What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
AI is not yet producing a clear, economy-wide unemployment shock, but it is already changing which work is technically reachable and may be affecting who gets hired into exposed occupations. Anthropic researchers Maxim Massenkoff and Peter McCrory’s March 5, 2026 report measures observed exposure: the share of an occupation’s task time that shows sufficient, work-related use of Claude for automation or assistance. Actual use remains far below technical capability. Computer and Math occupations show about 33% observed coverage against 94% theoretical capability, while computer programmers lead the listed occupations at 75% coverage. Unemployment trends for highly exposed workers and unexposed workers have been broadly similar since late 2022. The early warning is among 22-to-25-year-olds, whose entry rate into exposed occupations is estimated to have fallen 14% after ChatGPT, a result the authors describe as barely statistically significant.
What does Anthropic’s new AI exposure measure actually measure?
Observed exposure is intended to show where AI-capable tasks are being used in real work, not merely where a language model could theoretically help. It combines five ideas: what an LLM can do, whether people are using it for work, whether the use automates or augments a task, and how much of an occupation’s working time consists of that task.
That distinction matters because a job can contain many theoretically automatable activities without employers or workers adopting automation. Conversely, a small amount of AI use in a high-time-share task can give an occupation meaningful observed exposure. The measure therefore sits between a laboratory capability score and a head-count measure of jobs eliminated.
“This report introduces a new measure for understanding the labor market effects of AI and studies impacts on unemployment and hiring.” — Maxim Massenkoff and Peter McCrory, Anthropic, 2026
How is observed exposure calculated?
The authors combine three main data sources and then apply explicit weighting choices.
- Map work into tasks. O*NET descriptions cover roughly 800 US occupations and list the tasks associated with each occupation, including the estimated time spent on those tasks.
- Measure real use. Anthropic Economic Index data identify Claude interactions that have sufficient work-related usage for a task to count as covered.
- Start with theoretical capability. The report uses exposure scores from Eloundou and co-authors’ 2023 framework. A β score of 1 means an LLM alone could at least double the speed of the task; β=0.5 means tools are needed to achieve that threshold; β=0 means the task is not exposed under the framework.
- Weight the type of use. A fully automated implementation receives full weight. An augmentative use, in which a person remains involved, receives half weight.
- Average by task time. Task-level coverage is averaged using the share of an occupation’s time spent on each task, producing an occupation-level observed-coverage percentage.
In simplified form, an occupation’s percentage is not “the percentage of workers replaced.” It is a time-weighted estimate of how much of the occupation’s task mix is showing qualifying AI use, after distinguishing automation from assistance.
How large is the gap between AI capability and actual workplace use?
The report finds a substantial adoption gap. Ninety-seven percent of observed Claude tasks fell into the β=0.5 or β=1 categories, meaning they were theoretically feasible under the capability framework. Yet observed coverage in Computer and Math occupations was about 33%, compared with 94% theoretical capability.
Recommended Free Tools
The two percentages answer different questions. The 94% figure asks whether the tasks could meet the theoretical speed-up threshold with an LLM, alone or with tools. The 33% figure asks how much qualifying work-related Claude use was actually observed after applying the automation and augmentation weights and the occupation’s task-time mix. A high capability score can therefore coexist with limited adoption because of workflow integration, data access, reliability requirements, cost, policy, or employee choice.
This gap also explains why capability headlines should not be read as near-term employment forecasts. Technical feasibility is a ceiling-like measure; observed exposure is an adoption-sensitive measure that can rise as organizations change processes and as models improve.
Which occupations have the highest and lowest observed coverage?
Among the occupations highlighted in the report, computer programming has the highest observed coverage. Customer service and data-entry work also rank high, while several hands-on or in-person service jobs show zero observed coverage in the measured Claude traffic.
Rank #2
| Occupation | Observed coverage | How to read the figure |
|---|---|---|
| Computer programmers | 75% | About three-quarters of task time, after the report’s weighting, had qualifying observed coverage. |
| Customer service representatives | 67% | A large share of the occupation’s measured task time showed work-related coverage. |
| Data-entry keyers | 67% | Routine information-handling tasks rank among the most covered listed work. |
| Cooks | 0% | No observed coverage in the report’s listed examples. |
| Motorcycle mechanics | 0% | No observed coverage in the report’s listed examples. |
| Lifeguards | 0% | No observed coverage in the report’s listed examples. |
| Bartenders | 0% | No observed coverage in the report’s listed examples. |
| Dishwashers | 0% | No observed coverage in the report’s listed examples. |
| Dressing-room attendants | 0% | No observed coverage in the report’s listed examples. |
About 30% of workers are in occupations with zero observed coverage. Zero coverage does not mean that technology can never affect those jobs; it means the measure found no qualifying Claude use for the tasks represented in those occupations at the time and under the report’s thresholds.
Is AI already causing unemployment?
The report does not find a systematic rise in unemployment for workers in highly exposed occupations since late 2022. The researchers matched occupation-level exposure to Current Population Survey respondents and compared unemployment trends for highly exposed and unexposed groups. Those trends are largely similar, with no statistically clear increase for the exposed group.
That is a narrower conclusion than “AI has had no labor-market effect.” The analysis tests unemployment patterns over an early period, and a person can be affected without becoming unemployed: hours, duties, pay, occupation, hiring prospects, or productivity may change first. The absence of a clear unemployment increase also does not prove that future adoption will be harmless.
Business-cycle conditions and trade policy can affect the same occupations at the same time. The report identifies those forces as potential confounders, so the observed relationship should not be treated as a clean estimate of layoffs caused by AI.
Are young workers being shut out of AI-exposed jobs?
The clearest adverse signal appears in entry into exposed occupations among people ages 22 to 25. Their entry rate falls by about 0.5 percentage points, corresponding to an estimated 14% post-ChatGPT reduction relative to 2022.
The estimate is tentative and barely statistically significant. A lower transition rate can reflect several behaviors besides permanent exclusion: young people may stay in jobs they already hold, switch into less-exposed occupations, postpone a move while searching, or return to school. Survey-based transition measures can also be noisy, particularly when occupation coding changes between interviews.
For that reason, the result is best read as an early warning about hiring pipelines rather than proof that an entire generation is being displaced. It suggests that entry-level pathways deserve separate monitoring even when aggregate unemployment remains stable.
Does greater AI exposure predict slower job growth?
Yes, in the report’s comparison with Bureau of Labor Statistics projections, but the relationship is an association rather than a causal forecast. A 10-percentage-point increase in observed coverage is associated with a 0.6-percentage-point lower projected employment-growth rate for 2024–2034.
For example, an occupation with 40% observed coverage would be compared with one at 30% coverage using a difference of 10 percentage points, not a 10% relative increase. The 0.6-point estimate describes the difference in the BLS projection relationship analyzed by Anthropic; it does not say that 0.6% of workers will lose jobs, nor that AI alone determines the projection.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallOutdated 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 matchBLS projections themselves incorporate many assumptions about demographics, demand, technology, and industry structure. Exposure may be correlated with other characteristics—such as routine digital workflows—that also influence projected growth.
Why does the automation-versus-augmentation distinction matter?
Automation and augmentation have different immediate labor implications, so the report does not count them equally. A fully automated implementation receives full weight, while an augmentative use receives half weight because a worker remains part of the process.
This choice prevents a high volume of drafting, search, or editing assistance from being interpreted as the same thing as a task performed without human involvement. It also means the percentage is sensitive to how an interaction is classified. Two occupations with similar amounts of Claude traffic could receive different observed-coverage values if one uses the system mainly for autonomous execution and the other mainly for support.
The weighting is a judgment call, not an independently observed law of economics. The authors identify automation-versus-augmentation weights, usage thresholds, and similarity matching as areas where reasonable methodological choices could change the number.
Free tools Windows power users keep installed
One-click scans. No signup required.
How should workers interpret an occupation’s coverage percentage?
Use the percentage as an indicator of where work processes are already engaging with AI, not as a personal probability of redundancy.
- It is occupation-level. A programmer’s individual duties may differ substantially from the average task mix used for the occupation.
- It is task-time weighted. A task that occupies more working time has more influence than an occasional task.
- It includes assistance. Half-weight augmentation contributes to coverage even when a human remains responsible.
- It is US-centered and time-specific. The result reflects the occupations, usage data, and thresholds studied, not every country or future model.
- It is not a wage or hiring measure. Coverage does not reveal whether employers reduced pay, increased output, added workers, or changed job quality.
For a career decision, the useful question is which recurring tasks are changing and which human responsibilities remain scarce, rather than whether the occupation has crossed a single danger percentage.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How can employers and policymakers use the measure responsibly?
Employers can use observed exposure to identify workflows that merit review, then check quality, security, accountability, and training requirements before changing staffing. A high score is a prompt to examine processes, not a justification for automatic head-count cuts. A zero score can indicate that current usage has not reached the occupation, while leaving open the possibility of later adoption.
Workforce planners can separate entry-level hiring from employment of established workers. The report’s young-worker result shows why a stable unemployment rate may conceal weaker access to first jobs in exposed fields. Tracking transitions, schooling, occupation changes, hours, and earnings together would provide a fuller picture than unemployment alone.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Researchers should compare theoretical capability, observed use, and employment outcomes as separate layers. Collapsing them into one “AI risk” number hides the adoption gap that is central to this report.
What are the report’s main limitations?
- Geography: The framework is built around US occupations and US labor-market data, so it should not be generalized automatically to other countries.
- Observational usage: Claude traffic shows where qualifying use occurred, not every AI system, internal deployment, or offline workflow.
- Threshold choices: The definition of sufficient work-related usage, similarity matching, β scores, and the automation/augmentation weights all affect coverage.
- Confounding: Business cycles, trade policy, industry demand, and other shocks may move employment in the same occupations.
- Survey error: Current Population Survey transition measures can misclassify occupation changes or miss complex paths through school and nonemployment.
- Changing technology: Observed use can expand as capabilities, prices, interfaces, and organizational adoption change.
Anthropic says it plans to reuse the method with new usage data and extend it to other countries. Such updates can show whether the capability-to-adoption gap is closing and whether the early hiring signal persists.
What is the most defensible conclusion in 2026?
AI has measurable exposure in specific occupations, but the evidence does not yet show broad AI-driven unemployment. The strongest current facts are the separation between theoretical capability and observed use, the concentration of coverage in occupations such as programming, customer service, and data entry, and the tentative decline in young workers’ entry into exposed occupations. Observed exposure is therefore best treated as an early labor-market indicator—useful for tracking where change is happening, insufficient on its own to predict who will lose a job.
Frequently Asked Questions
Can this exposure measure be applied to workers outside the United States?
Not directly with the evidence available here. The framework uses US O*NET occupations and US labor-market data; the authors say they intend to extend it to other countries, which will require comparable task and employment measures.
Does 75% coverage mean that 75% of computer programmers’ jobs will disappear?
No. The figure is a task-time-weighted measure of observed, work-related AI coverage, including half-weight augmentative use. It does not estimate layoffs, replacement rates, wages, or the number of jobs that employers will retain.
How certain is the reported 14% decline for young workers?
It is an estimate that is barely statistically significant, so the result is less precise than a robust trend. The authors also note alternative explanations such as staying in an existing job, switching occupations, returning to school, and noisy survey transitions.
How often will the observed-exposure numbers change?
They can change whenever usage patterns, model capabilities, organizational adoption, or the classification thresholds change. Anthropic says it plans to repeat the method with new usage data rather than treat the March 2026 estimates as permanent.
What evidence would show that exposure is causing job losses rather than merely correlating with them?
Researchers would need longer time series, better controls for business cycles and trade policy, and outcome measures beyond unemployment, such as hiring, hours, earnings, occupation changes, and firm-level staffing. The current report presents early observational evidence rather than a definitive causal estimate.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →
