US Workforce AI Exposure

arXiv: 59% of O*NET tasks digitally feasible for RL automation

40.7% of tasks fail the physical feasibility gate and receive a score of zero, while gate-passing tasks have a conditional mean of 45.5.

NYT/OpenAI: GDPVal models reach 80%+ win rate vs human pros

When we originally released GDPVal, which was just a few months ago, none of the models were yet on par with human experts. Months later, we have over an 80 percent win rate compared to human professionals.

LLM model agreement on occupational AI exposure as low as 57%

Replicating the dominant rubric with three frontier models on identical tasks, we find a 3.6-fold divergence in mean exposure with agreement as low as 57%.

Anthropic: top-quartile exposure workers 3x more likely to fear job loss

For every 10-percentage-point increase in exposure, perceived job threat increased by 1.3 percentage points. People in the top 25% of exposure mentioned the worry three times as often as those in the bottom 25%.

OpenAI: 66pp capability overhang — 90% theoretical vs 23.8% realized exposure in high-automation-risk jobs

Jobs at high automation risk: gap 66.2 pp, 23.8% realized, 90.0% theoretical. Jobs that grow with AI: gap 49.7 pp, 22.7% realized, 72.4% theoretical. Jobs that will reorganize: gap 52.3 pp, 14.9% realized, 67.1% theoretical. Jobs with less immediate change: gap 21.0 pp, 6.4% realized, 27.4% theoretical. Across every job category, current usage lags behind the possible. Exposure alone is a weak predictor of immediate labor market pressure.

Census/QWI: Monetary policy explains ≤¼ of early-career employment gap; AI effect persists

At best, differences in monetary policy sensitivity can explain about one fourth of the regression-adjusted gap in employment, but monetary policy sensitivity does not appear to predict that AI-exposed firms would reduce their early career hiring activity more than others.

Yale Budget Lab: Unemployed workers in occupations where 25-35% of tasks are AI-performable, invariant to duration

The share of workers in the lowest, middle, and highest occupational exposure groups stay stable. Irrespective of the duration of unemployment, unemployed workers were in occupations where about 25 to 35 percent of tasks, on average, could be performed by generative AI. Both samples indicate that observed usage is more likely to be associated with automation than augmentation.

NYT: Economists shifting from dismissive to 'it's coming'; policy unprepared

Most still do not see much evidence that A.I. is disrupting the job market. But they are starting to take seriously the possibility that it could someday soon.

Fed SBU: 78% of US labor force works at an AI-adopting firm (employment-weighted)

The SBU estimates an employment-weighted firm AI adoption rate of around 78 percent and an LLM adoption rate of about 54 percent. In this context, employment weighting approximates the share of the labor force working at firms that have adopted AI.

Brookings: STARs are 43% of all US workers in top AI exposure quartile

Of America's ~70M STARs (workers skilled through alternative routes, no four-year degree): 15.6M work in roles in the top quartile of AI exposure (43% of all top-quartile workers); 11M of those are in Gateway occupations — the stepping-stone roles connecting entry-level to higher-wage work — with 6 Gateway occupations alone accounting for ~8M of them. STARs are 62.3% of all Gateway-occupation workers. Across Destination occupations, 12.9M workers (~1/3) are highly exposed, including sales reps, accountants, financial managers. Only 51% of Gateway-to-Destination career pathways AVOID high AI exposure. 3.5M STARs are both highly exposed AND have low adaptive capacity (67% of all such workers). 23M STARs have low adaptive capacity overall (68% of all such workers). Highest pathway-exposure metros: Palm Bay FL (35.5%), Cape Coral FL (34.7%), Jacksonville (33%), Albany NY (32.8%), Harrisburg (32.6%), Providence (30.1%). 73% of US workers live and work in the same county, so disruption — and remediation — will be place-specific. Uses Anthropic's observed-exposure measure on Opportunity@Work pathway taxonomy.

Hosseini/Lichtinger: 21.3% of O*NET tasks >80% automatable; avg 10pp labor pool expansion

We define a binary automation exposure indicator that equals one for tasks in categories 4–5, i.e., more than 80 percent automatable. Under this definition, 4,107 tasks (21.3 percent) are classified as exposed. The mean PSS is about 0.11, implying that for the average occupation, GenAI expands the qualified labor pool by 10 percentage points of the workforce.

FRI: 69 economists assign 61% prob to moderate/rapid AI capabilities by 2030

69 economists assign 61.4% probability to moderate or rapid AI capability progress by 2030. In the rapid scenario, AI surpasses humans on most cognitive and physical tasks.

MIT/CCI: 92% of AI apps map to only 6.8% of 39,603 work activities; exposure deep but narrow

92% of AI applications map to only 6.8% of 39,603 classified work activities. AI apps grew 6x from 2022-2024 but activity coverage expanded only 1.2x. 75% of AI market value concentrated in software/information tasks. Based on O*NET 29.1 and TAAFT dataset.

Otgonsuren: structured-task industries face sharpest AI employment cuts

Industries with high volumes of structured, repetitive tasks face the sharpest near-term employment contractions.

Tufts Digital Planet: 4.9M 'tipping point' workers in 33 occupations swing to >40% displacement

There are 4.9 million 'tipping point' workers -- spanning 33 occupations that swing from <10% to >40% displacement in the next 2-5 years across the US.

Anthropic: 49% of jobs have 25%+ of tasks performed using Claude

About 49% of jobs have seen at least a quarter of their tasks performed using Claude.

Imas/Shukla: Exposure != displacement; O-ring complementarities can raise wages under partial automation

Two jobs with identical exposure scores can have completely opposite displacement risks depending on whether their tasks are complements, whether demand for their output is elastic or inelastic, and the incentives of the firm to invest in automation.

Fed: >98% of coder employment in top AI-exposure quintile

The coding occupations are overwhelmingly in the high exposure group according to both metrics, with more than 98 percent of coder employment in the highest quintiles.

Yale Budget Lab: OpenAI exposure quintiles stable; no shift in worker distribution since ChatGPT

The share of workers in the lowest, middle, and highest occupational exposure groups stay stable. Even when specifically examining the unemployed population, there is no clear growth in exposure to generative AI.

GS Research (Briggs): AI can automate tasks accounting for 25% of all US work hours

In the US, AI can potentially automate tasks that account for 25% of all work hours, Briggs' team finds.

Fed/Duke: Large firms expect AI workforce cuts; small firms expect modest gains

Employment effects heterogeneous by firm size: large companies expect AI-driven workforce reductions while smaller firms anticipate modest employment growth. Compositional reallocation of labor both within and across firms, with routine clerical roles declining and skilled-technical roles increasing.

Anthropic: ~70% of workers have some observed AI task coverage; theoretical far exceeds actual

At the bottom end, 30% of workers have zero coverage, as their tasks appeared too infrequently in our data to meet the minimum threshold.

CMU/Stanford: Agent benchmarks target Computer/Math domain — only 7.6% of US employment

Across all benchmarks, examples collectively cover a limited 56.5% of the domain taxonomy but a substantially broader 85.4% of the skill taxonomy.

KPMG: 57% expect humans to manage/direct AI agents

87% Upskilling/reskilling current workforce

MIT FutureTech: 80-95% AI task success projected by 2029 (N=17,205 evals)

AI models successfully complete tasks that take humans approximately 3-4 hours with about a 50% success rate, increasing to about 65% by 2025-Q3. Most text-based tasks projected to reach 80-95% AI success by 2029. 62.6% of O*NET tasks classified as having at least 10% LLM time-savings potential. Based on 17,205 expert evaluations across 3,000+ O*NET tasks and 40+ LLMs.

MIT/USC: Pro se federal filings 11%→16.8%; AI-text in 18% of complaints by 2026

the national non-prisoner pro se filing share rose sharply from its approximately 11% historical steady state to 16.8% in fiscal year 2025, a gain that has no precedent in 25 years of administrative records. ... the share of complaints containing AI-generated text has been rising monotonically from 1.0% in 2023 to 18.0% in early 2026.

Yale Budget Lab: 6 exposure metrics agree on who's exposed, disagree on magnitude (778 occupations)

AI exposure metrics broadly agree with each other, but that they disagree with each other more on highly exposed occupations. The key point of disagreement between different AI exposure metrics is in the magnitude of exposure, not whether an occupation is exposed.

Lightcast: Most-exposed roles 70-78% (office/editorial); least 6-14% (healthcare)

Many kinds of writing and editing jobs are among those most exposed, with over 70% of their core skills potentially affected. On the other end of the spectrum, healthcare and first responder jobs dominate the low end of the rankings.

Indeed: 45% of data/analytics postings and 20%+ of dev/IT/R&D now mention AI

45% of data & analytics job postings mentioned AI, the highest among all sectors analyzed...software development, IT systems & solutions, and scientific research & development each mentioned AI 20% or more of the time

Brookings: Admin support has highest exposure (52.5%) + lowest adaptability

Administrative support occupations have lower adaptive capacity (0.360) combined with the highest AI exposure of any major occupation group (0.525).

the overall Generative AI exposure of the tasks in all white-collar occupations is 40% versus 9% for blue-collar and service occupations

NBER: AI exposure and adaptive capacity positively correlated (r=0.502); most exposed workers best positioned

We find a positive correlation (r = 0.502) between AI exposure and a novel measure of worker adaptive capacity to displacement. Higher-income, highly skilled workers in professional occupations typically possess characteristics that enable successful navigation of job transitions.

Computer/mathematical tasks account for ~33% of all Claude.ai conversations and ~50% of API traffic, indicating concentrated AI impact on tech-adjacent roles.

Cognizant: exposure growing 9%/yr (up from 2%); 30% above 2032 projections

93% of all occupations analyzed have at least one task with significant AI exposure. Education task exposure jumped from 11% to 49% — a 4.5x increase. Based on reassessment of 18,000 tasks across 1,000 O*NET occupations.

In recent years, retail trade has the lowest share of tasks that have been automated — around 50% — whereas computers/electronics has the highest share — more than 85%.

Deloitte: 84% of firms have not redesigned jobs around AI (n=3,235)

Despite high expectations for automation, 84% of companies have not redesigned jobs or the nature of work itself around AI capabilities.

LLM adoption among U.S. workers increased from 30.1% to 38.3% between December 2024 and December 2025. Small effects on wages in exposed occupations; no significant effects on job openings or total jobs.

NBER (Gans/Goldfarb): Linear exposure indices overstate displacement when tasks are quality complements

Widely-used exposure indices, which aggregate task-level automation risk using linear formulas, will overstate displacement when tasks are complements. The relevant object is not average task exposure but the structure of bottlenecks and how automation reshapes worker time around them.

OECD: ~25% of OECD workers exposed to generative AI

Around a quarter of workers in OECD Member countries are exposed to generative AI, meaning 20% of their job tasks could be done at least 50% faster with the help of generative AI.

MIT Iceberg Index: 11.7% of US wage value exposed to AI across admin, finance, professional services

Analysis of Bureau of Labor Statistics skill taxonomies reveals that current AI systems can technically perform approximately 16 percent of classified labor tasks. The Iceberg Index for digital AI shows values averaging 11.7%—five times larger than the 2.2% Surface Index.

ILO review: exposure converges on high-wage jobs

Productivity gains 20-60% in controlled RCTs, 15-30% in field experiments; AI exposure measures converge toward high-wage jobs being most exposed.

In about 40% of employment in exposed occupations, at least 50% of tasks will be replaceable. Average labor-cost savings ~25% from current tools, potentially 40% as systems improve.

NBER (Brynjolfsson/Hitzig): AI expanding codifiable knowledge frontier, increasing exposure

Transformative AI sharply expands what counts as codifiable -- and therefore transferable -- 'local knowledge,' in three main ways: it makes explicit knowledge more accessible to decision-makers, it increasingly extracts tacit know-how once embedded in human perception and practice, and it generates machine-native knowledge.

MIT (Autor/Thompson): Cross-occupation expertise premium NEARLY DOUBLED 1980-2018: 1σ expertise = +16 log pts wages in 1980, +31 log pts in 2018

Introduces a content-agnostic expertise measure based on the Efficient Coding Hypothesis applied to 1977 DOT and 2018 O*NET task descriptions; uses OpenAI text-embedding-3-small with 0.95 caliper to identify which tasks were removed, retained, or added across 303 Census occupations. Findings on US task composition 1977-2018: routine share of tasks fell from 50.4% to 32.2%, abstract share rose from 33.2% to 53.6%, manual share fell from 16.4% to 14.2%. 66% of removed tasks were routine; 77% of added tasks were abstract. Cross-occupation expertise wage premium nearly doubled: 1σ expertise = +16 log pts wages in 1980, +31 log pts in 2018 (R² 0.32 → 0.49). Alternative framework to Anthropic/OpenAI/Webb exposure measures — measures expertise required by tasks rather than AI capability to perform them.

678 occupations having on average 23% of their tasks exposed to Generative AI

OECD: One in three OECD job vacancies have high AI exposure (2025)

With one in three job vacancies having high AI exposure, a significant share of jobs in OECD economies are influenced by the rise of AI. Only a small percentage of training courses currently deliver AI content.

PNAS Nexus: AI-exposed occupations face higher unemployment

Workers in AI-exposed occupations face significantly higher unemployment risk.

Brynjolfsson/Shao: Workers want 46% of tasks automated; preferences vary by task type

Workers want 46% of their tasks automated; preferences vary by task type and worker characteristics.

Oxford/OII: Complementary effects 1.7x larger than substitution; AI demand doubling → 5% rise in complementary skill demand

When AI demand doubles in a job ad, demand for complementary skills rises by 5%. Complementary effects are 1.7x larger than substitution effects. Based on 12 million US job vacancies 2018-2023.

17-36% of worker skills exposed at moderate-to-high AI capability level.

JEMS/Census ABS: 18.2% of workers at AI-using firms by 2017 (n=850K firms)

Employment-weighted adoption was just over 18%. AI use in production was found in every sector of the economy. Manufacturing and information led at roughly 12% each. Based on the 2018 Annual Business Survey of 850,000 firms.

OpenAI/UPenn: 80% of workers have ≥10% tasks affected

~80% of the US workforce could have at least 10% of their tasks affected by GPTs. Legal, accounting, and financial analysis are among the highest-exposure occupations.

Almost 40% of global employment is exposed to AI, with advanced economies more affected.

LinkedIn: 55% of members hold jobs impacted by generative AI

55% of LinkedIn members hold jobs that stand to be impacted by generative AI. By 2030, the skills required for jobs will change by up to 65%.

IMF (Pizzinelli): complementarity adjustment shrinks AE–EM exposure gap

AEs face higher exposure than EMs due to a higher employment share in professional and managerial occupations. However, when accounting for potential complementarity, differences in exposure across countries are more muted. Within countries… Women and highly educated workers face greater occupational exposure to AI, at both high and low complementarity. Workers in the upper tail of the earnings distribution are more likely to be in occupations with high exposure but also high potential complementarity.

27% of jobs are in occupations at high risk of automation across OECD countries.

Roughly two-thirds of current jobs are exposed to some degree of AI automation; generative AI could substitute up to one-fourth of current work.

SMJ (Felten/Raj/Seamans): AIOE — occupation×industry×geography exposure dataset

The authors create and validate a new measure of an occupation's exposure to AI that they call the AI Occupational Exposure (AIOE). They use the AIOE to construct a measure of AI exposure at the industry level (AIIE) and a measure of AI exposure at the county level (AIGE).

AEA (Brynjolfsson/Mitchell/Rock): SML — most occupations partially automatable

We apply the rubric evaluating task potential for ML… to build measures of 'Suitability for Machine Learning' (SML) and apply it to 18,156 tasks in O*NET. We find that (i) ML affects different occupations than earlier automation waves; (ii) most occupations include at least some SML tasks; (iii) few occupations are fully automatable using ML; and (iv) realizing the potential of ML usually requires redesign of job task content.

Oxford (Frey & Osborne): 47% of US jobs at risk of computerisation

We examine how susceptible jobs are to computerisation… implementing a novel methodology to estimate the probability of computerisation for 702 detailed occupations, using a Gaussian process classifier… According to our estimates, about 47 percent of total US employment is at risk.