier graduation cohorts in their first year post-graduation. As a kind of placebo test, Figure S100 compares performance in the jobs with least AI-exposure. This analysis serves as robustness check on our main analysis in Figure 2. But this alternative approach does not control for the differences in labor market conditions in month tg1t_{g1} and tg2t_{g2} which are in fact different points in time.

Image not extracted

Figure S100. One-sided pointwise difference p-values testing if graduates from recent years secured a significantly lower share of jobs with LLM exposure quintile 1. Lines are colored by reference graduation cohort which are selected to be ones that did not experience the launch of ChatGPT during their early years in the workforce post-graduation.

Image not extracted

Figure S104. One-sided pointwise difference p-values testing if graduates from recent years secured a significantly lower share of jobs with LLM exposure quintile 5. Lines are colored by reference graduation cohort which are selected to be ones that did not experience the launch of ChatGPT during their early years in the workforce post-graduation.

Image not extracted

Figure S101. One-sided pointwise difference p-values testing if graduates from recent years secured a significantly lower share of jobs with LLM exposure quintile 2. Lines are colored by reference graduation cohort which are selected to be ones that did not experience the launch of ChatGPT during their early years in the workforce post-graduation.

Image not extracted

Figure S102. One-sided pointwise difference p-values testing if graduates from recent years secured a significantly lower share of jobs with LLM exposure quintile 3. Lines are colored by reference graduation cohort which are selected to be ones that did not experience the launch of ChatGPT during their early years in the workforce post-graduation.

Image not extracted

Figure S103. One-sided pointwise difference p-values testing if graduates from recent years secured a significantly lower share of jobs with LLM exposure quintile 4. Lines are colored by reference graduation cohort which are selected to be ones that did not experience the launch of ChatGPT during their early years in the workforce post-graduation.

Workers’ Education Exposure to AI

We augment worker profiles with a dataset [23] of 3 million US higher education course syllabi from the Open Syllabus Project. The syllabus data are the third and final observation period in this study: the most recent courses available are those offered in 2017, so education exposure is measured from syllabi offered up to and including 2017, as described among the limitations below. For each syllabus ss, we compute a vector t⃗s∈R∣T∣\vec{t}_s \in\mathbb{R}^{|T|} of semantic similarities between the learning objectives of the course and each O*NET task t∈Tt \in T. Let e⃗∈[0,1]∣T∣\vec{e} \in[0,1]^{|T|} be the task-level exposure vector [2] under GPT-4, Rubric 1 β\beta, containing a 1 if a task is exposed to automation from LLMs. We define the syllabus-level education exposure as the dot product edus=t⃗s⋅e⃗edu_s = \vec{t}_s \cdot\vec{e}. Finally, given a single worker profile, we estimate the worker’s education exposure as the average edusedu_s for syllabi from the profile’s education field of study, degree-granting institution, and the four years leading up to and including their graduation year.

How to read the education exposure score

Because edusedu_s is a dot product of similarities with exposure weights, it is not a count of tasks, and we describe it here in units a reader can interpret. Each element of t⃗s\vec{t}_s is a similarity, not an indicator: across the 3,074,606 syllabi in the corpus these similarities are non-negative and never exceed 0.66, with a median across tasks below 0.21 for every syllabus. A task therefore contributes a fraction of its exposure weight rather than a whole task, and edusedu_s should be read as a similarity-weighted measure of how much of the LLM-exposed task space a course’s stated objectives cover.

Two reference points make the scale concrete. The exposure vector sums to ∑τeτ=6,164.6\sum_{\tau} e_{\tau} = 6{,}164.6 over the 17,992 O*NET tasks, of which 9,869 (54.9%) carry non-zero exposure and 2,463 (13.7%) are fully exposed; this sum is the value a syllabus would attain if it matched every exposed task perfectly, and it is the natural denominator for expressing edusedu_s as a share. Observed scores occupy the lower half of that range: across syllabi, edusedu_s has mean 1,745.2 and standard deviation 386.1, a median of 1,801 (29.2% of the maximum) and a maximum of 3,170 (51.4%). Tables S15 and S16 report both normalizations alongside the raw score, and the regression tables report eduedu as a zz-score so that its coefficients are per standard deviation.

It follows that ratios of raw scores are not ratios of task counts. The most exposed institution in Table S16 scores 1.93 times the least exposed, but this does not mean its graduates encountered twice as many LLM-exposed tasks. It means that its courses resemble the exposed task space about 1.6 standard deviations more than the average course, against about 1.4 standard deviations less for the least exposed, and that the two differ by roughly 18 percentage points of the attainable maximum. The ordering of Table S16 is informative in this light: community and technical colleges score highest and research universities lowest, because the measure rewards coursework whose objectives are phrased as occupational tasks, which is what the O*NET task statements describe.

University NameAvg. edusedu_sSyllabus Count% of ∑teτ\sum_t e_\tauzz-score
New Mexico Junior College235516,63338.2%+1.58
Modesto Junior College215910,77935.0%+1.07
Park University203227,09633.0%+0.74
Fresno City College201724,78632.7%+0.70
Kentucky Community and Technical College System201510,24432.7%+0.70
Rowan-Cabarrus Community College199733,40332.4%+0.65
Hartnell College198716,51632.2%+0.63
South Texas College197711,99832.1%+0.60
Clark State Community College197446,09432.0%+0.59
Santa Rosa Junior College195935,61331.8%+0.55
⋮\vdots⋮\vdots⋮\vdots⋮\vdots⋮\vdots
Loyola University New Orleans163615,39826.5%-0.28
Midwestern State University162225,20226.3%-0.32
Minnesota State Colleges and Universities System159222,86925.8%-0.40
Alamo Colleges1588159,99525.8%-0.41
University of Alabama, Tuscaloosa156053,36825.3%-0.48
Galveston College147618,90823.9%-0.70
Wilkes University133810,00821.7%-1.05
University of Michigan–Ann Arbor125323,79820.3%-1.27
University of California, Irvine125211,83420.3%-1.28
Princeton University122311,66719.8%-1.35

Table S16. The ten universities with highest and lowest average edusedu_s rounded to the nearest integer. The last two columns are defined as in Table S15. Because edusedu_s is a similarity-weighted sum rather than a tally of tasks, the ratio of two universities’ scores is not a ratio of task counts: New Mexico Junior College scores 1.93 times Princeton University, but the interpretation is that its courses sit about 1.6 SD above the average course and Princeton’s about 1.4 SD below, a gap of roughly 3 SD in the syllabus distribution. Community and technical colleges occupy the top of this table and research universities the bottom because the measure rewards coursework whose stated objectives resemble workplace tasks.

Table S17 reports the corpus underlying these scores. The Open Syllabus extract covers 2,599 institutions, 38 two-digit CIP fields and cohort years 2000 through 2024, giving 42,828 distinct university-by-field programs. Evidence per program is highly uneven: the median program is backed by 6 syllabi, the first quartile by 2 and the ninety-fifth percentile by 253, and 8,906 programs (20.8%) rest on a single syllabus. Large programs dominate the graduate sample, but scores for thinly covered programs should be read as noisy estimates of curricular content. Coverage across cohort years is sparser still. Of the 875,955 university-by-field-by-year cells, 410,927 (46.9%) contain no syllabus within the four-year window ending in that cohort year; every one of these carries the score that the same program takes in a year whose window is non-empty, and a quarter of programs (25.6%) take a single value across all 25 cohort years. Educational exposure in this study should therefore be read as a stable characteristic of a program rather than as a year-by-year measurement, which is consistent with our use of it as a baseline attribute measured before generative AI.

Panel A. Coverage
Syllabi3,074,606
Institutions2,599
Fields of study (2-digit CIP)38
University-by-field programs42,828
Cohort years2000–2024
Program-by-year cells875,955
empty in their four-year window410,927 (46.9%)
programs with a single score across years10,946 (25.6%)
Panel B. Distinct syllabi per program
Meanp25Medianp75p95Max
Syllabi65.6262625365,006
Panel C. Distribution of edusedu_s
NNMeanSDp5Medianp95
Per syllabus3,074,6061,745.2386.11,011.31,801.22,279.6
Per program42,8281,682.1316.01,123.91,710.92,134.6
Per graduate1,258,0681,739.5325.91,150.01,778.62,177.8
Per field of study381,747.4121.91,523.81,759.71,899.8

Table S17. Table S17: The syllabus corpus behind the education exposure score. Panel A reports coverage. Panel B reports how many distinct syllabi back a university-by-field program; because num_syllabi counts syllabi within the four-year window ending in each cohort year, consecutive years overlap and the distinct count is the sum over years divided by four, which reproduces the syllabus counts in Table S16 to within 0.4%. Panel C reports the distribution of edusedu_s at each level of aggregation; the median corresponds to 29.2% of ∑τeτ=6,164.6\sum_{\tau} e_{\tau} = 6,164.6 per syllabus, 27.8% per program and 28.9% per graduate, and the per-graduate row is the estimation sample of Table 1. Of the 42,828 programs, 8,906 (20.8%) rest on a single syllabus and 20,470 (47.8%) on five or fewer.

The derived extract we analyze records each syllabus only through its similarity vector and the resulting scores, so we cannot report syllabus length or word counts. What it does support is a comparison of the distribution of course-level exposure across institutions and fields, which is what Table S17 provides.

This approach has several limitations. First, the syllabus data is a subset of all US higher education. Second, the most recent syllabi available are for courses offered in 2017. For workers who graduated after 2017, we default to the relevant syllabi from years 2014-2017 and we consider their education exposure in the context of status quo educational activity from before the COVID pandemic and before generative AI. Despite these limitations, to the best of our knowledge, there is no better data for estimating the difference in teaching across fields of study, educational institutions, and graduation years (see Table S15 for average edusedu_s across fields of study and Table S16 for the most and least exposed universities.)

CIP CodeCIP TitleAvg. edusedu_sSyllabus Count% of ∑τeτ\sum_{\tau} e_{\tau}zz-score
44PUBLIC ADMINISTRATION AND SOCIAL SERVICE PROFESSIONS.199622,17132.4%+0.65
52BUSINESS, MANAGEMENT, MARKETING, AND RELATED SUPPORT SERVICES.1970229,51232.0%+0.58
41SCIENCE TECHNOLOGIES/TECHNICIANS.18995,97030.8%+0.40
51HEALTH PROFESSIONS AND RELATED PROGRAMS.1870218,64230.3%+0.32
47MECHANIC AND REPAIR TECHNOLOGIES/TECHNICIANS.186225,36430.2%+0.30
19FAMILY AND CONSUMER SCIENCES/HUMAN SCIENCES.184954,32030.0%+0.27
46CONSTRUCTION TRADES.184326,30329.9%+0.25
04ARCHITECTURE AND RELATED SERVICES.184017,77929.8%+0.25
12CULINARY, ENTERTAINMENT, AND PERSONAL SERVICES.183816,25629.8%+0.24
29MILITARY TECHNOLOGIES AND APPLIED SCIENCES.183718,29029.8%+0.24
43HOMELAND SECURITY, LAW ENFORCEMENT, FIREFIGHTING AND RELATED PROTECTIVE SERVICES.1832100,07229.7%+0.22
31PARKS, RECREATION, LEISURE, FITNESS, AND KINESIOLOGY.183279,19829.7%+0.22
09COMMUNICATION, JOURNALISM, AND RELATED PROGRAMS.182968,89129.7%+0.22
13EDUCATION.1813391,91829.4%+0.18
15ENGINEERING/ENGINEERING-RELATED TECHNOLOGIES/TECHNICIANS.181258,48829.4%+0.17
25LIBRARY SCIENCE.18077,27629.3%+0.16
10COMMUNICATIONS TECHNOLOGIES/TECHNICIANS AND SUPPORT SERVICES.179534,79129.1%+0.13
42PSYCHOLOGY.177381,60328.8%+0.07
01AGRICULTURAL/ANIMAL/PLANT/VETERINARY SCIENCE AND RELATED FIELDS.176128,31228.6%+0.04
49TRANSPORTATION AND MATERIALS MOVING.175823,42228.5%+0.03
11COMPUTER AND INFORMATION SCIENCES AND SUPPORT SERVICES.1754126,67828.5%+0.02
48PRECISION PRODUCTION.17414,24428.2%-0.01
03NATURAL RESOURCES AND CONSERVATION.17338,92428.1%-0.03
14ENGINEERING.170734,15927.7%-0.10
26BIOLOGICAL AND BIOMEDICAL SCIENCES.1693142,11827.5%-0.14
50VISUAL AND PERFORMING ARTS.1687185,89827.4%-0.15
24LIBERAL ARTS AND SCIENCES, GENERAL STUDIES AND HUMANITIES.16798,66727.2%-0.17
23ENGLISH LANGUAGE AND LITERATURE/LETTERS.1678186,19927.2%-0.17
22LEGAL PROFESSIONS AND STUDIES.166044,00826.9%-0.22
39THEOLOGY AND RELIGIOUS VOCATIONS.165931,96026.9%-0.22
45SOCIAL SCIENCES.1642133,58026.6%-0.27
30MULTI/INTERDISCIPLINARY STUDIES.163270,52726.5%-0.29
40PHYSICAL SCIENCES.1627133,96426.4%-0.31
27MATHEMATICS AND STATISTICS.1620204,84826.3%-0.32
38PHILOSOPHY AND RELIGIOUS STUDIES.152640,16124.8%-0.57
54HISTORY.152351,18624.7%-0.58
05AREA, ETHNIC, CULTURAL, GENDER, AND GROUP STUDIES.151072,98124.5%-0.61
16FOREIGN LANGUAGES, LITERATURES, AND LINGUISTICS.149785,92624.3%-0.64

Table S15. Average edusedu_s by field of study (two-digit CIP code) rounded to the nearest integer. Fields are ordered from highest to lowest average edusedu_s. edusedu_s is not a count of tasks. It is the sum, over all 17,992 O*NET tasks, of the semantic similarity between a course’s learning objectives and the task, weighted by that task’s LLM exposure eτe_{\tau}. The fourth column expresses it as a percentage of ∑τeτ=6,164.6\sum_{\tau} e_{\tau} = 6,164.6, the value a course would attain if it matched every LLM-exposed task perfectly, which is the exposure-weighted average similarity between the field’s courses and the exposed task space. The fifth column is a zz-score against the distribution of edusedu_s over all 3,074,606 syllabi (mean 1,745.2, SD 386.1).

Reading the three-way interaction

The models in Table 1 interact education exposure with both the LLM exposure of the first job and the launch of ChatGPT, and a three-way interaction is difficult to read off a coefficient table. Figure S105 therefore plots what Models 2 and 4 imply for a graduate’s first job. Each panel fixes one outcome and one period, one line fixes the first job’s LLM exposure at one standard deviation below the mean and the other one standard deviation above it, and education exposure varies along the horizontal axis. Because every predictor and both dependent variables are zz-scores, the slope of a line is the association between a one standard deviation increase in eduedu and the outcome, measured in standard deviations of that outcome.

Each panel is drawn relative to its own reference graduate, who has average education exposure and enters a job one standard deviation below the mean in LLM exposure in that same period. This removes only the level shift common to every job that starts after the launch, which for months to first job is 2.338 standard deviations and would otherwise compress everything else in the panel; the association of llmllm with the outcome, its change at the launch, and every term involving eduedu are all retained. The three-way interaction is then the change, from the left panel to the right panel, in the gap between the two slopes.

For salary, education exposure is associated with higher pay in jobs of low LLM exposure and lower pay in jobs of high LLM exposure, and the gap between those two slopes narrows from 0.038 standard deviations before the launch (+0.018 against −0.020) to 0.016 after it (+0.012 against −0.005). For months to first job, the slopes are +0.000 and +0.010 before the launch and −0.015 and −0.023 after it, so search times fall with education exposure only in the later period and fall fastest in the most exposed jobs. Every interaction term in Models 2 and 4 is significant at the 5% level or better; the weakest is edu×gptedu \times gpt in Model 2, at p=0.029p = 0.029. The figure also makes the size of these associations visible: across the four standard deviations of education exposure spanned by each panel, every line moves by less than a tenth of a standard deviation of its outcome, which is the sense in which we describe the three-way interactions as modest in the main text.

Sample size and effect size

The models in Table 1 are estimated on 1,258,068 graduates. That sample is the 1,732,499 first jobs held by 2019 through 2024 graduates within three years of graduation, restricted to those whose program carries an education exposure score and whose first-job occupation carries a work-from-home share. It is smaller than the 2,665,306 graduates of SI Table S11 because that analysis covers the 2017 through 2024 cohorts and requires neither a matched syllabus nor a work-from-home score. At this size the standard error on a standardized coefficient is of order 0.001, so statistical significance no longer discriminates between findings and the interpretive burden falls on magnitude. This section therefore reports the estimates in the units of the dependent variables, the share of variance the education exposure terms add, and the same models re-estimated on small subsamples.

The dependent variables have standard deviations of 0.488 log points and 6.31 months, so a coefficient of bb standard deviations corresponds to 100(e0.488b−1)%100(e^{0.488b} - 1)\% of salary or 189b189b days of job search. Read this way, Model 2 says that a one standard deviation increase in the first job’s LLM exposure is associated with 4.4% higher salary before the launch of ChatGPT and 1.3% after it, while the terms involving education exposure are −0.03% for eduedu, +0.20% for edu×gptedu \times gpt, −0.92% for edu×llmedu \times llm and +0.52% for edu×llm×gptedu \times llm \times gpt. Model 4 says that a one standard deviation increase in the first job’s LLM exposure is associated with a search 2.7 days shorter before the launch and 3.4 days longer after it, with +1.0 day for eduedu, −4.6 days for edu×gptedu \times gpt, +1.0 day for edu×llmedu \times llm and −1.8 days for edu×llm×gptedu \times llm \times gpt. The largest association involving education exposure is therefore about half a percent of salary or a working week of search time, against 4.4% of salary for the largest association in the table, the LLM exposure premium itself.

The same point can be made as variance explained. Dropping the four education exposure terms from Model 2 lowers R2R^2 from 0.4423 to 0.4420, an increment of 0.0003 (Cohen’s f2=0.0005f^2 = 0.0005), while dropping the four LLM exposure terms instead lowers it to 0.4381, an increment of 0.0042 (f2=0.0075f^2 = 0.0075). In Model 4 the two increments are 0.0001 and 0.0002. These are small by any convention: first-job salary and search time are overwhelmingly explained by the fixed effects for university, field of study, degree, graduation year, state and sector and by the occupation’s wage level. We report the education exposure terms as evidence about the sign and the timing of an association, not as an account of individual outcomes.

Finally, we re-estimate both specifications on 50 independent random subsamples of 50,000 graduates each, drawn without replacement with seeded draws, using the identical specification including every university indicator present in the draw (Table S19). We match the design of SI Table S12 so that the two exercises can be compared. The standard errors behave as sampling theory requires: a sample 25 times smaller should inflate them by 1,258,068/50,000=5.02\sqrt{1,258,068/50,000}=5.02, and the mean within-draw standard error is 5.07 times the full-sample standard error for both outcomes. That is direct evidence that the small standard errors of Table 1 reflect the size of the sample rather than an understatement of uncertainty.

What survives resampling differs across the table, and we state it plainly rather than claim uniform robustness. For months to first job every term except eduedu reproduces its full-sample sign in at least 94% of draws (eduedu does so in 84%), and edu×gptedu \times gpt, which carries the finding that education exposure is associated with shorter searches after the launch, does so in every draw and remains significant at the 5% level in 76% of them. For log salary the two interactions with the job’s LLM exposure are comparably stable, edu×llmedu \times llm reproducing its sign in every draw and edu×llm×gptedu \times llm \times gpt in 96%. The two weakest terms in the table do not survive: eduedu, which is indistinguishable from zero in the full sample, and edu×gptedu \times gpt, the weakest interaction at p=0.029p=0.029, reproduce their signs in 58% and 60% of draws and are significant in 6% and 4%. An estimate of 0.004 standard deviations cannot be resolved against a standard error of 0.009, so this is the expected consequence of the referee’s test rather than a failure of it, and it is the reason we do not rest the argument on any single coefficient.

Two caveats belong with the exercise. Fitting roughly 1,230 parameters to 50,000 observations makes the design nearly singular in places, and R2R^2 rises from 0.442 to a mean of 0.456 for salary and from 0.375 to 0.391 for months, which is overfitting rather than improved explanation. The draws are also not exactly centered on the full-sample estimates: the mean of edu×llm×gptedu \times llm \times gpt in Model 4 is −0.014 against a full-sample −0.009, the largest such gap in either panel, which is smaller than the spread of that term across draws but larger than the standard error of their mean and so reflects the behavior of the term at n=50,000n=50,000 rather than sampling noise alone. And the draws are subsets of the same data, so they test whether the estimates are an artifact of precision, not whether they would replicate in a different sample. Unlike the job seeking subsamples of SI Table S12, all 800 subsample estimates here had a computable standard error.

Dependent Variable:Log Salary (Real 2015 $) | | Months to First Job |
VariableModel 1Model 2Model 3Model 4
Job Opening Rate0.014***0.014***2.994***2.994***
Major SOC Log Annual Wage (2020)0.330***0.312***-0.224***-0.202***
Work-From-Home Score (%)0.010***0.006***-0.016***-0.010***
Post ChatGPT Indicator (gptgpt)0.017***0.015***14.743***14.764***
Education Exposure (eduedu)-0.000-0.0000.000***0.000***
LLM Exposure (llmllm)0.214***-0.450***
gpt×llmgpt \times llm-0.030***0.205***
edu×gptedu \times gpt0.003***0.002*-0.155***-0.153***
edu×llmedu \times llm-0.009***0.032***
edu×llm×gptedu \times llm \times gpt0.005***-0.060***
State FEYesYesYesYes
Sector FEYesYesYesYes
Terminal Degree Type FEYesYesYesYes
Graduation Year FEYesYesYesYes
College Major FEYesYesYesYes
University FEYesYesYesYes
R2R^20.4380.4420.3750.375
adj. R2R^20.4370.4420.3740.374
Observations1,258,0681,258,0681,258,0681,258,068
pval<0.1∗p_{val} < 0.1^{*}, pval<0.01∗∗p_{val} < 0.01^{**}, pval<0.001∗∗∗p_{val} < 0.001^{***}

Table S18. Education exposure to LLM exposed tasks is associated with shorter job searches after ChatGPT and partially offsets salary penalties in highly exposed occupations. gpt=1gpt = 1 if job starts on or after December 2022. eduedu is education exposure averaged over matched syllabi, a similarity-weighted measure rather than a count of tasks (see SI Sections S12 and S12.1). llmllm is first-job occupation six-digit SOC code LLM exposure [2]. All models include fixed effects for the state and two-digit NAICS code of the job as well as for the terminal degree type (e.g., bachelors or masters), graduation year, college field of study (i.e., two-digit CIP Code), and degree-granting university.

Weighting Revelio Graduates to Match NCSG Survey Graduates

The National Survey of College Graduates (NSCG) is a biennial survey sponsored by the U.S. National Science Foundation’s National Center for Science and Engineering Statistics (NCSES) and conducted by the U.S. Census Bureau. The survey provides nationally oriented data on the characteristics of U.S. college graduates, with particular emphasis on the science and engineering workforce. In the 2023 cycle, the target population is non-institutionalized residents of the United States or Puerto Rico who have earned at least a bachelor’s degree and are younger than 76 as of the survey reference date. The NSCG is drawn from the American Community Survey (ACS) sampling frame and uses a four-panel rotating panel design in which each new panel completes a baseline interview followed by up to three biennial follow-ups before rotating out. Data are collected via a tri-modal approach (web, mail, and CATI), and the public-use files include final survey weights designed to support unbiased population inference under the complex design; for the 2023 cycle, NCSES reports a weighted response rate of 61% and notes that sampling errors are computed via successive difference replication.

In studies using LinkedIn worker profiles—which are typically a non-probability sample whose composition can differ systematically from the population of college graduates—the NSCG can serve as an external benchmark to re-weight analyses toward population-representative estimates for outcomes observable on LinkedIn. Concretely, we compare the distribution of graduates by college field of study in the Revelio/LinkedIn data to the distribution in the corresponding graduation year in NSCG (see Figure S106-Figure S110). Typically, compared to the NSCG, the Revelio worker profile data over samples Business and Communications majors and under samples Engineering and Education majors.

Image not extracted

Figure S106. Comparing the Revelio distribuion of fields of study of bachelors recipients in graduation cohort 2013to the distribuion according to the NSCG survey.

Image not extracted

Figure S110. Comparing the Revelio distribuion of fields of study of bachelors recipients in graduation cohort 2023to the distribuion according to the NSCG survey.

We can then weight our sample of LinkedIn profiles for the biennial graduation years in the NSCG so the distribution college fields of study are identical and repeat our analysis of earnings and job seeking time in graduates first job after degree completion (compare main text Table 1 to the weighted analysis on a subset of graduation cohorts in Table S20). In the weighted analysis, key results from the main regression remain unchanged or are perhaps stronger. Education exposure (eduedu) is only weakly related to salaries and job seeking times before the launch of ChatGPT, at no more than 0.013 SD in either outcome, but becomes more predictive afterward. In particular, the three-way interaction (edu×llm×gptedu \times llm \times gpt) continues to indicate the added value of educational exposure for graduates whose first job after degree completion is LLM-exposed after ChatGPT.

Image not extracted

Figure S107. Comparing the Revelio distribuion of fields of study of bachelors recipients in graduation cohort 2015to the distribuion according to the NSCG survey.

Image not extracted

Figure S108. Comparing the Revelio distribuion of fields of study of bachelors recipients in graduation cohort 2017to the distribuion according to the NSCG survey.

Image not extracted

Figure S109. Comparing the Revelio distribuion of fields of study of bachelors recipients in graduation cohort 2021to the distribuion according to the NSCG survey.

Graduation Cohort Job Gaps by Education Exposure Decile

Finally, we repeat the disaggregation of SI § 15 using deciles of the graduate’s educational LLM exposure (eduedu, see SI § 18) in place of the LLM exposure of the first job (see Figure S112 through Figure S121). This asks whether the cohort gap is organized by what graduates were trained on rather than by the jobs they entered. As in SI § 15, panel A of each figure gives the month-by-month series and panel B averages the same statistic over the months before and after the launch of ChatGPT, and Figure S111 collects those summaries for all ten deciles. The education deciles carry large and persistent differences, but almost none of them move at the launch, which is what we would expect if the cohort gap is organized by the jobs graduates enter rather than by what they studied.

Image not extracted

Figure S111. Summary of the decile figures in this section. As in Figure S78, but for deciles of educational LLM exposure. Each cell reports the difference in mean jobGapjobGap between a graduation cohort and the 2013–2020 cohorts over the same months since their own graduation, in percentage points; the third row of each block is the change in that difference at the launch of ChatGPT. Stars mark two-sample tt-tests at the 5%, 1%, and 0.1% levels.

Image not extracted

Figure S112. (A) For each graduation cohort by month from May of their graduation year, the share of jobs started by graduates with education AI-exposure decile 1 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S121. (A) For each graduation cohort by month from May of their graduation year, the share of jobs started by graduates with education AI-exposure decile 10 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S113. (A) For each graduation cohort by month from May of their graduation year, the share of jobs started by graduates with education AI-exposure decile 2 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S114. (A) For each graduation cohort by month from May of their graduation year, the share of jobs started by graduates with education AI-exposure decile 3 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S115. (A) For each graduation cohort by month from May of their graduation year, the share of jobs started by graduates with education AI-exposure decile 4 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S116. (A) For each graduation cohort by month from May of their graduation year, the share of jobs started by graduates with education AI-exposure decile 5 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S117. (A) For each graduation cohort by month from May of their graduation year, the share of jobs started by graduates with education AI-exposure decile 6 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S118. (A) For each graduation cohort by month from May of their graduation year, the share of jobs started by graduates with education AI-exposure decile 7 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S119. (A) For each graduation cohort by month from May of their graduation year, the share of jobs started by graduates with education AI-exposure decile 8 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Image not extracted

Figure S120. (A) For each graduation cohort by month from May of their graduation year, the share of jobs started by graduates with education AI-exposure decile 9 started compared to the overall share of jobs. Vertical lines indicate 95% confidence intervals. Dashed lines indicate the launch of ChatGPT for graduation cohorts 2021 and 2022. (B) The same statistic averaged over the months before and after the launch of ChatGPT, separately for the 2021 and 2022 graduation cohorts, each compared against the 2013–2020 cohorts averaged over the same months since their own graduation. Error bars are 95% confidence intervals.

Time to First Job: Observation Windows and Treatment Timing

This section examines two features of the time-to-first-job analyses that could shape their results: the observation window available to each graduation cohort, and the definition of the post-ChatGPT indicator in Table 1.

A common observation window

Job start dates in the Revelio extract end in December 2024, so the most recent cohorts cannot be followed for the full three years used in the main analysis: the 2023 cohort is observed for about 19 months after May of its graduation year and the 2024 cohort for about seven. Truncation removes the slowest job searches, and it can remove more of them from one group of jobs than another. Graduation-year fixed effects absorb differences across cohorts in the level of job delays, but not truncation that differs within a cohort between more and less exposed jobs. We therefore re-estimate the model behind Figure 2A (Eq. (15)) on windows that every included cohort fully covers: first jobs starting within 12 months of May of the graduation year for the 2017–2023 cohorts (2,161,375 graduates), and within 6 months for the 2017–2024 cohorts (1,707,969 graduates). Fixed effects and controls are unchanged.

Figure S122 reports the cohort-specific LLM-exposure coefficients βg,llm\beta_{g,llm}. Coefficients are smaller in every cohort under the shorter windows because the range of possible delays is shorter, so they should be compared within a specification rather than across specifications. Within the 12-month window, graduates whose first job was highly exposed found jobs faster than their peers in every cohort. The advantage is largest for the 2020–2022 cohorts (−0.314, −0.277, and −0.585) and returns in 2023 (−0.106, SE 0.037) to the range of the 2017–2019 cohorts (−0.106 to −0.151). Within the 6-month window the 2023 coefficient is again negative (−0.166, SE 0.020), and the 2024 coefficient is small and not statistically distinguishable from zero (0.059, SE 0.036). The positive coefficients for the 2023 and 2024 cohorts in the full sample (0.486 and 0.582) therefore reflect the shorter observation window available to those cohorts rather than a reversal in the job search advantage of exposed jobs.

Image not extracted

Figure S122. On a common observation window, the job search advantage of LLM-exposed first jobs narrows after 2022 but does not reverse. Cohort-specific coefficients βg,llm\beta_{g,llm} from Eq. (15): the difference in months to first job between graduates whose first job had high versus low LLM exposure. Black: the full sample of Figure 2A (first jobs within three years of graduation). Blue: first jobs within 12 months of May of the graduation year, 2017–2023 cohorts. Orange: first jobs within 6 months, 2017–2024 cohorts. Bars are 95% confidence intervals; the dashed line marks the launch of ChatGPT.

A cohort-level post-ChatGPT indicator

In Table 1 the indicator gptgpt equals one when the first job starts in or after December 2022. For the time-to-first-job models (Models 3 and 4) this indicator is partly determined by the outcome: within a graduation cohort, a graduate whose first job starts after the launch has necessarily searched longer than a classmate whose first job starts before it. Table S21 instead assigns post-ChatGPT status by graduation cohort, setting gptcohort=1gpt_{cohort} = 1 for cohorts graduating in 2023 or later, whose job search began after the launch. Because gptcohortgpt_{cohort} is collinear with the graduation-year fixed effects, only its interactions enter the model. Models 1 and 2 use every cohort in the Table 1 sample; Models 3 and 4 additionally restrict to first jobs within 12 months of May of the graduation year and to the 2019–2023 cohorts, a window that every included cohort fully covers. All other terms and fixed effects are those of Table 1, and variables are standardized in the same way.

The two associations the main text relies on hold under the cohort-level indicator. Higher education exposure is associated with shorter job searches after the launch (edu×gptcohortedu \times gpt_{cohort} between −0.015 and −0.022 across the four models, compared with −0.024 in Table 1), and more exposed first jobs are associated with longer searches after the launch (gptcohort×llmgpt_{cohort} \times llm of 0.043 in the full sample and 0.013 on the common window, compared with 0.032). The three-way interaction is not robust: it is −0.005 (p<0.1p < 0.1) in the full sample and 0.001 (not significant) on the common window, compared with −0.009 in Table 1. We therefore do not rely on the three-way interaction for time to first job.

References

  1. [1]Edward W Felten, Manav Raj, and Robert Seamans. Occupational heterogeneity in exposure to generative ai. Available at SSRN 4414065, 2023.DOI
  2. [2]Tyna Eloundou, Sam Manning, Pamela Mishkin, and Daniel Rock. GPTs are GPTs: Labor market impact potential of LLMs. Science, 384(6702):1306–1308, 2024.DOI
  3. [3]Reuters. ChatGPT sets record for fastest-growing user base. Reuters Technology News, February 2023. Published February 1, 2023.
  4. [4]Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen. Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence. Stanford Digital Economy Lab Working Paper, August 2026. Revised August 12, 2026.
  5. [5]Guy Lichtinger and Seyed Mahdi Hosseini Maasoum. Generative AI as seniority-biased technological change: Evidence from us resume and job posting data. Available at SSRN, 2025.DOI
  6. [6]José Azar, Mireia Gine, and Javier Sanz-Espín. AI Is Already Eroding Wages: Quasi-Experimental Evidence From Occupational Exposure. Available at SSRN 5842084, 2025.DOI
  7. [7]Federal Reserve Bank of New York. The Labor Market for Recent College Graduates. https://www.newyorkfed.org/research/college-labor-market, 2024. Accessed 16 December 2025.
  8. [8]METR. Time horizon 1.1. https://metr.org/blog/2026-1-29-time-horizon-1-1/, January 2026. Published January 29, 2026.
  9. [9]Jean-Stanislas Denain and Alexander Barry. Have AI capabilities accelerated? Epoch AI, April 2026. Published April 16, 2026.
  10. [10]Yoshua Bengio et al. International AI safety report 2026. Technical Report DSIT 2026/001, UK Department for Science, Innovation and Technology, February 2026. arXiv:2602.21012.arxiv.org/abs/2602.21012
  11. [11]Eduard Bruill, Samuel Mäurer, and Davud Rostam-Afschar. Beliefs about bots: How employers plan for AI in white-collar work. arXiv preprint arXiv:2510.21959, 2025.
  12. [12]Anders Humlum and Emilie Vestergaard. Large Language Models, Small Labor Market Effects. Technical report, National Bureau of Economic Research, 2025.DOI
  13. [13]Andrew Johnston and Christos Makridis. Ai, output, and employment. 2026.
  14. [14]Martha Gimbel, Molly Kinder, Joshua Kendall, and Maddie Lee. Evaluating the impact of AI on the labor market: Current state of affairs. The Budget Lab at Yale, 1, 2025.
  15. [15]Zanna Iscenko and Fabien Curto Millet. Looking for the Ladder: Is AI Impacting Entry-Level Jobs? Economic Innovation Group, Washington, DC. https://eig.org/wp-content/uploads/2026/01/TAWP-Iscenko-Millet.pdf, 2026.
  16. [16]Anders Humlum and Emilie Vestergaard. Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. Research Insight 28/26, ROCKWOOL Foundation Berlin (RFBerlin), July 2026.
  17. [17]Jonathan Hartley, Filip Jolevski, Vitor Melo, and Brendan Moore. The labor market effects of generative artificial intelligence. Technical report, SSRN Working Paper 5136877, 2026.
  18. [18]Jessica Liu and Douglas Webber. AI adoption and firms’ job-posting behavior. FEDS Notes, Board of Governors of the Federal Reserve System, March 2026. Published March 27, 2026.DOI
  19. [19]Board of Governors of the Federal Reserve System (US). Federal funds effective rate (fed-funds). Retrieved from FRED, Federal Reserve Bank of St. Louis, November 2025. Accessed November 14, 2025.
  20. [20]Gregor Zens, Maximilian Böck, and Thomas O Zörner. The heterogeneous impact of monetary policy on the us labor market. Journal of Economic Dynamics and Control, 119:103989, 2020.DOI
  21. [21]Federal Reserve Bank of St. Louis. Indeed, Software Development Job Postings in the United States. https://fred.stlouisfed.org/series/IHLIDXUSTPSOFTDEVE, 2025. FRED economic data series IHLIDXUSTPSOFTDEVE. Accessed December 17, 2025.
  22. [22]Diego Mendez-Carbajo. Was there a tech-hiring bubble? Indeed.com job postings data suggest so. FRED Blog.
  23. [23]Alireza Javadian Sabet, Sarah H Bana, Renzhe Yu, and Morgan R Frank. Course-Skill Atlas: A national longitudinal dataset of skills taught in US higher education curricula. Scientific Data, 11(1):1086, 2024.DOI
  24. [24]Morgan R Frank, Yong-Yeol Ahn, and Esteban Moro. AI exposure predicts unemployment risk: A new approach to technology-driven job loss. PNAS Nexus, 4(4):pgaf107, 2025.DOI
  25. [25]Maxim Massenkoff and Peter McCrory. Labor market impacts of ai: A new measure and early evidence. Technical report, 2026.
  26. [26]Edward W Felten, Manav Raj, and Robert Seamans. A Method to Link Advances in Artificial Intelligence to Occupational Abilities. In AEA Papers and Proceedings, volume 108, pages 54–57, 2018.DOI
  27. [27]Ruth Appel, Peter McCrory, Alex Tamkin, Michael Stern, Miles McCain, and Tyler Neylon. Anthropic Economic Index Report: Uneven Geographic and Enterprise AI Adoption, 2025.
  28. [28]Jin Wang and Wenxiang Fan. The effect of ChatGPT on students’ learning performance, learning perception, and higher-order thinking: Insights from a meta-analysis. Humanities and Social Sciences Communications, 12(1):1–21, 2025.DOI
  29. [29]Stephen Hansen, Peter John Lambert, Nicholas Bloom, Steven J Davis, Rafaella Sadun, and Bledi Taska. Remote work across jobs, companies, and space. Technical report, National Bureau of Economic Research, 2023.
  30. [30]Francesco Filippucci, Peter N Gal, and Matthias Schief. Aggregate Productivity Gains from Artificial Intelligence: a Sectoral Perspective. Available at SSRN 5260318, 2025.DOI
  31. [31]Peter John Lambert and Yannick Schindler. The broken ladder: AI, remote work, and early-career hiring. 2026.
  32. [32]Fangyan Wang, Zaiyan Wei, and Yang Wang. Generative AI and the reorganization of labor demand. arXiv preprint arXiv:2605.23159, 2026.arxiv.org/abs/2605.23159
  33. [33]Richard Audoly, Miles Guerin, and Giorgio Topa. Do job postings show early labor-market effects of AI? Liberty Street Economics, Federal Reserve Bank of New York, May 2026. Published May 14, 2026.DOI
  34. [34]R. Maria del Rio-Chanona, Ekkehard Ernst, Rossana Merola, Daniel Samaan, and Ole Teutloff. AI and jobs: A review of theory, estimates, and evidence. arXiv preprint arXiv:2509.15265, 2025.
  35. [35]Shakked Noy and Whitney Zhang. Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654):187–192, 2023.
  36. [36]Sida Peng, Eirini Kalliamvakou, Peter Cihon, and Mert Demirer. The impact of ai on developer productivity: Evidence from github copilot. arXiv preprint arXiv:2302.06590, 2023.arxiv.org/abs/2302.06590
  37. [37]Erik Brynjolfsson, Danielle Li, and Lindsey R Raymond. Generative AI at work. The Quarterly Journal of Economics, 140(2):889–942, 2025.
  38. [38]Fabrizio Dell’Acqua, Edward McFowland III, Ethan R Mollick, Hila Lifshitz-Assaf, Katherine Kellogg, Saran Rajendran, Lisa Krayer, François Candelon, and Karim R Lakhani. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. Organization Science, 2025. https://doi.org/10.1287/orsc.2025.21838.
  39. [39]Jacob Dominski and Yong Suk Lee. Advancing AI capabilities and evolving labor outcomes. arXiv preprint arXiv:2507.08244, 2025.arxiv.org/abs/2507.08244
  40. [40]U.S. Department of Labor, Employment and Training Administration. UI Reports Handbook No. 401: ETA 203, Characteristics of the Insured Unemployed (IV-4-4). PDF, 2001. Accessed 2025-12-25.
  41. [41]U.S. Department of Labor, Employment and Training Administration, Office of Unemployment Insurance. Unemployment Insurance Data Downloads. Web page, 2025. Accessed 2025-12-25.
  42. [42]U.S. Department of Labor, Employment and Training Administration. Unemployment Insurance Program Letter No. 22-11: Elimination of Sampling as a Method for Producing the ETA 203 Report. Web page, June 2011. Accessed 2025-12-25.
  43. [43]U.S. Bureau of Labor Statistics. How the Government Measures Unemployment. PDF, June 2014. Accessed 2025-12-25.
  44. [44]Guido Matias Cortes and Eliza Forsythe. Impacts of the COVID-19 pandemic and the CARES act on earnings and inequality. Working paper, York University and University of Illinois Urbana-Champaign, May 2021. May 12, 2021 version; accessed 2025-12-25.

Paper details

vlnc:W_rbs5bs76
Version
v2 · latest
Published
Added
License
CC BY 4.0
Usage

views · 1

downloads · 8

A view counts once per browser per day, a download each time the PDF is requested.

  • Economics, Econometrics and Finance

Not recorded for this paper

View statement
Looking for related papers…
Contents