Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problemsSome links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
AI is not yet producing a broad, measurable unemployment shock in the United States, but its effects are beginning to appear unevenly. Anthropic’s March 5, 2026 report finds that actual work-related Claude use covers only a fraction of the tasks that language models could theoretically perform. Exposure is concentrated in occupations built around digital, text, and information-processing work, while many hands-on jobs show no observed coverage. The report also finds similar unemployment trends for highly exposed and unexposed workers since late 2022, alongside a tentative decline in young workers’ entry into exposed occupations. Its new measure, called observed exposure, is designed to distinguish what AI could do from what workers and employers are actually doing with it.
What is observed exposure to AI?
Observed exposure measures the share of an occupation’s task time that is both theoretically feasible for a large language model and actually being performed with work-related Claude use. It is intended to answer a practical question that capability scores alone cannot: where is AI already being used on the job, rather than where might it be useful in principle?
As an Amazon Associate I earn from qualifying purchases.
Recommended Free Tools
Maxim Massenkoff and Peter McCrory introduced the measure in Labor market impacts of AI: A new measure and early evidence, published March 5, 2026. The index combines five ideas:
- Model capability: whether an LLM could complete or accelerate a task.
- Observed use: whether the task appears often enough in work-related Claude traffic to count as covered.
- Work context: consumer or casual use is excluded from the work-use calculation.
- Automation versus augmentation: fully automated implementations receive more weight than uses where a person remains in control.
- Task importance: tasks are averaged using the share of an occupation’s working time spent on them.
A high score therefore does not mean that every worker in an occupation will be replaced. It means a substantial portion of the occupation’s task time is already being handled, at least partly, through an AI workflow represented in the usage data.
#1 Best Overall
How is AI exposure calculated?
The calculation joins three datasets and applies explicit weighting choices. The result is an occupation-level estimate, not a prediction of any individual’s probability of losing a job.
- Map occupations to tasks. The authors use task descriptions for roughly 800 US occupations from O*NET, the federal occupational information system.
- Measure real usage. Anthropic’s Economic Index identifies work-related Claude activity and matches the observed tasks to O*NET descriptions. A task counts as covered only when usage passes the report’s sufficiency threshold.
- Apply theoretical exposure scores. The study uses Eloundou and colleagues’ 2023 beta score. A beta of 1 means an LLM alone could at least double task speed; 0.5 means tools beyond the model are needed for that level of improvement; 0 means the task is not theoretically exposed.
- Weight the type of assistance. A fully automated implementation receives full weight. An augmentative workflow, in which a worker directs, checks, or completes the work, receives half weight.
- Average by time spent. Task-level coverage is aggregated according to the portion of an occupation’s time devoted to each task, producing the occupation’s observed-coverage percentage.
These choices make the measure more conservative than a capability-only index. A task can be technically possible for an LLM and still contribute nothing to observed exposure if it is not appearing in qualifying work use.
How different are theoretical capability and real-world use?
The gap is large. In Anthropic’s 2026 data, 97% of observed Claude tasks fell into theoretical exposure categories with beta values of 0.5 or 1, yet actual observed coverage in Computer and Math occupations was about 33%, compared with 94% theoretical capability.
This contrast matters because a capability score describes a ceiling, not adoption. It can include tasks that are blocked by reliability concerns, privacy rules, missing system integrations, approval requirements, or a lack of time to redesign a workflow. Observed exposure asks whether those barriers have been overcome enough for a task to appear in real work patterns.
The two measures answer different questions:
| Measure | What it asks | What it cannot establish |
|---|---|---|
| Theoretical exposure | Could an LLM perform the task at the specified speed-improvement threshold? | Whether an employer has adopted it or whether workers’ jobs have changed. |
| Observed exposure | Is qualifying, work-related AI use currently covering a meaningful share of task time? | Whether the use will persist, raise productivity, reduce headcount, or improve job quality. |
Which occupations have the highest observed AI exposure?
Digital information work leads the listed occupations, but the percentages describe task coverage, not the percentage of jobs eliminated.
| Occupation or group | Observed coverage | How to read the figure |
|---|---|---|
| Computer programmers | 75% | Three-quarters of estimated task time is covered by qualifying observed use under the report’s weighting method. |
| Customer service representatives | 67% | Many communication and information tasks are represented in work-related use. |
| Data-entry keyers | 67% | Structured text and information handling create substantial potential for coverage. |
| Computer and Math occupations | About 33% | The group-level observed figure is far below its 94% theoretical capability estimate. |
| Cooks, motorcycle mechanics, lifeguards, bartenders, dishwashers, and dressing-room attendants | 0% | No observed coverage was recorded for these occupations in the report’s Claude-usage measure. |
Zero observed coverage is a statement about the dataset and period, not a guarantee that a job is permanently insulated. A task may be absent because it is difficult to perform through a text interface, because employers use other systems, or because adoption has not yet reached the threshold used by the study.
Free tools Windows power users keep installed
One-click scans. No signup required.
Rank #2
Does exposure mean that AI is automating jobs?
No. The index deliberately separates automation from augmentation. Automation means the AI system carries out the implementation with little continuing human involvement; augmentation means the worker uses AI to assist, draft, search, summarize, or check work while retaining a meaningful role.
Full automation receives full weight in the coverage calculation, while augmentative use receives half weight. This prevents an occupation from receiving the same exposure score when AI is independently completing tasks as when it is functioning as a tool under close human direction.
The distinction still does not reveal the eventual employment effect. Augmentation can increase output without reducing staff, allow a smaller team to handle more demand, change the mix of skills employers seek, or make difficult work less costly. Automation can have similar outcomes, depending on whether demand expands, prices fall, or firms redeploy workers.
Is AI already causing unemployment?
The report does not find a statistically clear, broad increase in unemployment among workers in the most exposed occupations since late 2022. Its authors match occupation-level exposure to Current Population Survey respondents and find that unemployment trends for highly exposed and unexposed groups are largely similar over that period.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →That result should be read as an early observational finding, not proof that AI has had no labor-market effect. Occupation exposure is measured at an aggregate level, and unemployment can miss reduced hours, slower wage growth, changed duties, lower-quality new jobs, or people who leave the labor force. The comparison also cannot cleanly separate AI from business-cycle conditions, trade policy, interest rates, and other forces affecting hiring.
In practical terms, the evidence supports a narrower conclusion: through the period studied, there is no clear statistical signal of an economy-wide unemployment jump concentrated in the most AI-exposed occupations.
Are young workers being shut out of AI-exposed jobs?
The strongest warning in the report concerns entry into exposed occupations for people ages 22 to 25. Their rate of finding work in an exposed occupation fell by about 0.5 percentage points after ChatGPT’s arrival; the authors estimate that as a 14% decline relative to the 2022 level, with statistical significance only barely crossing conventional thresholds.
This pattern differs from a simple story in which experienced workers are immediately displaced. Younger applicants are disproportionately represented in hiring pipelines, internships, junior programming, support roles, and other positions where employers may decide that AI-assisted productivity reduces the need for entry-level recruitment.
The estimate has several possible interpretations. Young workers may remain in their existing jobs rather than switch into exposed occupations, move into less-exposed work, delay entry while returning to school, or be affected by ordinary sampling noise in survey transitions. The measure tracks movement into occupations, so it should not be treated as a count of young people permanently unemployed.
What does the relationship with BLS growth projections show?
Across occupations, a 10-percentage-point increase in observed coverage is associated with a 0.6-percentage-point lower employment-growth projection for 2024–2034 in an Anthropic analysis of Bureau of Labor Statistics projections.
This is an association between two occupation-level measures, not a causal estimate that AI will remove 0.6% of jobs. BLS projections incorporate expected demand, demographics, industry structure, and other economic assumptions. Highly exposed occupations may also be concentrated in sectors facing unrelated restructuring, making it impossible to assign the projected difference to AI alone.
The comparison is nevertheless useful as a screening signal: occupations with more observed AI coverage tend to have weaker projected growth, so analysts should examine both technology adoption and the underlying demand outlook rather than relying on either measure by itself.
What are the measure’s main limitations?
The authors describe observed exposure as an early, US-centered framework with several judgment calls and sources of uncertainty:
- Usage thresholds: The cutoff for sufficient work-related use determines whether a task is counted as covered.
- Similarity matching: Claude conversations must be matched to O*NET task descriptions, and close wording does not always mean identical work.
- Beta scores: The theoretical categories inherit assumptions about how much an LLM can speed a task and when additional tools are required.
- Automation weights: Giving augmentation half the weight of automation is analytically useful but remains a modeling choice.
- Representativeness: Claude traffic is not a census of every AI system, employer, country, or occupation.
- Confounding: Business cycles, trade policy, and sector-specific shocks can affect employment at the same time as AI adoption.
- Survey transitions: Current Population Survey responses about movement between occupations can contain reporting and sampling error.
- Moving baseline: Observed use can rise as model capabilities, integrations, employer policies, and worker habits change.
These limitations mean that the percentages are best used to compare occupations consistently at a point in time, not to forecast an individual worker’s outcome.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How should workers and employers use the findings?
The measure is most useful as a prompt for task-level planning. Workers should examine which parts of their job are routine, digital, and easy to verify, then build complementary skills in problem framing, domain judgment, communication, process ownership, and quality control. Those capabilities can remain valuable even when drafting or information retrieval is automated.
Employers can use the same decomposition without assuming that an occupation is a single block of work:
- List the occupation’s recurring tasks and estimate time spent on each.
- Separate tasks that can be safely assisted from tasks requiring human judgment, physical presence, confidentiality, or legal accountability.
- Pilot AI on low-risk tasks with measurable checks for accuracy, privacy, and turnaround time.
- Track whether productivity gains increase output, improve service, or reduce staffing needs; do not assume one outcome in advance.
- Provide training and redesign entry-level pathways if automation removes the tasks through which new workers traditionally learn.
This approach follows the report’s central insight: exposure belongs to tasks, while employment outcomes emerge from how organizations combine those tasks with demand and human skills.
What should readers watch in future updates?
The authors plan to repeat the method as new usage data arrive and extend it to other countries. Future releases should be judged by whether the same occupation’s observed coverage rises, whether the automation and augmentation mix changes, and whether employment transitions move in the same direction.
Best Value
A stronger labor-market conclusion would require several signals to align: sustained adoption across more than one AI provider, changes in hiring and hours rather than only task use, effects that survive business-cycle controls, and evidence from countries with different institutions. Until then, observed exposure is best understood as an adoption-sensitive early-warning measure rather than a job-loss forecast.
What is the bottom-line reading of the 2026 evidence?
AI’s practical reach is currently much smaller than its theoretical capability, and the labor-market effects are uneven. High exposure identifies occupations where workflows are changing fastest, not occupations where mass unemployment has already occurred. The clearest early concern is a possible weakening of young workers’ entry into exposed roles, while the broader unemployment signal remains inconclusive.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Frequently Asked Questions
Can observed exposure predict whether my company will lay off staff?
No. The measure is calculated for occupations across a large US workforce, not for individual employers. A company may automate more, less, or differently than the occupation-wide pattern because of its software, customer demand, regulations, and management choices.
Why might two people with the same job title face different AI risks?
Job titles hide differences in industry, seniority, client contact, and responsibility for final decisions. One worker may spend most of the day on repeatable digital tasks, while another with the same title handles exceptions, negotiations, or physical work that is not represented in the observed-use score.
Does zero observed coverage mean an occupation is safe from AI?
No. It means the study did not detect qualifying Claude use for the occupation during the measured period. Other AI products, future multimodal systems, or changes in employer practice could produce exposure that this particular index does not capture.
How can policymakers tell whether entry-level hiring is being displaced?
They should track vacancy postings, internship and apprenticeship volumes, occupational transitions, hours, wages, and school enrollment together. A fall in transitions alone cannot distinguish AI substitution from young people staying in existing jobs, returning to education, or responding to a weak hiring market.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
