In 2026, researchers made AI exposure easier to measure, but not individual job outcomes easier to predict. Capability maps, observed-use indicators, employer signals and labor studies describe different things. Use an occupation score to identify tasks to inspect, then look for repeated workflow changes and relevant local demand before making a costly move. Exposure is not adoption, displacement or a personal job-loss probability.
What is the 2026 answer in one minute?
In 2026, researchers improved the map of work exposed to AI; they did not make an individual job outcome predictable. The International Labour Organization’s “Workers’ exposure to AI: What indicators tell us – and what they don’t” says exposure measures estimate whether AI could affect tasks, but do not establish whether adoption is profitable or how employment and wages respond. The U.S. Bureau of Labor Statistics’ “Artificial Intelligence (AI) exposure categories” combines theoretical capability measures with observed-use indicators to compare occupations; it says these categories do not measure job loss, productivity gains, automation probability, or wage effects. Ask what a score measures: capability fit, observed use, or a labor-market outcome. Use an occupation indicator to choose tasks in your work to examine, not as evidence that your employer has adopted AI or that your role will disappear. The practical recommendation is an inference from these limits: begin with evidence gathering and a reversible test or learning step. Consider a larger move only when repeated workflow changes and relevant local demand evidence point in the same direction, and the option fits your salary, time, and constraints. Better evidence should precede a costly decision.
Sources: Workers’ exposure to AI: What indicators tell us – and what they don’t; Artificial Intelligence (AI) exposure categories
What changed in exposure research during 2026?
The clearest 2026 change is greater clarity about what exposure measures can answer. The International Labour Organization’s “Workers’ exposure to AI: What indicators tell us – and what they don’t” compares the choices behind exposure indicators and stresses that they do not establish what employers will adopt or what will happen to employment, wages, or demand. The OECD’s “The OECD AI Exposure Measure” maps AI capability indicators to occupational requirements across nine domains. The U.S. Bureau of Labor Statistics’ “AI exposure categories” combines theoretical capability measures with evidence of observed use to form relative occupational groups. Together they provide wider lenses, not a personal forecast. The ILO brief makes a recurring weakness of rankings visible: they can look like measurements of one settled property even though their inputs and assumptions differ. Indicators use different task lists, exposure judgments, and aggregation methods. Occupations also contain varied work: two people with the same title may handle different tasks or checking responsibilities. An occupation-level label compresses variation relevant to an individual decision. The ILO summary also points to constraints beyond technical possibility: profitability, institutions, and work reorganization affect whether a capability becomes automation in practice. Its contribution is a caution about interpretation, not a replacement score that resolves uncertainty. The OECD asks a capability-to-requirement question. It maps indicators across nine cognitive, social, and physical domains against occupational requirements through a capability-gap index. In the accessible abstract, smaller gaps indicate that current AI systems are closer to the capability profile associated with an occupation. The abstract identifies routine information processing, administrative work, and codifiable tasks as areas with smaller gaps, while contextual judgment, interpersonal understanding, complex decisions, and responsibility are among areas with larger gaps. This describes relative fit; it does not mean every task in a job can be automated or that a worker will be displaced. The OECD frames its measure over five to ten years. This is an analytical horizon, not a deployment schedule or job-loss forecast. Adoption depends on organizational choices, regulation, and social decisions as well as capability. The useful career question is narrower: which recurring tasks in my role resemble the mapped capabilities, and what conditions affect their suitability for change? The abstract cannot supply my workflow, employer policy, or local demand. For U.S. occupations, BLS combines three theoretical inputs and two observed-use inputs, treating those families separately before grouping occupations by relative scores. BLS says the scores are normalized as percentile ranks within each source and that the categories are relative groupings. They locate occupations within the covered comparison; neither a high category nor a low one predicts a particular job outcome. BLS explicitly warns that categories are not forecasts of employment, wages, adoption, or replacement. BLS documents complications that a headline category can hide: inputs must be crosswalked to occupational classifications, and some information is imputed. The agency notes that theoretical inputs conceptualize capabilities no later than mid-2023, while observed inputs may reflect early adopters or particular platforms. The categories also do not distinguish whether AI assists a worker or automates a task. Check these limits before interpreting a changed rank: sources and dates, occupations covered, and whether the score describes capability, observed use, or both. The ILO’s 2025 “Generative AI and Jobs: A 2025 Update” provides task-level baseline context. But a 2026 indicator should not be read as a directly comparable update unless measures, task definitions, occupation coverage, and data windows align. A category may change with capabilities, observed use, revised inputs, or aggregation. These publications are complementary, not a validated time series for measuring personal risk. For a career decision, ask what the indicator measures and which occupations it covers. Use it to select tasks for inspection, then compare them with your workflow and evidence about employer use or local demand. Broader measurement and clearer vocabulary help workers ask precise questions; they do not establish whether an employer will reorganize work or an individual will lose a job.
Sources: Workers’ exposure to AI: What indicators tell us – and what they don’t; The OECD AI Exposure Measure; Artificial Intelligence (AI) exposure categories; Generative AI and Jobs: A 2025 Update
How are the measures different, and why can they disagree?
A theoretical exposure score estimates whether AI capabilities could assist or complete tasks associated with an occupation. An observed-use measure asks where a particular system has actually been used in a selected sample. A capability-gap measure compares systems’ capabilities with occupational requirements. Labor-outcome measures look at hiring, employment, wages or job postings. They answer related but distinct questions, so disagreement is often a clue about scope and method rather than evidence that one number is wrong. The U.S. Bureau of Labor Statistics (BLS), in “Artificial Intelligence (AI) exposure categories,” makes this distinction visible by combining five external sources: three theoretical measures and two observed-use measures. The observed measures draw on Claude and Copilot activity mapped to occupational tasks or work activities. BLS first converts each input’s occupation scores into percentile ranks within that source’s covered occupations. It then calculates two medians for each occupation: one across the three theoretical sources and another across the two observed current-evidence sources. A clustering algorithm groups occupations using those dimensions into four relative categories: low, moderate, high and very high. The percentile step matters when comparing occupations. A percentile rank tells where an occupation’s score sits relative to other covered occupations under that specific source. BLS says normalization puts source results on a common zero-to-one scale, but the meaning changes from raw score to relative position. The comparison depends on the occupation universe and source: a rank can change because the comparison set or method changes, even if a worker’s daily tasks do not. Coverage and mapping add further uncertainty. BLS maps source data into its occupational classification system; some inputs do not cover every occupation. Across 4,155 possible occupation-source combinations, it reports 3,944 observed and 211 imputed, affecting 75 occupations. Its categories also do not distinguish automation from augmentation. The published limitations say theoretical inputs describe capabilities no later than mid-2023, while observed-use data may lean toward early adopters or particular platforms. Anthropic’s “Labor market impacts of AI: A new measure and early evidence” shows why observed use can tell a different story from theoretical possibility. Its March 2026 measure combines task ratings of theoretical feasibility with first-party Claude usage data, including work-related interactions. It gives more weight to automated patterns than augmentative ones, then aggregates task coverage to occupations using estimated task time. In its own data, the analysis reports that Claude use covered 33% of tasks in the Computer and Mathematics occupational category, while theoretical estimates indicated a much broader feasible scope. Those figures describe one provider’s platform and its measurement choices, not all AI use or an employer-wide implementation rate. The OECD’s “The OECD AI exposure measure” takes a different route. Its May 2026 working paper maps AI capabilities across nine cognitive, social and physical domains to occupational requirements, then constructs an AI Capability Gap index. Lower gaps mean current systems are closer to the capability profile required by an occupation, which the paper interprets as higher potential exposure. The abstract frames the measure as forward-looking over five to ten years and notes that actual effects depend on adoption, regulation, organizational change and social choices. This is a capability-to-requirement map, not a record of observed workplace use or a schedule for job change. These measures can disagree because they vary in task definitions, occupation coverage, AI systems, observation periods, data sources and aggregation. A capability map may flag work that a specific product has rarely been observed doing. A BLS relative rank compares occupations inside a defined source and universe; an OECD capability gap compares a different capability profile to requirements. Neither can be averaged responsibly with an employment or hiring result, which measures another link in the chain. A category is a summary of selected inputs, not a direct measure of how much work an employer has automated. A worker should look at actual recurring tasks, including review and handoffs, instead of treating an occupational average as a description of every role bearing that title. Check the question each indicator answers, its population and date, and its known omissions before interpreting a change. Compare like with like; do not turn unlike measures into a synthetic personal risk score.
Sources: Artificial Intelligence (AI) exposure categories; Labor market impacts of AI: A new measure and early evidence; The OECD AI Exposure Measure
What do 2026 labor studies show so far?
The 2026 labor evidence gives workers reason to watch early-career hiring and task allocation, but it does not produce one verdict about jobs. Studies measure unemployment among U.S. workers, payroll employment for younger workers, graduate outcomes, and junior-versus-senior shares at firms across countries. These findings can coexist. None estimates the chance that a particular reader will lose a job. In “Labor market impacts of AI: A new measure and early evidence,” Anthropic combines theoretical task feasibility with work-related Claude use observed on its platform, weighting automated use more heavily than assistance. It matches exposure to U.S. Current Population Survey data and compares highly exposed workers with those having no measured exposure. The authors report no statistically distinguishable post-ChatGPT rise in the unemployment gap. That weighs against a large relative unemployment increase in this comparison and period; it does not show that no tasks or hiring practices changed. Survey unemployment can miss people who never enter the labor force or whose transitions are misreported, and platform activity is not a census of workplace use. The study also examines job-finding among workers ages 22–25. It reports that the monthly rate of starting work in the most exposed occupations fell about half a percentage point relative to the no-exposure group, and estimates a 14% post-ChatGPT decline against its 2022 baseline. The result is barely statistically significant. Young people not hired into those occupations may remain in current jobs, take different work, or return to school; survey transition measurement is another limitation. This is a tentative hiring signal, not a displacement rate. Stanford’s revised August 2026 “Canaries in the Coal Mine?” uses ADP payroll records covering millions of U.S. workers through June 2026. It reports no widespread economy-wide displacement, but estimates employment for 22–25-year-olds in exposed occupations was 19% below the level it would have reached had it kept pace with less-exposed peers. The divergence operates mainly through slower hiring; experienced workers show no comparable gap. This is a relative comparison, not a finding that one in five young workers lost a job. The authors call the results early descriptive indicators, not causal estimates. They report attenuation after education controls, some divergence predating generative AI, and a stronger result in ADP than national survey benchmarks. These limits matter for applying it to a particular occupation or region. The U.S. Census Bureau working paper “Graduating into Disruption” examines administrative records for college graduates by major. Its September 2026 abstract reports that the most AI-exposed decile of majors experienced a regression-adjusted five-percentage-point decline in initial employment likelihood and a 13% decline in full-quarter initial earnings after ChatGPT’s introduction. The earnings difference reflects lower earnings within employing sectors and movement toward lower-wage sector mixes; effects attenuate as graduates move farther from labor-market entry. This concerns U.S. graduates entering work, not established employees or every graduate in an exposed field. It is a working paper, and timing after ChatGPT does not alone establish AI as the cause. A September 2026 Stanford working paper, “How Does AI Change Labor Demand? Evidence from 41 Countries,” shifts to firm-level workforce composition. It studies 1.25 billion job postings and 154 million employment records across 41 countries, inferring adoption from advertisements involving generative AI. Its abstract reports a relative decline in junior workforce share at adopting firms, primarily because senior employment grew, with suggestive evidence of modest overall employment growth. Senior employment shifted toward AI-exposed occupations, while estimates suggest juniors shifted away. This is not a forecast for each junior worker. Adoption is inferred from job ads; a cross-country result does not mean every employer follows the pattern. Taken together, the studies need not conflict. Unemployment can show no measurable increase while firms hire fewer entrants into certain exposed roles; graduate outcomes can weaken at entry while overall employment grows through senior hiring. The studies use different populations, periods, outcomes and exposure definitions. Their shared practical signal is narrower than “AI is eliminating jobs”: early-career entry and task division between junior and experienced workers warrant observation. The evidence does not establish that AI alone caused these patterns or that an established worker’s role is about to disappear. New entrants can inspect which tasks build experience and whether employers still hire for them. Established workers can check changing responsibilities and local opportunities before a costly pivot.
Sources: Labor market impacts of AI: A new measure and early evidence; Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence; Graduating into Disruption; How Does AI Change Labor Demand? Evidence from 41 Countries
How should a worker interpret exposure alongside employer demand?
Treat task fit and employer demand as separate evidence. An exposure map asks where AI capabilities may intersect with occupational tasks; hiring data asks what employers advertise. Neither establishes implementation, layoffs, or a durable local opportunity. The International Labour Organization's “Workers’ exposure to AI: What indicators tell us – and what they don’t” describes exposure as possible technological susceptibility, not a forecast of employment, wages, or demand. A divergence can be informative: technology may fit tasks employers are not hiring to change, or employers may request AI skills in occupations a task map labels less exposed. The New York Fed's May 2026 “Do Job Postings Show Early Labor-Market Effects of AI?” illustrates how to read the signals together. It combines a task-level exposure measure with U.S. Lightcast online postings, comparing employment and vacancies across exposure bands and tracking posting changes around ChatGPT's late-2022 release. Under its chosen measure, less than 10 percent of workers and vacancies were in occupations scoring at least 0.4; 40 percent of workers were in occupations with zero measured exposure. In its event study, vacancies in more exposed occupations were already declining relative to less exposed ones before late 2022; it found little evidence of a distinct additional AI-related decline after release. Junior and senior postings within highly exposed occupations showed no clear divergence. The Fed authors note a single less-exposed task can constrain automation of an occupation, and firms in the New York Fed's Second District reported retraining more often than reducing hiring in exposed occupations. The evidence is U.S.-based; postings are not a full census of vacancies or internal job changes. This shows no distinct hiring break attributable to AI in these postings during the study window, not what one employer has implemented. The study also notes that a single less-exposed task can limit automation of a whole occupation, and that firms in the New York Fed's Second District reported retraining more often than reducing hiring in exposed occupations. These are regional survey responses, not a universal employer pattern. A different signal appears in the Federal Reserve Board's September 30, 2026 FEDS Note, “AI on the Factory Floor: Evidence from Manufacturing Job Postings.” Task-based measures often classify manufacturing as relatively less exposed because production work includes physical activity. Yet Lightcast ads showed AI-related skill requirements reaching 11 percent of manufacturing postings in the second half of 2025, compared with 8 percent economy-wide, using data through July 2026. This mismatch is not evidence that factories automated particular jobs or production workers generally need machine-learning training. The measures answer different questions. Task maps identify where capabilities may matter; postings record requested skills among advertised openings. The Federal Reserve note says ads reflect employers' stated requirements, not skills of people hired, and may miss other hiring channels or internal promotions. Hiring requests can signal changing skill language without proving broad adoption, realized pay gains, or displacement. For a career decision, use divergence as a prompt to investigate your own work. Identify which tasks are changing and whether AI is being tried, approved, or built into ordinary workflow. Then inspect several recent postings in your region or for remote roles you could realistically pursue. Look for the duties around repeated AI requirements: is AI a preferred tool, a core production skill, or one element alongside domain judgment and responsibility? Check whether roles meet your salary floor and fit your credentials, location, schedule, and care obligations. One unusual ad is weak evidence; a recurring requirement across relevant openings is a better reason to investigate a skill gap, . National series can orient a search, but cannot price local vacancies or reveal one employer's hiring standards. When signals diverge, compare the task bundle, employer practices, and repeated demand in the market you can access. That can justify a bounded learning or workflow experiment; a costly field change needs stronger evidence that target work is reachable and compatible with your constraints.
Sources: Do Job Postings Show Early Labor-Market Effects of AI?; AI on the Factory Floor: Evidence from Manufacturing Job Postings; Workers’ exposure to AI: What indicators tell us – and what they don’t
Which career move fits the evidence and your constraints?
When the evidence is only an occupation-level exposure indicator, begin with a low-cost review of your own tasks. Move toward a larger change only as the evidence becomes more specific: repeated changes in the work itself, a plausible role that uses your existing strengths, and signs of demand in the market where you can actually work. That sequence is an editorial decision rule, not a validated prediction. The International Labour Organization's “Workers’ exposure to AI: What indicators tell us – and what they don’t” emphasizes that exposure measures describe technological susceptibility, not whether adoption is feasible or what happens to employment, wages, workflows, or demand. Consider an example: a recurring report-summary task contains several steps. A tool may help assemble a first draft from permitted material. But source checking, handling exceptions, protecting confidential information, and accepting responsibility for what is sent remain part of the workflow. If review takes as long as drafting, or an error carries consequences that require extensive checking, the apparent time saving may not translate into useful capacity. If the draft is reliable enough for a qualified person to review, the task may change without the whole role disappearing. The relevant question is what happens to the full sequence, including the work around the draft, not whether one step looks automatable. The first option is a bounded workflow test. It is the most reversible and asks for the least evidence: identify one recurring, low-risk task; check workplace rules and data restrictions; then, if permitted, compare a tool-assisted result with the usual process. Track corrections, omissions, review time, and who remains accountable. Stop if quality drops or checking erases the benefit. An in-role upgrade is a reasonable next comparison when the task mix is changing but your domain judgment and responsibility still matter. The OECD’s “The OECD AI exposure measure” maps capability indicators across nine cognitive, social, and physical domains to occupational requirements; its abstract finds smaller capability gaps around routine information processing and codifiable work, and larger gaps where contextual judgment, interpersonal understanding, complex decisions, and responsibility matter. It does, however, support inspecting where your role requires context and accountability before assuming that faster drafting makes your experience irrelevant. An adjacent move deserves attention when your current tasks overlap with neighboring roles and local evidence repeatedly points to relevant requirements. Compare actual postings across several employers and note recurring duties, credentials, tools, and levels of responsibility. The Federal Reserve Board note “AI on the Factory Floor: Evidence from Manufacturing Job Postings” offers a useful example of why task exposure and demand can diverge: its U.S. Lightcast analysis found AI-related skill requirements rose in manufacturing postings, including a reported 11 percent of manufacturing postings versus 8 percent economy-wide by July 2026. The note treats online advertisements as a timely hiring-demand indicator, not a direct count of openings; postings show stated requirements, may miss offline hiring and internal promotion, and do not establish that a worker with a new certificate will be hired. A larger career change has the highest cost and lowest reversibility. Consider it when changes to your task bundle persist, an adjacent option does not preserve enough of your experience or income, and evidence of relevant demand is stronger than a general exposure score. Before committing, test the plan against your salary floor, location, credentials, training time, tuition and other costs, health, and family responsibilities. If the target requires deep technical preparation, a structured degree may provide depth and feedback; if the goal is to use AI better in an existing field, a bounded project or focused learning may answer the need with less disruption. Durable skills such as problem framing, data literacy, evaluation, domain knowledge, and verification can travel across tools. Vendor interfaces change faster. A course completion alone does not establish workplace capability or hiring value. The practical ranking is therefore: inspect and, where allowed, test one task; compare an in-role upgrade; investigate adjacent work when local demand and task overlap support it; reserve a larger transition for persistent evidence and a feasible plan.
Sources: Workers’ exposure to AI: What indicators tell us – and what they don’t; The OECD AI Exposure Measure; AI on the Factory Floor: Evidence from Manufacturing Job Postings
What evidence can you gather before committing to retraining?
Before paying for extensive retraining, gather two kinds of evidence: a record of what changes in one recurring task, and signs of what nearby employers request. The Bureau of Labor Statistics says its occupation categories are relative comparisons, not forecasts of job loss, adoption, wages or productivity. Ask whether a tool changes work you do, and whether a skill you might learn is relevant to roles you could realistically pursue. Choose one recurring task, such as preparing a report summary. Record its steps from source material to approved result: inputs, judgment calls, exceptions, quality checks, reviewer and approximate time. Keep the current method as a baseline. Without it, a faster first draft can look like a gain even when correction and approval take longer. Check workplace rules before testing a tool, including requirements for approved systems, confidential information, customer data and human review. If the task is permitted, compare one low-risk output with the ordinary process. Log omissions, corrections, review time and who remains accountable. Stop or adjust if quality falls, sensitive information could be exposed, or review consumes the time saved. This is an editorial decision procedure, not a validated assessment. One trial can reveal friction; it cannot establish lasting productivity, employer-wide adoption or whether a role will shrink. Examine labor demand separately. The Federal Reserve Bank of New York’s “Do Job Postings Show Early Labor-Market Effects of AI?” analyzed U.S. Lightcast postings alongside an occupation exposure measure. The authors found little indication of a distinct AI-driven decline in postings for exposed occupations. The relative vacancy decline began before ChatGPT’s release, complicating a causal explanation. They caution that postings do not equal layoffs or completed hires. This national pattern is context, not a substitute for the market where you can work. Compare several recent ads for roles you could reach while respecting your salary floor and location. Note recurring duties, skills, credentials, seniority and place. Separate repeated requirements from one unusually specific listing, and required qualifications from preferred ones. The September 2026 Federal Reserve note “AI on the Factory Floor” shows why local demand can add information: manufacturing is often classified as relatively less exposed by task measures, yet AI-related skills rose in online postings. The note measures advertised requirements, not skills of workers hired; it may miss internal promotion and offline recruitment. It cannot show that a course secures a job. Use the combined record to choose learning. A specific gap might call for a small project or focused course with practice and feedback. A certificate is more relevant when target employers repeatedly request it; check cost, prerequisites and curriculum before enrolling. For work requiring deeper technical foundations, compare formal study with staged projects against actual role requirements. No route guarantees hiring. If evidence is thin, keep the step reversible: document the task, ask a manager what review and accountability would be required, then revisit the decision when local evidence is stronger.
Sources: Artificial Intelligence (AI) exposure categories; Do Job Postings Show Early Labor-Market Effects of AI?; AI on the Factory Floor: Evidence from Manufacturing Job Postings
What is the next step this month?
Choose one recurring task and one local demand signal to examine this month. Note where AI might assist and what still needs checking or approval. Then review current openings in your region for roles you could reach; note which skills or duties recur across listings. This compares your work with stated employer needs without treating either signal as a forecast. The International Labour Organization’s “Workers’ exposure to AI: What indicators tell us – and what they don’t” explains why: exposure measures describe possible task change, but do not establish whether adoption is profitable or how employment, wages, workflows, or demand will adjust. The 2026 verdict is narrow: measurement improved and labor studies add signals, but they do not settle your personal outcome. Reconsider a larger move when repeated task changes and relevant local demand point in the same direction, and the path fits your salary floor, finances, health, time, and family responsibilities. The free task checker at /ai-job-risk-checker can organize change-pressure signals; it does not estimate a validated probability of displacement. To compare a stay-and-redesign option with adjacent or larger changes, the paid career roadmap can structure scenarios. It does not guarantee employment or income. Document one task and check one demand signal before costly retraining.
Sources: Workers’ exposure to AI: What indicators tell us – and what they don’t
Sources and notes
- Workers’ exposure to AI: What indicators tell us – and what they don’t
Explains the limits of exposure indicators and why they do not establish adoption or labor-market outcomes.
- The OECD AI Exposure Measure
Maps AI capability indicators to occupational requirements across nine domains and describes capability gaps.
- Artificial Intelligence (AI) exposure categories
Describes BLS’s theoretical and observed-use inputs, relative occupational categories, and their limitations.
- Generative AI and Jobs: A 2025 Update
Provides task-level baseline context for interpreting later exposure indicators.
- Labor market impacts of AI: A new measure and early evidence
Reports a U.S. labor analysis combining theoretical task feasibility, first-party Claude use, and CPS outcomes.
- Do Job Postings Show Early Labor-Market Effects of AI?
Compares U.S. job postings with an exposure measure and discusses limits of postings as labor-demand evidence.
- Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence
Summarizes descriptive U.S. payroll evidence on employment patterns among younger workers in exposed occupations.
- Graduating into Disruption
The working-paper abstract reports initial employment and earnings differences by AI exposure among U.S. college-major cohorts.
- How Does AI Change Labor Demand? Evidence from 41 Countries
The abstract reports firm-level workforce composition findings using job advertisements and employment records across 41 countries.
- AI on the Factory Floor: Evidence from Manufacturing Job Postings
Uses U.S. manufacturing postings to illustrate that advertised AI skill requirements and task exposure are distinct signals.
Apply this to your own work
See the whole job market at once.
Explore which occupations AI may reshape, then turn the signal into a practical response.
Explore the job map