Study library · selected research in the current map

The studies behind the AI job claims

Each card says what was measured. That is the difference between a warning and a claim you can use.

Start with the Norwegian register-data check, then compare hiring, payroll and wage studies below.

Norway: no clear broad exposure-linked employment decline

Source and version: Hernæs and Kostøl, AI Labor Market Index, October 2026 release; the June discussion paper is an earlier vintage of the same research.
Population and place: Norwegian private-sector wage earners aged 21–60 across 397 scored occupations.
Period: Monthly employer registers through July 2026; pre-ChatGPT baseline August–October 2022.
AI measure: Occupation task exposure to language models, not employer adoption.
Outcome: Paid employment, with separate new-hire and nominal pay series on the dashboard.
Method: Three-month smoothed, seasonally adjusted employment growth in highest versus lowest exposure fifth; descriptive comparison.
Finding: Ages 21–60: +1.5 percentage points, standard error 2.7. Ages 21–30: −0.8 points, standard error 4.0. From February 2025 the young gap is −2.9 points, standard error 2.0.
Causal status: No employer-use comparison or causal AI displacement estimate.
Limit: Imprecise group gaps; a decade-wide young group can conceal narrower age changes. The dashboard and paper are not independent studies.

Read the bounded update

Spain: registered unemployment rose in more AI-exposed occupations

Source and version: González-González and coauthors, IEB Working Paper 2026/17, October 2026 version.
Population and place: 502 four-digit Spanish occupations.
Period: Monthly records, January 2021–March 2026.
AI measure: Occupation exposure mapped from Claude use and tasks; not Spanish firm adoption.
Outcome: Registered-unemployment counts and new-contract flows, not unemployment rates or linked worker transitions.
Method: Within-broad-occupation-family monthly difference-in-differences comparison.
Finding: Per 10-point higher exposure: 1.2% more registered unemployed workers in November 2022–November 2024; 1.7% later, imprecise. No persistent new-contract decline.
Causal status: Exposure-linked association, not verified AI-caused job loss.
Limit: Full-period pre-trend test rejects; the count cannot separate unemployment entries from slower exits. A different Spanish employment analysis finds growth across exposure groups through 2025.

Read the bounded update

Seven-country firm survey: little reported net employment change

Source and version: Yotzov and coauthors, NBER Working Paper 34836, revised October 2026; expands the February version, not an independent replication.
Population and place: Nearly 9,000 CFOs, CEOs and senior executives at firms in the U.S., U.K., Germany, Australia, Canada, Mexico and South Africa.
Period: Surveyed in summer 2026; reports on the preceding three years and expectations for the next three.
AI technology: Firm-reported AI use, not one named model or an occupation exposure score.
Outcome: Executive-attributed past firm employment effects and expected future effects.
Method: Multi-country executive survey; the accessible abstract does not show full country weights or response rates.
Finding: Seventy-eight percent of firms report AI use. Respondents report essentially no net AI employment effect so far, while more than 40% expect a positive or negative effect at their own firm.
Inference: Self-reported attribution and forecasts, not verified causal job counts.
Limit: Aggregate net effects can conceal firm and worker losses. Future effects have not happened; detailed sampling and country weights have not been verified.

Read the bounded update

U.S. employment: a gender gap within AI-exposed occupations

Source and version: Pew Research Center, 22 October 2026; technical appendix.
Population and place: U.S. CPS employment counts across 480 occupations, separated by sex and education.
Period: 2017–19 compared with January 2023–June 2026; 2020–22 excluded.
AI technology: Occupation task exposure to generative AI, not measured employer use.
Outcome: Employment growth, not layoffs, hiring or wages.
Method: Within-occupation pre/post regression with occupation and year controls, using O*NET task scores and CPS counts.
Finding: Per one-point higher exposure, men’s employment grew 6.3% and women’s 0.2%; for workers with high school or less, +8.8% versus −9.5%.
Causal status: Association. Other changes may explain the gap.
Limit: Exposure does not measure AI adoption, and the comparison cannot isolate the effect of AI.

Read the bounded update

U.S. job ads: more new AI duties, not proven new jobs

Source and version: Harris, SSRN working-paper abstract, revised 15 October 2026. Full paper not inspected.
Population and place: 18.6 million duty statements from U.S. ATS job pages archived by Common Crawl.
Period: 2019–2026; novelty benchmarked against early-2019 postings.
AI technology: Generative-AI-related duties and occupation-level AI use.
Outcome: Share of advertised duties classified as new; posted-pay comparison.
Method: Text comparison with a 2019 baseline and language-model classification of genuine new tasks versus wording changes.
Finding: In the most AI-using third of occupations, new-task statements rose from under 1% in 2022 to nearly 3% in 2026. Working-with-AI statements rose from 4% to 21% of AI-related duty statements. An apparent AI-duty posted-pay premium largely disappears within the same firm and title.
Inference: Descriptive change in advertised tasks, not a causal labor-market effect.
Limit: Archived ATS pages may miss vacancies; only the abstract was accessible, so classifier validation was not checked. No actual task performance, hire, job creation or wage receipt is measured.

A hiring warning changed ratings, not jobs

Source: Boyd-Swan and Reynolds, Annenberg EdWorkingPaper 26-1604, October 2026
Population and place: 1,750 U.S. for-profit professionals with hiring responsibilities recruited through Prolific.
Period: May–June 2026 survey.
AI technology: Generative AI discussed in a vignette about student coursework; actual student use was not observed.
Outcome: Stated willingness to interview and negotiate starting salary for constructed recent-graduate resumes.
Method: Random assignment to neutral, AI-as-aid or AI-as-substitute messages before two resume ratings.
Finding: The substitution message reduced ratings, especially for the high-GPA resume. The aid message produced no consistent penalty.
Inference: Causal effect of the message on hypothetical ratings, not a measured labor-market effect.
Limit: Opt-in panel, fictional candidates, redacted school and major; no real callback or wage offer.

U.S. payroll: the September exposure gap

Source and version: ADP Research, 21 October 2026, updating the Stanford–ADP Canaries Dashboard; not an independent replication of its earlier study.
Population and place: Workers mapped to occupations at U.S. ADP-payroll firms with five years of records; the panel excludes firm entry and exit and is not nationally representative.
Period: September 2026 versus September 2025.
AI measure: Occupation-level potential exposure, not observed use by employers.
Outcome: Employment on payroll, not hires, unemployment or AI-caused layoffs.
Comparison: Year-over-year change in high- and least-exposure occupation groups, split by age.
Finding: Ages 22–25: −3.7% versus +0.2%; ages 26–30: −2.8% versus +0.9%; all ages: −0.5% versus +1.0%.
Inference: The young-worker exposure gap remains visible in this panel; the update does not identify AI as its cause.
Limit: Selected firms, exposure rather than adoption, and no causal control for remote work or demand in this month’s descriptive comparison.

Read the update.

German full-time employment: a smaller early-career signal

Source and version: IAB Forum, 13 October 2026.
Population and place: About 22.3 million German full-time social-insurance employees aged 20–64; not trainees, marginal employees or the unemployed.
Period: Monthly records through December 2025, compared with 2022.
AI measure: Occupation task exposure to language models, not employer adoption.
Outcome: Full-time employment stock, not hiring or layoffs caused by AI.
Method and comparison: High versus lower exposure; adjusted models include remote-work suitability and cyclical vulnerability.
Finding: Ages 20–29: −3.8% versus −2.6% unadjusted. The adjusted AI-exposure association is negative but not statistically significant; cyclical weakness is stronger.
Inference: No broad full-time employment collapse in highly exposed German jobs through 2025.
Limit: Different country, period, employment coverage and age bands from the U.S. payroll update; no direct AI-use measure.

Remote work complicates the junior-hiring AI story

Source and version: Lambert and Schindler, LSE CEP Discussion Paper 2193, June 2026.
Population and place: Resume-derived new hires and online vacancy ads in the U.S., U.K., Canada and Australia.
Period: 2017–2025.
AI measure: Occupation-level generative-AI task exposure; not observed use by every employer.
Outcome: Junior share of new hires and share of ads requiring no more than three years of experience; not unemployment or total jobs.
Method and comparison: Observational difference-in-differences models compare AI and work-from-home exposure separately, then jointly; a further comparison uses remote or hybrid ads in 2021–2022.
Finding: The GenAI association with a lower junior share weakens sharply in joint models and often becomes statistically indistinguishable from zero; work-from-home exposure remains predictive.
Inference: Remote work is a serious competing explanation for some exposed-job hiring patterns, not proof that AI has no effect.
Limit: The sources are selected resume and posting panels, the exposure measures overlap strongly, and the study ends before the newer French hiring-flow result.

Read the comparison.

Swedish AI-skill hires: pay gap comes from firm sorting

Source and version: Hellsten, Ratio Working Paper 393, 16 October 2026; previously circulated as Returns to AI skills, not a separate study.
Population and place: External hires matched to Swedish vacancy ads and population registers; pay analyses restrict to selected private-sector firms and hires.
Period: 2014–2023 hires, with following-year earnings measured where available.
AI measure: Advertised skills for developing or implementing AI; ordinary chatbot use is explicitly excluded.
Outcome: Annual earnings of matched hires; no layoff or unemployment estimate.
Method and comparison: Observational vacancy-to-hire matching, occupation and worker controls, then firm fixed effects.
Finding: About 5–6% more earnings than occupational peers after basic controls; no remaining average gap within the same firm.
Inference: AI-skilled hires sort into higher-paying firms. This does not identify an individual return to acquiring a skill.
Limit: Restricted external-hire sample, imperfect possible vacancy matching, and a period ending in 2023. Not a test of chatbot training or current displacement.

AI access and education-based task performance

Source: Cruces and coauthors, full May 2026 paper; NBER lists an October revision not inspected in full.
Population and place: 1,174 Argentine adults aged 25–45 recruited online.
Period: Experiment reported in 2026.
AI measure: Randomized access to a GPT-4.1 assistant.
Outcome: Score on one incentivized business problem-solving task, plus an immediate unassisted follow-up; no hiring or wage outcome.
Method and comparison: Preregistered randomized access within education groups, outside firms.
Finding: The higher-education score gap fell from 0.548 to 0.139 standard deviations with access.
Inference: Causal for task access in this experiment, not for labor-market outcomes.
Limit: One simulated task; no observed employer decisions or long-run career outcomes.

France: young permanent hires in exposed occupations

Source: Palladino and coauthors, LIFT Working Paper 2026-02 and its policy note with exact percentages; the HEC policy note and Banque de France commentary describe the same study, not independent replications.
Population and place: Private-sector contract movements in metropolitan France, excluding agriculture, temporary agency work and household employers; focus on ages 21–29.
Period: January 2019–March 2026; comparison to November 2022.
AI measure: Occupation exposure score derived from French skill descriptions; one-time 2026 firm survey reports use intensity.
Outcome: Monthly permanent-contract hires and endings, not occupation-specific employment stock or unemployment.
Method: Descriptive comparison of highest and lowest exposure quintiles; linked observational firm comparison.
Finding: Young permanent hires fell 28.3% versus 7.5%; contract endings fell 23% versus 6%. At firms reporting intensive AI use, junior hiring was relatively weaker without a clear total-headcount difference.
Inference: Concentrated entry friction is observed; an AI-caused effect is not identified.
Limit: Post-pandemic tech correction, remote work, demand and apprenticeship changes may differ across occupations. Exposure does not observe use.

Anthropic: junior share of technical hires fell

Source: Wilmers and coauthors, Anthropic Institute technical report, October 2026.
Population and place: One U.S. AI company’s technical departments.
Period: Since early 2025, as reported in October 2026.
AI measure: Internal use telemetry and employee-directed agent sessions.
Outcome: Junior share of technical headcount and hires, not national jobs or layoffs.
Method: Internal descriptive records, manager survey and interviews; no causal hiring design.
Finding: Junior share of technical hires fell from 27% to 9%; technical hiring continued growing in absolute terms.
Limit: Firm growth, role mix and self-reporting limit attribution and generalization.

Read the 17 October task-use check.

U.S. worker generative AI use by occupation and task

Source: Bick, Blandin, Deming and Schumacher, Federal Reserve Bank of St. Louis Working Paper 2026-017; author explanation and abstract inspected, full PDF not accessed.
Population and place: Nearly 14,000 surveyed U.S. workers.
Period: Four quarterly waves, August 2025–May 2026.
AI measure: Self-reported generative AI use for regularly assisted job tasks.
Outcome: Adoption shares by occupation and task; no employment, wage or hours outcome.
Method and comparison: Worker survey using each occupation's ten most important O*NET tasks. Compares use rates across occupations and tasks, and with exposure scores.
Finding: At least 20% of workers reported AI use in more than 80% of occupations; fewer than 3% of measured tasks had use by a majority of workers performing them.
Inference: Descriptive adoption, not a causal employment effect.
Limit: Self-report, selected tasks and no measured staffing response.

Read the 16 October doctoral-career check.

U.S. STEM PhD writing and career paths

Source: Gross, Shvadron and Zhang, NBER Working Paper 35859, October 2026; author abstract inspected, full paper inaccessible.
Population and place: Near-census of U.S. STEM PhD graduates in the authors’ dataset.
Period: Dissertations from 2019–2026; subsequent employment linked.
AI measure: Detected AI-written text in dissertations, not verified student self-report.
Outcome: Academic research, non-tenure-track and industry career destinations; not unemployment or layoffs.
Method and comparison: Association among graduates in the same program and cohort, according to the abstract.
Finding: Detection in 29% of 2026 dissertations; detected writing is associated with less entry to academic research and more entry to other roles.
Inference: Career-path association, not a causal effect of AI assistance.
Limit: Career intentions and language background may differ. Full detector and model details, and exact outcome effect sizes, were not accessible.

Read the 14 October sector check.

U.S. sector task exposure and employment

Source: Johnston and Makridis, CESifo Working Paper 12579, March 2026; full paper inspected.
Population and place: U.S. state-by-industry groups; QCEW employment covers most payroll jobs.
Period: 2017–2024, compared with 2020.
AI measure: Pre-2020 occupational task mix mapped to potential LLM exposure, split into tasks AI might assist and tasks it might perform more independently; not observed adoption.
Outcome: Relative state-industry employment.
Method and comparison: Difference-in-differences event study comparing more- and less-exposed groups, with state-year and industry-year controls.
Finding: About 4% higher relative employment by 2024 per standard deviation of augmentable-task exposure; automatable-task estimate not statistically distinguishable from zero.
Inference: Exposure-linked result, not proof AI created jobs or caused no losses.
Limit: The automatable-task employment series had a positive raw pre-trend that was detrended; verified AI deployment and worker transitions are missing.

Read the 13 October version check.

German AI-skill ads and employment

Source: Peede and Stops, Labour Economics, 2026 abstract; 2024 IAB working-paper abstract is an earlier version, not separate evidence.
Population and place: German establishments with job ads linked to administrative establishment records.
Period: 2015–2023; mainly early AI adoption.
AI measure: Share of ads requiring AI-related skills, a proxy for activity rather than verified deployment.
Outcome: Employment growth overall and by job skill group; non-AI skill demand in ads.
Method and comparison: Establishment-level association. The latest full model was not accessible.
Finding: The 2026 abstract reports no employment-growth association and a small fall in some non-AI analytical skill requirements. The 2024 abstract had reported slightly higher high-skill job growth.
Inference: Descriptive association, not a causal estimate or a current generative-AI job-loss test.
Limit: Full 2026 methods and the reason for its changed skill-group summary are unverified.

U.S. AI use and earlier employment paths

Source: Makridis, Highfill and Samuels, BEA Working Paper WP2026-18, August 2026.
Population and place: U.S. state-industry cells in Gallup worker-use and official employment data.
Period: AI use measured 2025–2026; employment and output histories 2016–2024.
AI measure: Worker-reported frequent use; separate employer-reported industry measure.
Outcome: Employment, earnings and real output paths.
Method and comparison: Retrospective dynamic specifications comparing state-industry cells with different later use levels.
Finding: Higher-use cells had stronger post-2020 output paths and positive but imprecise employment differences. Industry-only employer-use analysis did not reproduce the output pattern.
Inference: Descriptive, not causal; later use cannot prove earlier job growth came from AI.
Limit: Reverse causality and local differences can explain part of the association.

Read the 12 October report on who could struggle to recover.

U.S. AI exposure and capacity to recover after job loss

Source: Manning and Aguirre, NBER Working Paper 34705, January 2026; author exposition at Brookings describes the same study.
Population and place: 356 occupations covering 95.9% of U.S. employment; occupation averages, not individual forecasts.
Period: Published January 2026; combined source datasets have different reference periods.
AI measure: Estimated task exposure to large language models, not observed employer adoption.
Outcome: Index of potential capacity to find work and limit earnings loss if displaced; no actual layoffs, transitions or post-displacement wages measured.
Method and comparison: Combines exposure with standardized liquid wealth, transferable skills, geographic labor-market density and age. Compares high-exposure occupations by average capacity.
Finding: 6.1 million workers, 4.2% of the covered workforce, are in high-exposure and low-capacity occupations; clerical and administrative roles feature heavily.
Inference: Conditional vulnerability map, not a causal AI employment effect.
Limit: Occupation averages conceal individual differences; future demand and newly created jobs could change the estimates.

Read the 11 October comparison of the two Copilot studies.

Copilot contract timing: fewer technical hires at selected large firms

Source: Dillon and Raymond, NBER Summer Institute 2026 conference abstract; full paper and slides not accessed.
Population and place: 5,425 large multinational firms with 45 million workers worldwide; estimate applies to firms induced to adopt by contract renewal.
Period: Within one year of adoption; exact calendar window not verified from the abstract.
AI technology: GitHub Copilot coding assistant.
Outcome: Number of technical-worker hires, separations and seniority mix; not the chance of any hire or wages.
Method and comparison: Adoption timing shifted by pre-existing enterprise software renewal cycles; full model and precision unavailable.
Finding: About 35% fewer technical hires among induced adopters within one year, with no rise in separations. Code-intensive hiring shifted toward senior workers.
Inference: Author-reported local adoption estimate, not a measured economy-wide or all-firm effect.
Limit: Conference abstract only. Full comparison, standard errors and robustness are unverified; one author is affiliated with Microsoft Research.

Copilot-linked firms: more months with an engineer hire

Source: Baird et al., Contemporary Economic Policy, April 2026; publisher abstract inspected, full text unavailable.
Population and place: Firms represented in LinkedIn and GitHub data; exact sample and geography not verified from the abstract.
Period: Exact study window not verified.
AI technology: Firm adoption of GitHub Copilot.
Outcome: Monthly probability of hiring a software engineer, not count of total hires, separations or wages.
Method and comparison: Adopters versus comparison firms in proprietary LinkedIn/GitHub records; full adjustment design unavailable.
Finding: Adoption associated with a 3–5% higher monthly probability of a software-engineer hire, driven by entry-level hires.
Inference: Association, not a causal employment effect or 3–5% more hires.
Limit: Authors work at LinkedIn and GitHub; proprietary data cannot be independently checked here. Different outcome and sample from the NBER conference study.

U.S. hiring professionals: a strong AI warning lowered survey ratings

Source: Boyd-Swan and Reynolds, EdWorkingPaper 26-1604, October 2026.
Population and place: 1,750 U.S. full-time, for-profit-sector professionals with hiring responsibilities recruited through Prolific; two fictional recent-graduate resumes each.
Period: Survey collected 21 May–8 June 2026.
AI measure: Randomly assigned information about student generative-AI use, not observed applicant or employer use.
Outcome: Stated interview likelihood on a seven-point scale and willingness to negotiate an initial salary offer upward; no actual invitations, hires or wages.
Method and comparison: Randomized control, mild AI-assistance frame or strong AI-as-substitute frame before resume ratings; high- versus low-GPA resumes.
Finding: The strong frame lowered ratings, especially for high-GPA resumes. For white-male resumes, the high-versus-low GPA interview-rating gap narrowed from 1.04 to 0.62 scale points.
Inference: Causal effect of the message on survey ratings in this selected sample, not on real hiring.
Limit: Prolific respondents and generic resumes do not represent all employers or real vacancies. Some subgroup estimates are imprecise; the study does not show how often the belief exists or whether AI caused job loss.

London applications: an AI credential improved interview invitations

Source: Drydakis, IZA@LISER Discussion Paper 18964, September 2026.
Population and place: Fictitious women applicants to 1,016 London private-sector, full-time vacancies not requiring a degree.
Period: Applications sent January–June 2025.
AI measure: A six-month sector-specific AI-related digital-skills certificate and CV skill description, not observed workplace AI use.
Outcome: Interview invitations; advertised pay of inviting vacancies. No offers, hires or realized wages.
Method and comparison: Three matched applications per vacancy: majority-group woman without AI certificate, underrepresented woman with nondigital certificate, and underrepresented woman with AI-related certificate. Four identity-signal experiments studied separately; pooled estimate reported.
Finding: Underrepresented applicants received invitations for 14.2% of applications with the AI-related certificate versus 11.1% without it; majority-group applicants without the certificate received 25.0%. Adjusted pooled signal difference: about 4 percentage points.
Inference: Randomized CV signals identify an early-stage invitation response in this setting, not a completed employment or wage effect.
Limit: No generic digital-skills comparator or majority-group applicant with the AI credential. The race-specific signal difference was not statistically significant. London non-degree vacancies may not generalize elsewhere; the experiment cannot value a real course.

Danish firm adopters: slower hiring relative to trend

Source: Bonin, Darougheh and Kuchler, Danmarks Nationalbank Working Paper 223, 16 September 2026; authors’ 8 October exposition of the same study.
Population and place: Surveyed Danish private non-financial firms with at least 10 employees, linked to monthly employer–employee records.
Period: First-time 2023 adoption tracked through September 2025; a 2024 cohort provides a timing check.
AI technology: Survey-reported use of AI, increasingly including generative tools for text and code; intensity varies and is not measured here.
Outcome: Full-time-equivalent employment relative to each firm’s prior trend, plus hiring versus incumbent retention.
Method and comparison: First-time adopters versus non-adopters within detailed industry-month groups; firm-specific pre-adoption trend and a second-cohort check. Observational, not randomized.
Finding: Authors’ CEPR account: about 9% below trend for 2023 adopters by September 2025, concentrated in firms below 100 employees and reduced hiring rather than layoffs. The central-bank abstract says 11% slower employment growth by late 2025; exact specification difference unverified.
Inference: An adoption-linked hiring warning in a narrow firm group, not proof AI caused the gap or a national jobs collapse.
Limit: Employer self-selection, demand and financing may confound the comparison. The small 2023-adopter group employed under 2% of Danish workers. The study does not estimate personal risk, wages or a causal economy-wide effect.

Danish chatbot encouragement: no extra junior-share decline through 2024

Source: Humlum and Vestergaard, Brookings, 6 October 2026; author account of the already tracked NBER working paper, not an independent study.
Population and place: About 25,000 Danish workers in 11 AI-exposed occupations, linked to administrative records.
Period: Survey and records through December 2024.
AI technology: AI chatbots; employer encouragement is the adoption indicator for this comparison.
Outcome: Early-career share of workplace employment.
Method: Compare chatbot-encouraging workplaces with others over time; not randomized adoption.
Finding: No differential decline; the authors rule out a gap larger than about one-third of a percentage point.
Limit: Does not explain the exposed occupations’ overall decline, later hiring or findings in other countries.

Manufacturing AI-skill ads: high posted pay, no measured worker raise

Source: Tito, Federal Reserve FEDS Note, 30 September 2026.
Population and place: U.S. manufacturing online job ads in Lightcast; production occupations examined separately.
Period: Skills and advertised wages through July 2026; broad AI-skill requests rose in the second half of 2025.
AI measure: Ads requesting broad AI-related skills, including machine learning; explicit generative-AI skills are a separate, much smaller category.
Outcome: Share of ads listing skills and average advertised wages, not accepted wages, hires or layoffs.
Method and comparison: Descriptive shares and unadjusted average posted pay for ads requesting AI skills versus all manufacturing ads.
Finding: Broad AI skills reached about 11% of manufacturing ads versus 8% across sectors in late 2025; explicit GenAI skills remained under 1%. AI-skill manufacturing ads showed about 70% higher average advertised pay, and production-occupation AI-skill ads about 30% since 2023.
Inference: Manufacturers are advertising for AI skills, including in some production roles. The pay difference is not a worker raise or a causal AI wage premium.
Limit: Different occupations, employers and locations may explain the wage gap. Pay is missing from many ads, and coverage changed over time. Online postings are recruiting activity, not completed hires or employment.

U.S. job ads: task text changes within comparable roles

Source: Wang, Wei and Wang, arXiv preprint, May 2026; full PDF inspected.
Population and place: 9,373,092 sampled U.S. Lightcast online postings from a raw file of more than 188 million ads across sectors.
Period: January 2021 to June 2025; decomposition shares reported from 2023Q3 onward.
AI measure: LLM-extracted posting tasks scored for potential GenAI assistance, not observed employer adoption.
Outcome: Average AI-exposure score of advertised tasks and its change through role mix, within-group task text and interaction. No hires or layoffs.
Method and comparison: Repeated sample within occupation, industry and seniority cells; Kitagawa decomposition of posting-mix versus within-cell exposure changes over time.
Finding: Across the absolute decline in measured exposure after 2023Q3, 52.01% comes from posting mix, 39.46% from within-cell task text and 8.54% from interaction. Junior postings split 60.15%, 18.22% and 21.63%.
Inference: Fixed occupation scores can miss changing advertised tasks. The percentages describe index movement, not displaced workers.
Limit: Not the same employer or job tracked over time; ad text may differ from real work. Exposure rose before ChatGPT. No causal AI effect or worker outcome is identified.

More startups, no aggregate startup-jobs gain

Source: Gupta, Qian, Simintzi and Sun, working paper, revised 2 July 2026
Population and place: U.S. existing startups and new firms in exposed markets.
Period: First two quarters after ChatGPT for incumbent employment; exact full window not verified.
AI measure: Pre-release generative-AI task exposure, not observed use at every firm.
Outcome: Existing-startup employment, new-firm entry, aggregate startup employment and workforce mix.
Method: Post-release comparison by pre-release exposure, per the abstract.
Finding: Exposed existing startups reduced employment, mainly junior roles. Smaller new entrants offset the loss; aggregate startup employment was unchanged in the study.
Inference: A bounded exposure-based market result, not a national AI-caused jobs estimate.
Limit: Only the latest abstract was inspected. Full sample, estimates, confidence intervals and trend checks remain unverified.

UK entry did not show a clear rise; founding teams shrank

Source: Agarwal, working paper, posted 25 September 2026
Population and place: New UK firms grouped by industry LLM exposure.
Period: Post-2022 entry; founding-team shift reported from 2025.
Outcome: Firm incorporation and founding-director counts, not employee jobs.
Method: Industry exposure comparison around generative-AI diffusion, per the abstract.
Finding: No statistically significant firm-entry change. A one-standard-deviation exposure increase was associated with 2.1% fewer directors and a 14.3% lower share of firms with three or more directors.
Inference: UK organizational contrast, not proof of no entry effect or a jobs result.
Limit: Only the abstract was inspected; full design, power and precision were not verified.

More startups in GenAI-exposed industries

Source: Bena, Bian and Giannetti, CEPR Discussion Paper 21982, 28 September 2026. Abstract and author summary inspected; full paper not accessed.
Population and place: Industry-level startup comparisons; geography and data coverage are not specified in the accessible abstract.
Period: After generative-AI diffusion; exact observation dates not stated in the accessible abstract.
AI measure: Industry task-level exposure, not observed adoption by each startup.
Outcome: Startup formation and reported industry-level employment and wage growth from new entry.
Method and comparison: Higher- versus lower-exposure industries after diffusion; identification details were not available in the inspected summary.
Finding: Twenty percent more startup formation in more-exposed industries. Individual entrants were smaller, while the authors report aggregate employment and wage growth from entry.
Inference: New firms may offset some displacement pressure. This is not a net-jobs verdict for all workers.
Limit: No causal claim, confidence interval, geography or exact labor effect can be verified from the accessible abstract. Incumbent losses and worker transitions need separate checks.

U.S. firms: exposure scores only partly predict AI use

Source: Dey and coauthors, U.S. Census Working Paper CES-26-61, September 2026.
Population and place: Matched single-location U.S. firms. About 15,500 firm-year observations under the revised AI-use question and 57,000 under the earlier wording.
Period: Occupation mix from 2018–2020 linked to AI-use reports from 2023 into early 2026; the main AI supplement ran November 2025–February 2026.
AI technology: Survey-reported AI use broadly, not generative AI alone. Fourteen task/occupation exposure measures.
Outcome: Whether firms reported AI use; no employment, hiring, wage or layoff outcome.
Method and comparison: Linked BLS occupational data and Census business survey responses; regression associations before and after year and detailed-industry controls.
Finding: A one-standard-deviation exposure increase corresponded to 4–11 percentage points higher reported adoption before controls and 1–8 points after controls. Exposure alone explained about 1–6% of cross-firm variation in use.
Inference: Exposure contains some information about adoption but cannot stand in for observed firm use.
Limit: Single-location firms only; survey wording changed in November 2025; sample linkage and self-reported use can limit generalization. No causal labor-market effect is estimated.

U.S. software ads: fewer junior openings relative to senior ones

Source: Westby, Modestino and Cheng, IZA Discussion Paper 18723, June 2026.
Population and place: U.S. online software-developer postings in Lightcast; the main regressions use 575,397 underlying ads.
Period: Main comparison: 12 months before and after ChatGPT’s November 2022 public release; longer descriptive data run January 2019–March 2025.
AI technology: ChatGPT public release is the time marker; employer-level AI adoption is not observed.
Outcome: Junior and senior vacancy counts and advertised experience and skill requirements; not hires, employment or layoffs.
Method and comparison: Event study and difference-in-differences comparing ads requiring 0–3 years with ads requiring 4+ years, plus other-occupation comparisons and a mechanical-engineering placebo.
Finding: Junior vacancies fell roughly 14–15% relative to senior vacancies in the main post-release window. Required experience rose within some job titles; remaining junior ads shifted toward problem solving, communication and detail rather than mainly AI-specific skills.
Inference: A software-specific entry-vacancy warning in an early transition period.
Limit: Ads are not completed hires. The wider post-pandemic software slowdown may still matter. The main model omits ads without stated experience, though a classification check gives a similar pattern. No causal AI or personal risk estimate follows.

Platform logs: the AI exposure score depends on who uses the service

Source: Yin and Ogut, arXiv version 2, May 2026; methodological preprint.
Population and place: U.S. ACS wage-and-salary worker panel, about 13.1 million person-years across 2015–2024; platform conversation shares from Claude and Copilot releases.
AI technology: Occupation scores that combine task applicability with AI-platform conversation shares.
Outcome: Sensitivity of an employment association to the platform input; not jobs lost or gained.
Method and comparison: The authors held the employment design, outcome, sample and task rubric fixed, then changed platform usage shares. They also reweighted platform occupation shares to the U.S. workforce mix.
Finding: Coefficients differed by a factor of 1.9 across tested inputs; consumer and enterprise channels within one vendor disagreed in sign. Workforce reweighting attenuated tested estimates by 42–93%.
Inference: Platform user selection can materially alter an exposure-based employment estimate.
Limit: This is a measurement diagnostic, not a causal AI job effect. Reweighting cannot remove within-occupation selection. It does not invalidate every type of exposure or adoption measure.

ChatGPT-assisted pitches: harder to judge experience in a controlled test

Source: Cowgill, Hernández-Lagos and Wright, IZA Discussion Paper 18442, March 2026; listed as published in Management Science in 2026.
Population and place: 343 pitch writers and 801 evaluators recruited through Prolific from English- and non-English-speaking contexts; not a representative labor-market sample.
Period: Spring and summer 2023.
AI technology: ChatGPT access for revising short pitches.
Outcome: Accuracy in estimating whether a pitch writer had prior domain experience; pitch ratings and stated demand for further verification.
Method and comparison: Controlled hiring and startup-investment exercises with 6,440 pooled evaluations. Writers produced an unassisted pitch first and then an assisted version; evaluators were assigned pitches and knew whether assistance was available.
Finding: Screening accuracy fell 4–9% in the pooled settings when ChatGPT assistance was available on average. For writers from non-English-speaking contexts, accuracy improved.
Inference: AI-written text can change how informative a pitch is; the direction depends on who receives the stronger boost.
Limit: Not a study of real job applications, employer decisions, offers, hires, wages or layoffs. The pooled estimate is not hiring-only, and fixed drafting order leaves a possible practice effect.

AI cover letters: better writing, no reported invitation gain

Source: Abbas Nejad, Musillo, Wicker and Zaccaria, Journal of Labor Economics, 2026.
Population and place: Job seekers and recruiters in two field experiments; the university summary identifies recruiters at four Dutch companies. Full sample details were not accessible in this run.
Period: Published in 2026; experiment dates not verified here.
AI technology: ChatGPT assistance for cover letters.
Outcome: Cover-letter quality and the study’s interview-invitation measure.
Method and comparison: Two field experiments comparing assisted and unassisted letters, plus a calibrated matching model. Only the official abstract and university summary were inspected.
Finding: Assistance improved letter quality, especially for lower-quality applicants, but did not raise the reported interview-invitation outcome. Recruiters valued strong unassisted letters more when AI use was disclosed.
Inference: Better-looking applications do not automatically improve a hiring-stage outcome.
Limit: The full accepted paper was unavailable to verify sample and outcome details. Less efficient matching is a model result; neither actual hires nor employment losses are established.

U.S. exposed occupations: higher posted pay, lower incumbent pay

Source: Rinz, Cleveland Fed Working Paper 26-24, September 2026
Population and place: U.S. occupations in Lightcast full-time job ads and Current Population Survey worker records.
Period: Worker-flow and earnings comparisons through 2026Q2; ad language through August 2026.
AI technology: Potential exposure of occupation tasks to large language models, plus AI words in ads; not verified firm deployment.
Outcome: AI mentions, posted wages, job ads, hires, separations and realized weekly earnings by worker status.
Method and comparison: Descriptive event-study and difference-in-differences comparisons across occupations with different LLM exposure, relative to pre-ChatGPT periods.
Finding: Real posted wages were 3.2% higher per exposure standard deviation by 2026Q2. In the later period, real wages for continuously employed workers were about 0.9% lower. New-hire pay had a positive but statistically imprecise estimate. Hires and separations both rose; net flows were not statistically distinguishable from zero.
Inference: A pay and churn split that advertised wages alone miss.
Limit: Exposure is not adoption. Other occupation-level shocks can explain the pattern; the paper does not identify a causal AI effect or national net job loss.

Philippine AI voice interviews: more offers in one randomized setting

Source: Jabarian and Henkel, revised working paper, September 2026
Population and place: 67,056 eligible applications for entry-level customer-service jobs at one recruitment firm across 19 Philippine cities.
Period: March to June 2025.
AI technology: AI voice agent conducting interviews; human recruiters made hiring decisions.
Outcome: Job offers, starts, short-term retention and productivity among hires.
Method and comparison: Random assignment to AI interviewer, human interviewer or choice arm; the main comparison is AI versus human interviews.
Finding: AI interviews raised offer likelihood by 12% relative to human interviews. Starts and short-term retention rose; measured productivity among hires did not fall.
Inference: Causal effect of changing the interviewer in this firm, not of AI on total employment.
Limit: One recruiter and occupation. It does not test AI-written applications, wages, layoffs or net job creation.

AI-assisted applications: a possible screening barrier, not a measured effect

Source: Ashlagi, Johari, Kleinberg and Murthy, arXiv preprint, September 2026
Population and place: None. This is a formal hiring model, not a study of a country, workers or employers.
Period: Not applicable.
Outcome: Predicted employer screening and candidate access to screening under stated assumptions.
Method: A theoretical model of applicants with different experience and job fit, noisy application materials and costly employer screening.
Finding: If AI makes application materials less informative, employers in the model rely more on prior experience. Inexperienced applicants who would fit the job can lose screening opportunities. A lower-cost intermediate assessment can restore a path to screening.
Inference: A plausible mechanism for an entry barrier when AI changes application writing.
Limit: The paper does not measure AI use, application quality, employer screening, interviews, offers, hires, wages, employment or layoffs. Its result depends on model assumptions.

U.S. AI-exposed job ads: faster advertised pay, a thinner junior share

Source: Kennedy, Indeed Hiring Lab, 17 September 2026
Population and place: Millions of U.S. job postings with an advertised salary, grouped by occupation and seniority.
Period: 2021 through mid-2026, with post-ChatGPT comparisons.
AI technology: Generative AI. The exposure score is the share of a role’s skills judged open to hybrid or full transformation.
Outcome: Advertised salary and the share of postings described as entry, mid or senior level.
Method: Difference-in-differences models with occupation or job-title and month fixed effects; a separate specification holds seniority mix constant within occupation.
Finding: Advertised pay in the most-exposed occupations rose about 46% from 2021, versus 25% in the least-exposed. Their entry-level posting share fell from 29% to 10%, while senior share rose from 22% to 47%. The post-ChatGPT pay premium is 5.7% in the occupation model, 4.7% by job title and 2.4% when seniority mix is held constant; the last estimate is not statistically significant.
Inference: A signal that U.S. advertised opportunity in exposed work may be tilting toward senior roles. It limits any claim that a higher pay headline means junior access improved.
Limit: Job advertisements are not accepted wages, hires, employment or layoffs. Exposure is not employer adoption. Seniority, occupation and industry composition can explain part of the pay pattern, so this is not a causal AI wage estimate.

Texas graduates: lower first-year employment and pay in more-exposed majors

Source: Dodini and Smith, Federal Reserve Bank of Dallas analysis, 22 September 2026
Population and place: Graduates of Texas public universities, joined to Texas workforce employment and earnings records.
Period: Graduate cohorts around ChatGPT’s late-2022 release; the reported results compare first-year outcomes through 2024.
AI technology: Generative AI. Major exposure is the average automatable-task share of jobs that requested the major in pre-ChatGPT online postings, using a Claude-task-based occupation score.
Outcome: Employment in Texas within one year of graduating, first-year earnings among graduates who found Texas employment, graduate-school enrollment and major enrollment.
Method: Event-study comparison of cohorts and majors with different pre-ChatGPT exposure. Student education records are linked to Texas workforce records; the authors report stable pre-ChatGPT differences.
Finding: After ChatGPT’s release, a 10-percentage-point higher automatable-task share was associated with a 1.7-point lower probability of Texas employment within a year and roughly 5% lower first-year earnings among employed graduates.
Inference: Association consistent with a weaker early entry route in the Texas sample. It independently echoes the broader Census graduate warning.
Limit: Task exposure is not observed employer adoption, and another post-2022 shock correlated with major exposure could help explain the gap. The records do not cover every U.S. graduate or work outside Texas; they do not establish a personal or national result.

AI adopters: a lower junior share, not a proven junior-job collapse

Source: Chandar and Klein Teeselink, Stanford Digital Economy Lab working paper, September 2026
Population and place: Foreign affiliates of multinational companies in 41 countries, measured in Revelio/LinkedIn employment records.
Period: Monthly panel from January 2021 through March 2026; adoption is tracked after ChatGPT’s November 2022 launch.
AI technology: Generative AI. A company is classified as an adopter after it posts a job advertisement involving GenAI.
Outcome: Junior and senior employment, workforce composition and total employment at an affiliate.
Method: Matched, instrumented event study. The instrument is GenAI adoption among other multinational parents near the focal parent’s headquarters.
Finding: By March 2026, adopting affiliates had a 1.9-percentage-point lower junior share and 6.7% higher senior employment than controls. Junior employment was estimated 2.5% lower but imprecisely; total employment was estimated 3.3% higher at marginal significance.
Inference: Evidence consistent with an adoption-linked shift toward senior labor in the sampled firms, rather than a broad layoff wave.
Limit: A GenAI job ad is a proxy, not the exact start of use. The causal claim depends on the instrument. LinkedIn records overrepresent managers and professionals, and the main estimates cover larger foreign multinationals.

Broad U.S. monitor: no clear AI jobs footprint yet

Source: Yale Budget Lab, “Tracking the Impact of AI on the Labor Market,” updated 14 October 2026
Population and place: Broad U.S. labor-market measures. The September version incorporated August 2026 Current Population Survey microdata; the October summary does not specify its updated data cutoff.
Period: Ongoing monitoring against historical and pre-AI patterns; the October page did not expose a verified latest data month.
Outcome: Occupational churn, AI exposure among unemployed workers, AI-use measures and an exposure-based synthetic difference-in-differences comparison.
Method: Ongoing descriptive monitoring plus a synthetic difference-in-differences analysis of occupational AI exposure.
Finding: Yale’s October summary reports no clear AI-related occupational shift, no connection between AI-use measures and employment or unemployment changes, and no clear exposure-based footprint.
Inference: A broad counter-signal to a claim of an already visible economy-wide U.S. AI jobs collapse.
Limit: The tracker does not randomize employer adoption or identify a named system’s causal effect. Broad measures can miss a concentrated effect on new entrants, a region, an occupation or job quality.

Swedish AI grants: more vacancy ads, no detected headcount effect

Source: Hellsten, Khanna, Lodefalk and Yakymovych, IZA Discussion Paper 18267, November 2025
Population and place: 190 Swedish small- and mid-sized firms that applied to Vinnova for first-time AI-project support: 53 were awarded and 137 were not.
Period: Grant decisions in 2019–2020; job-vacancy data from 2017–2024 and outcomes up to five years after award.
Outcome: Whether a firm posted a vacancy, number of vacancies, total employment, hiring and separations.
Method: Synthetic difference-in-differences comparing awardees with non-awarded applicants, using near-universal Swedish vacancy data and population registers.
Finding: Five years after award, recipients were 24 percentage points more likely to post a vacancy and 22.5 points more likely to post a white-collar vacancy. The study found no statistically detectable effect on total employment, hiring or separations.
Inference: A close-to-causal grant effect on recruitment effort among Swedish AI-project applicants, not a simple employment verdict.
Limit: The intervention is a grant, not random use of a named system, and it largely predates broad chatbot diffusion. The sample is small, vacancies are not hires, and Swedish labor-market frictions may not travel.

A 4% cognitive-employment decline—inside one 2030 scenario

Source: Korinek, Jones, Sacher, Cotter and McCrory, CEPR Discussion Paper 21939, September 2026
Population and place: Not a worker or firm outcome sample. The paper models an economy and separately surveys U.S. adults about AI expectations.
Period: Illustrated paths from 2026 to 2030.
Outcome: Modelled cognitive employment, GDP, labor share, wages, reallocation and unemployment.
Method: A structural scenario framework that maps assumed AI capability and automation paths into economic outcomes; it illustrates modest, substantial and extreme change.
Finding: In the authors’ substantial scenario, cognitive employment declines 4% by 2030. They say median U.S. survey expectations are consistent with that path.
Inference: A useful stress test for a possible future—not a measured 2026 labor-market result.
Limit: The figure depends on the scenario’s assumptions and has no stated probability on the accessible summary page. It does not count layoffs, identify a worker’s risk or establish that any observed hiring change was caused by AI.

Danish firms reporting AI adoption: slower growth through lower recruitment

Source: Bonin, Darougheh and Kuchler, Danmarks Nationalbank Working Paper 223, September 2026
Population and place: Danish firms reporting AI use, matched to monthly employer–employee records; the reported effect is concentrated in smaller firms.
Period: Firms first reporting AI adoption in 2023, followed through late 2025.
Outcome: Employment growth and recruitment, including recruitment in AI-exposed occupations.
Method: Firm-level event-study comparison of adopters and non-adopters using prior employment trends and industry differences.
Finding: By late 2025, 2023 adopters had employment growth about 11% below non-adopters relative to prior trends. The gap reflects lower recruitment; aggregate Danish employment had not shown a major change.
Inference: A high-value adoption-linked warning that slower hiring can appear before an economy-wide unemployment signal.
Limit: Adoption is not random, so firm differences or other concurrent changes may explain part of the gap. Denmark, firm size and the period through late 2025 limit generalization; this is not a layoff count or a personal forecast.

California claims: no August surge in high-exposure unemployment claims

Source: California Employment Development Department and California Policy Lab AI-Unemployment Tracker, August 2026 update
Population and place: California unemployment-insurance claimants, grouped by the AI exposure of their self-reported last occupation.
Period: Monthly claims through August 2026; the update reports three-month moving averages.
Outcome: New initial unemployment-insurance claims, not employment, hiring or all job losses.
Method: Descriptive tracking by potential AI exposure and a task-use measure derived from Claude conversations, with linked occupation crosswalks.
Finding: High-exposure three-month claims declined about 1.2% under the potential measure and 1.0% under the observed-use measure, within recent historical fluctuations.
Inference: A current California counter-signal to a broad exposure-linked unemployment surge.
Limit: Exposure and task-use scores are not employer adoption. Claims exclude many workers who do not apply or are ineligible, and self-reported occupation codes can be missing or inaccurate.

Recent graduates: no summer unemployment spike in a new U.S. check

Source: Fairlie and Wu, IZA Discussion Paper 18945, September 2026
Population and place: Recent U.S. college graduates in Current Population Survey microdata, compared with older college graduates and young adults without a college degree.
Period: June through August 2026, compared with earlier summer months.
Outcome: Unemployment, plus a broader measure that includes “sidelined unemployed.”
Method: Difference-in-differences and event-study interaction models using two comparison groups; occupation interactions test AI exposure and remote-work availability.
Finding: No statistically significant summer-2026 relative unemployment increase appears in the reported comparisons, including the broader unemployment measure. The paper finds some evidence of a positive relationship with remote-work availability in occupation interactions.
Inference: A material U.S. counter-signal to a broad graduate-unemployment claim.
Limit: The paper does not observe employer AI adoption or measure initial job matches, earnings, vacancies or later career progression. It is a short-window unemployment result, not an all-clear for entry-level work.

Norway: no robust young-worker displacement yet

Source: Facius and Iacono, CESifo Working Paper 12752, June 2026
Population and place: Norwegian private-sector workers with a full-time main job; population-wide employer-employee records, focused on ages 22–25.
Period: January 2015 to March 2025, around ChatGPT’s November 2022 release.
Outcome: Employment, wages and labor-market mobility.
Method: Within-firm composition difference-in-differences, occupation-level synthetic difference-in-differences and a firm shift-share design, using two occupational exposure indices.
Finding: Young-worker employment estimates are negative but not individually statistically significant across the three designs. Backdated placebos can be larger than the post-ChatGPT estimate.
Inference: A serious cross-country limit on a universal early-career AI displacement story.
Limit: Norway is not the U.S. The study maps U.S.-based exposure scores to Norwegian occupations and does not observe firm-level AI adoption; late estimates trend negative and the post-period is still short.

European firms: productivity up, no adverse headcount result

Source: Aldasoro and colleagues, EIB Working Paper 2026/02, January 2026
Population and place: More than 12,000 non-financial firms in the European Union and United States.
Period: Pooled firm cross-sections from 2019 to 2024.
Outcome: Firm labor productivity and employment.
Method: Matched EIBIS-ORBIS firm data; the paper instruments EU-firm AI adoption with adoption rates of matched U.S. peers.
Finding: The paper reports 4% higher labor productivity and no adverse firm-level employment effect in its instrumented specification.
Inference: A broad short-run counterpoint to a firm-level job-collapse claim.
Limit: The measure includes big-data analytics and AI, not generative AI alone. The pooled design does not measure entry-level hiring, occupation mix or a randomized rollout.

Actual AI spending: heavy adopters added headcount

Source: Kharazian, Simon and Stevens, Ramp Economics Lab and Revelio Labs working-paper presentation, June 2026
Population and place: More than 21,000 Ramp-linked U.S. firms, joined to Revelio workforce data.
Period: Firm outcomes over the two years after the study’s adoption date.
Outcome: Total and entry-level headcount.
Method: Firm-level comparison by recorded AI-vendor spending intensity; high intensity is the top third of spend per employee in the first three post-adoption months.
Finding: The high-intensity group grew total headcount 10.2% and entry-level headcount 12%; low-intensity adopters had no statistically significant change.
Inference: A valuable actual-spending counterpoint to exposure-only studies, but still an association.
Limit: The accessible report is company-authored and uses proprietary Ramp-linked data. High-intensity adopters were already larger, more technical, more venture-backed and faster growing; the result is not a causal or economy-wide estimate.

Young workers: a 19% relative payroll gap

Source: Brynjolfsson, Chandar and Chen, Stanford Digital Economy Lab, revised August 2026
Population and place: Millions of U.S. workers in ADP payroll data; focus on workers aged 22–25.
Period: Employment trends through June 2026, compared with the period after ChatGPT's release.
Outcome: Employment and base pay by age and occupational AI exposure.
Method: Descriptive payroll comparisons between more- and less-exposed occupations, with robustness checks and alternative controls.
Finding: Young workers in highly exposed occupations were about 19% below the employment level they would have reached by keeping pace with less-exposed peers. Experienced workers showed no comparable gap.
Inference: A serious descriptive early-career warning, not a causal estimate.
Limit: Exposure is not adoption. Pre-trends, education, firm-hiring controls and the difference between the ADP sample and national benchmarks leave attribution and generalization unresolved.

Remote work: a rival explanation for junior-hiring decline

Source: Lambert and Schindler, University of Warwick working paper, September 2026
Population and place: New hires and online postings in the U.S., U.K., Canada and Australia.
Period: 2017–2025.
Outcome: Junior share of new hires and share of ads requiring limited experience.
Method: Difference-in-differences at occupation, region and firm level, testing GenAI and working-from-home exposure separately and jointly.
Finding: Each exposure separately predicts a lower junior share. Jointly, the remote-work effect remains while the GenAI coefficient sharply attenuates and is often not statistically distinct from zero.
Inference: A credible attribution challenge to an exposure-only AI explanation.
Limit: A working-paper result with different countries, outcomes and exposure measures from the U.S. payroll and Census studies. It does not show that remote work fully explains those results.

New graduates: lower employment and pay in the most exposed majors

Source: Orr, Tucker and Warren, U.S. Census Bureau working paper, September 2026
Population and place: Roughly 6.7 million bachelor’s graduates from more than 350 institutions in 24 U.S. states; a subset of U.S. higher education.
Period: Graduation records from 2016–2024 and post-graduation outcomes around ChatGPT's late-2022 release.
Outcome: Initial employment, full-quarter earnings, industry and job switching.
Method: Event-study and fixed-effects comparisons of majors with different pre-ChatGPT occupational task-exposure scores.
Finding: The most-exposed major decile had a five-point lower initial-employment rate and 13% lower full-quarter initial earnings than the bottom six deciles after ChatGPT's release.
Inference: Association consistent with a concentrated disruption at labor-market entry.
Limit: Exposure is not adoption. The design cannot rule out every time-varying shock correlated with major exposure, and the institutions are not all U.S. colleges.

Early-career hiring: U.S. result with unresolved attribution

Source: Lee C. Tucker, U.S. Census Bureau working paper, April 2026
Population and place: Private-sector workers aged 22–24 in industry-state cells across 45 U.S. states.
Period: Quarterly data through 2025 Q2; compared around ChatGPT's November 2022 release.
Outcome: Hires, employment, earnings and job flows.
Method: Event studies and regression comparisons of industry-state cells with different occupational AI-exposure scores.
Finding: Early-career hires fell immediately by 9% relative to less-exposed cells; regression-adjusted employment in the most-exposed quintile was 12% lower after 10 quarters.
Inference: Association consistent with reduced relative demand; not a clean causal estimate of AI adoption.
Limit: Exposure is not observed use. Pre-existing pandemic-era shifts, remote work and education trends remain plausible contributors.

Texas online postings: a regional AI-related pullback

Source: Dodini and Smith, Federal Reserve Bank of Dallas analysis, September 2026
Population and place: Texas employers’ Lightcast online job postings; a balanced panel separately follows incumbent firms.
Period: Quarterly postings after ChatGPT’s November 2022 release, with incumbent-firm results through early 2026.
Outcome: Online job postings: a measure of advertised labor demand, not realized hiring or employment.
Method: Within-industry comparison of occupations with different shares of tasks classified as automatable by a Claude-use-based index; firm exposure is set from pre-ChatGPT posted occupations.
Finding: More-exposed positions had about 8% fewer postings relative to less-exposed positions by 2025 Q1. The authors calculate a 2.6% aggregate effect on Texas Lightcast postings in 2025; incumbent firms show an 8–9% pullback by early 2026.
Inference: A serious regional, exposure-linked warning for advertised labor demand.
Limit: Claude task use is not verified employer adoption. Lightcast underrepresents some work and postings are not hires, jobs or layoffs. A broad U.S. posting analysis finds little distinct AI decline, so this does not establish a national causal effect.

Broad U.S. postings: no distinct AI drop yet

Source: Audoly, Guerin and Topa, Federal Reserve research analysis, June 2026
Population and place: U.S. postings collected by Lightcast from career pages, job boards and listing aggregators.
Period: Compared after ChatGPT's late-2022 release.
Outcome: Labor demand as represented by job postings.
Method: Tests whether postings for more AI-exposed occupations declined disproportionately.
Finding: Overall hiring slowed, but the analysis finds little indication of a distinct AI-driven decline in postings.
Inference: A broad counter-signal to claims of an economy-wide AI hiring collapse.
Limit: Postings are not realized hires and do not isolate new graduates, college majors or unadvertised recruitment.

U.S. task adoption: broad reach, shallow use

Source: Bick, Blandin, Deming and Schumacher, St. Louis Fed analysis, September 2026
Population and place: Nearly 14,000 U.S. workers in four Real-Time Population Survey waves.
Period: August 2025 to May 2026.
Outcome: Reported workplace AI use by detailed occupation and task.
Method: National survey linked to occupations and the ten O*NET tasks rated most important for each occupation.
Finding: At least one in five workers used AI in more than 80% of occupations and on more than 40% of tasks, but most adoption rates stayed below 50%.
Inference: Actual use differs materially within exposed occupations.
Limit: Self-reported use, not a measure of employment, wages or causal labor-market effects.

Firm AI use: broad enough to watch, too early to call a headcount collapse

Source: Bonney and colleagues, U.S. Census Bureau working paper, April 2026
Population and place: U.S. employer firms in the Business Trends and Outlook Survey.
Period: November 2025 to January 2026 reference period.
Outcome: Reported AI use, business functions, worker tasks and AI-related employment changes.
Method: Nationally representative firm survey with firm- and employment-weighted estimates.
Finding: 18% of firms, or 32% weighted by employment, reported AI use in a business function. Among AI users, 66% reported augmentation only and 2% reported an AI-related employment decrease.
Inference: Descriptive survey evidence, not a time-series or causal employment estimate.
Limit: Firms report their own attribution during one period; indirect or later effects may not appear.

Danish pay and hours: no detected effect; task change

Source: Humlum and Vestergaard, NBER Working Paper 33777, revised March 2026
Population and place: Danish workers and workplaces in 11 exposed occupations, linked to adoption surveys.
Period: The first two years after ChatGPT's launch.
Outcome: Earnings, recorded hours, reported productivity, tasks and occupational movement.
Method: Difference-in-differences using linked administrative records and adoption surveys.
Finding: No detectable differential earnings or recorded-hours effect; the estimates rule out effects larger than 2%. New AI-related tasks and occupational movement appeared among adopters.
Inference: Strong short-run result for the study's outcomes and setting.
Limit: It does not estimate U.S. entry hiring or establish that later effects will remain small.

Global exposure measures task potential, not layoffs

Source: ILO Working Paper 140, May 2025
Population and place: Global occupations mapped to ISCO-08 tasks.
Outcome: Potential generative-AI exposure of occupational tasks.
Method: Task-level scoring using worker input, expert review and AI-assisted prediction.
Finding: One in four workers are in occupations with some exposure; 3.3% of global employment is in the highest exposure category.
Inference: Potential for task change across occupations.
Limit: No observed employment, unemployment, hiring, wage or hour outcome is measured.