Two numbers describe the same job and they do not agree. Not by a little. One number comes from a lab estimate of what a language model could theoretically handle under ideal conditions. The other comes from watching what Claude actually does inside real professional work, day after day, across paying customers. Every AI job-loss headline you have read rests on the first number. Almost none of them rest on the second, and the distance between the two is where your real career risk sits.
Three years after ChatGPT launched, Anthropic economists Maxim Massenkoff and Peter McCrory published that second measure on 5 March 2026, building it from the Anthropic Economic Index rather than from surveys or capability benchmarks. They filtered real Claude API (application programming interface) calls and product usage down to professional and work-related contexts, separated fully automated pipelines from human-assisted augmentation, weighted the automated share more heavily, then mapped every usage pattern onto O*NET task descriptions, the US government's standardized catalogue of what each job actually involves. They call the result observed exposure, and it measures what AI is doing in professional workflows right now rather than what it might do someday.
Pooya Golchian reproduced the paper's figures here as live charts and ran the complete 22-category dataset back through descriptive and inferential statistics. Start with the radar below, which reproduces Figure 2 of the paper directly. The outer ring plots theoretical capability by occupational category. The inner ring plots observed usage across those same 22 categories. The distance between the two rings carries the entire argument, and it does not close evenly across the economy. Four more charts follow it, covering the exposure gap category by category, the ten most exposed occupations, the demographics of the exposed workforce, and the Bureau of Labor Statistics employment growth model.
The short answer
Every headline figure from the study sits in one table. Each row below gets a full section further down explaining the method that produced it, the sample it rests on, and the caveats the authors attach to it.
| Finding | Data |
|---|---|
| Computer Programmers exposure | 75% of tasks observed in real Claude usage |
| Workers with zero AI exposure | ~30% of the workforce |
| Unemployment change since ChatGPT | No statistically significant increase |
| Young worker (22-25) hiring change | -14% job-finding rate in exposed roles |
| Bureau of Labor Statistics (BLS) impact per +10 percentage point (pp) increase in exposure | -0.6pp in 10-year employment growth projection |
| Wage premium for exposed workers | 47% higher average earnings |
| Graduate degree rate: high vs zero exposure | 17.4% vs 4.5% |
Why Every Previous AI Risk Number Was Probably Wrong
The dominant framework for measuring AI's threat to jobs was built by Eloundou et al. in 2023. Their method asked whether a large language model (LLM) could reduce the time needed for a given task by at least 50%. Applied to every task in every occupation in O*NET (the Occupational Information Network, a US government database of standardized job task descriptions), it produces alarming numbers. Computer and Math occupations reached 94% exposure. Office and Admin reached 90%.
Thousands of articles have cited these figures. They are also measuring the wrong thing.
They measure capability, meaning what models can do under ideal conditions with careful prompting. They do not measure deployment, meaning what AI is actually doing in professional workflows today.
Anthropic's researchers built a different measure using the Anthropic Economic Index, a dataset of real Claude API (application programming interface) calls and product usage. They filtered for professional and work-related contexts, separated automated pipelines from human-assisted augmentation (weighting the former more heavily), and mapped each usage pattern to O*NET task descriptions.
The result is observed exposure: a ground-truth measure of how deeply AI has actually penetrated real work, right now.
The Gap Between Theory and Reality
Across knowledge work the gap runs 34 to 65 percentage points. Computer and Math occupations showed 94% theoretical exposure against 33% observed, a 61 point drop. Office and Admin showed 90% theoretical against 25% observed, the widest gap in the dataset at 65 points. Physical and service work barely registers a gap at all, because neither measure climbs out of single digits there.
This gap exists for several real reasons:
- Integration cost. Connecting AI to existing systems, data pipelines, and approval workflows takes significant engineering work.
- Liability and compliance. Legal, finance, and healthcare tasks often require human sign-off regardless of AI capability.
- Trust and verification. Workers and managers require consistent, auditable outputs before delegating consequential decisions.
- Workflow inertia. Many organizations have not yet restructured roles around AI augmentation.
The gap does not mean AI will never close it. It means the gap exists today, and that theoretical forecasts dramatically overstate current disruption.
Theoretical vs Observed Exposure Across All 22 Occupation Categories
Below is the full dataset plotted in the radar chart above, ranked by theoretical capability. Theoretical values come from Eloundou et al. (2023). Observed values come from the Anthropic Economic Index as reported in the 2026 paper.
| Occupation Category | Theoretical (Eloundou et al.) | Observed (Anthropic 2026) | Gap |
|---|---|---|---|
| Computer & Math | 94% | 33% | -61pp |
| Office & Admin | 90% | 25% | -65pp |
| Business & Finance | 75% | 22% | -53pp |
| Legal | 68% | 15% | -53pp |
| Architecture & Engineering | 65% | 14% | -51pp |
| Life & Social Sciences | 55% | 8% | -47pp |
| Arts, Design & Media | 55% | 10% | -45pp |
| Sales & Related | 52% | 18% | -34pp |
| Management | 50% | 8% | -42pp |
| Education & Library | 48% | 8% | -40pp |
| Healthcare Practitioners | 45% | 8% | -37pp |
| Social Services | 28% | 4% | -24pp |
| Healthcare Support | 20% | 4% | -16pp |
| Production | 20% | 4% | -16pp |
| Protective Service | 18% | 4% | -14pp |
| Transportation | 18% | 4% | -14pp |
| Personal Care | 12% | 3% | -9pp |
| Installation & Repair | 12% | 3% | -9pp |
| Construction | 10% | 3% | -7pp |
| Food & Serving | 8% | 2% | -6pp |
| Agriculture | 8% | 2% | -6pp |
| Grounds Maintenance | 5% | 2% | -3pp |
Read down the Gap column and the shape of the problem appears. The widest gaps sit at the top, in language-heavy desk work. The narrowest sit at the bottom, in work that needs hands, presence, or physical judgment.
Statistical Analysis: What the Full 22-Category Dataset Shows
The radar chart above covers all 22 occupation categories in the paper. Running the full dataset through descriptive and inferential statistics produces findings that go beyond any individual data point.
Key Statistics Across All 22 Categories
| Statistic | Value | Plain-language meaning |
|---|---|---|
| Mean theoretical exposure | 38.9% | Average share of tasks an LLM could theoretically handle |
| Mean observed exposure | 9.3% | Average share of tasks AI is actually performing today |
| Overall deployment ratio | 24% | For every 4 tasks AI could do, only 1 is being done |
| Mean gap (theoretical minus observed) | 29.6 percentage points | The average "untapped" AI potential per occupation category |
| Standard deviation of gap | 19.9 pp | High variability: the gap is far larger in knowledge work than manual work |
| Pearson correlation (r) | 0.92 | Very strong positive relationship between theoretical and observed values |
| Coefficient of determination (R²) | 0.85 | Theoretical exposure explains 85% of the variance in observed exposure |
The 24% Deployment Ratio
Dividing mean observed (9.3%) by mean theoretical (38.9%) gives 0.239, or roughly one in four. Across every occupation category the paper measures, AI is deployed on approximately 24% of the tasks it is theoretically capable of handling.
This ratio is remarkably consistent. It appears whether you look at high-exposure knowledge work (Computer & Math: 33/94 = 35%) or lower-exposure professional roles (Legal: 15/68 = 22%). The uniformity implies a structural barrier, not a sector-specific one. The blockers are integration cost, compliance requirements, trust deficits, and organizational inertia, and they apply at approximately the same scale regardless of the occupation.
Pearson Correlation r = 0.92
A Pearson correlation of r = 0.92 between theoretical and observed values (p < 0.001) tells us something important: the ranking of occupations by AI capability is almost identical to their ranking by actual AI deployment. Industries where AI could do the most are also where it is being deployed the most.
However, the slope of that relationship is not 1.0. It is approximately 0.24. The linear model that best describes the relationship is:
Observed exposure ≈ 0.24 × Theoretical capability
This means the deployment gap is proportional, not additive. A doubling of theoretical capability corresponds to roughly a doubling of observed deployment, at 24 cents on the dollar. The gap does not close as theoretical capability increases: it scales up proportionally.
R² = 0.85 means theoretical exposure explains 85% of the variance in observed exposure. The remaining 15% is explained by sector-specific factors: regulatory density (healthcare), client accountability requirements (legal), or the specific mix of tasks within a role.
Bimodal Gap Distribution
The standard deviation of the gap (19.9 pp) is large relative to the mean (29.6 pp), giving a coefficient of variation of 67%. This reflects a bimodal distribution rather than a normal one.
Two clusters emerge clearly:
Cluster 1: Knowledge work (Management, Business & Finance, Computer & Math, Legal, Office & Admin). Gaps range from 34 to 65 pp. These sectors have high theoretical exposure because their tasks are language-based and reasoning-intensive, but deployment is constrained by compliance, verification requirements, and integration complexity.
Cluster 2: Physical and personal service work (Food & Serving, Grounds Maintenance, Agriculture, Construction, Installation & Repair). Gaps range from 3 to 9 pp. Both theoretical and observed exposure are near zero because the underlying tasks require physical presence, sensory judgment, or real-time human interaction.
The practical implication: the 24% deployment ratio will close first in knowledge work, not physical work, because the structural barriers are organizational rather than technological.
The BLS Linear Model
The BLS employment growth relationship the paper identifies can be expressed as:
Projected 10-year job growth ≈ 4.2% − (0.06 × observed_exposure_percentage)
At 0% observed exposure, projected growth is approximately 4.2% (physical and service roles). At 70% observed exposure (Computer Programmers), projected growth approaches zero. Every 10 percentage point increase in observed exposure reduces long-run employment growth by 0.6 percentage points.
This is a first-order linear approximation. The relationship likely becomes nonlinear at very high exposure levels, and the BLS projections pre-date the latest model generations. But as a working rule, the model provides a concrete, data-grounded way to translate observed exposure into a long-range career risk signal.
Which Jobs Are Most Exposed Right Now
The top ten occupations by observed exposure are entirely white-collar knowledge work roles. Physical, hands-on, and social roles have zero or near-zero exposure.
Top 10 Most AI-Exposed Occupations
| Rank | Occupation | Observed Exposure |
|---|---|---|
| 1 | Computer Programmers | 75% |
| 2 | Data Entry Keyers | 67% |
| 3 | Customer Service Representatives | ~65% |
| 4 | Financial Analysts | ~55% |
| 5 | Writers and Authors | ~48% |
| 6 | Insurance Underwriters | ~46% |
| 7 | Paralegals and Legal Assistants | ~42% |
| 8 | Software Quality Assurance (QA) Engineers | ~40% |
| 9 | Data Scientists | ~38% |
| 10 | Bookkeeping Clerks | ~36% |
Jobs with Zero or Near-Zero AI Exposure
By contrast, roughly 30% of all workers have zero observed AI exposure. These roles require physical presence, real-time sensory judgment, or human connection that AI does not currently replicate.
| Occupation | Why Exposure Is Zero |
|---|---|
| Cooks and Chefs | Physical dexterity, sensory judgment, real-time adjustment |
| Motorcycle Mechanics | Hands-on diagnosis, physical manipulation of components |
| Lifeguards | Real-time vigilance, physical rescue, outdoor environment |
| Bartenders | Social interaction, mixing, reading the room |
| HVAC (Heating, Ventilation, and Air Conditioning) Technicians | On-site diagnosis, physical installation and calibration |
| Electricians | Safety-critical physical labor, on-site code compliance |
| Dishwashers | Physical operational labor |
| Personal Trainers | Physical coaching, real-time human motivation |
Who Actually Works in AI-Exposed Jobs
The demographics of the most AI-exposed workforce reveal something counterintuitive. The workers most at risk are not the lowest paid or least educated. They are among the most skilled, best compensated, and most credentialed workers in the economy.
Worker Profile: High AI Exposure vs Zero Exposure
| Characteristic | Zero AI Exposure | High AI Exposure (top quartile) |
|---|---|---|
| % Female | ~42% | ~58% (+16pp) |
| Average Annual Wage | ~$53,000 | ~$78,000 (+47%) |
| % with Graduate Degree | 4.5% | 17.4% (nearly 4x) |
| % White or Asian | ~65% | ~78% |
This flips the popular narrative in two ways.
First, AI disruption risk is concentrated in higher-wage, higher-education work. The fear that AI would primarily eliminate low-wage service jobs is not supported by this data. Cooks, mechanics, and bartenders are largely unaffected. Financial analysts, programmers, and writers are most affected.
Second, the exposed workforce skews female. This matters for policy. Most discussions of automation risk assume male-dominated manufacturing and logistics roles are most threatened. The observed exposure data points in a different direction.
What Has Actually Happened to Employment
Massenkoff and McCrory matched their occupation-level exposure scores to individual respondents in the Current Population Survey, covering 2019 through 2025. They used a difference-in-differences framework comparing unemployment trends for the top quartile of exposed workers against the zero-exposure group.
The result: no detectable unemployment increase for highly exposed workers since November 2022.
The method is not blind. The researchers calibrated their approach using a difference-in-differences framework (a statistical technique that compares how two groups change over time relative to each other) and found it would detect a doubling of unemployment in the high-exposure group, from around 3% to 6%. It would detect a disruption equivalent in scale to the Great Recession concentrated in white-collar work. Neither signal has appeared.
This does not mean AI will never affect unemployment. It means it has not, yet, at detectable scale.
The Young Worker Signal
Workers aged 22 to 25 entering high-exposure occupations are finding jobs 14% less often than peers entering low-exposure roles. That drop does not appear at all for workers over 25. It is the clearest early employment signal in the paper, and it is specific to new entrants.
Earlier research by Brynjolfsson et al. (2025) found a 6 to 16% fall in employment for young workers in AI-exposed fields. Anthropic's team investigated the monthly job-finding rate: how often do workers in this age group report starting a new job in a high-exposure versus low-exposure occupation?
The two series track closely through 2022 and 2023. From 2024 onward, they diverge. Job entry into low-exposure roles holds steady at around 2% per month. Entry into high-exposure occupations drops by approximately half a percentage point. Across the full post-ChatGPT period, the average decline is 14%, statistically significant at the borderline level.
Crucially, this pattern does not appear for workers over 25. It is specific to new entrants.
Several explanations are possible:
- Employers may be pausing junior hiring in roles where AI handles entry-level tasks
- New graduates may be redirecting into other fields
- Some young workers may be staying in existing jobs rather than switching
- Survey measurement of job transitions carries noise that makes borderline results uncertain
The finding does not yet distinguish between these mechanisms, but it is the most forward-looking signal in the paper.
What the BLS (Bureau of Labor Statistics) Forecasts for the Next Decade
For every 10 percentage point increase in observed AI exposure, the BLS projects 10-year employment growth to fall by 0.6 percentage points. An occupation at 70% observed exposure is projected to see near-zero net job growth over the next decade. This relationship holds for the observed exposure measure and does not appear for theoretical capability scores alone, confirming that real usage data is the better predictor.
Beyond unemployment trends, the paper matches occupation-level exposure scores against BLS 10-year employment growth projections for each role. The result is a forward-looking signal that theoretical frameworks miss entirely.
How to Read This Chart
- 0% observed exposure: jobs here (cooks, mechanics, lifeguards) are projected to grow around 4% over ten years
- 30% exposure: many business and finance roles, projected around 2.4% growth
- 50% exposure: financial analysts, insurance underwriters, projected around 1.2% growth
- 70%+ exposure: computer programmers and data entry keyers, projected near-zero growth
Growth projecting to zero does not mean those jobs disappear immediately. It means the BLS expects no net new positions to emerge in those fields, as opposed to the 4%+ growth in unaffected occupations.
What This Means for Your Career
The research does not provide a simple red or green signal. It provides a framework for thinking clearly about risk and adaptation.
Career Strategy by Exposure Level
| Exposure Level | Example Occupations | Strategy |
|---|---|---|
| High (50%+) | Computer programmers, data entry clerks, customer service reps | Shift toward architecture, oversight, tooling, and judgment-intensive work AI cannot handle |
| Medium (20-50%) | Financial analysts, writers, paralegals, data scientists | Build specialized domain depth in areas requiring novel interpretation and human accountability |
| Low (10-20%) | Sales engineers, healthcare practitioners, architects | Adopt AI as a productivity tool; competitive advantage comes from speed and capacity, not replacement |
| Zero | Cooks, mechanics, lifeguards, trainers, electricians | No near-term AI risk from automation; focus on traditional career development |
For developers specifically: Computer Programmers rank first at 75% observed coverage. This is not hypothetical. It reflects real Claude usage in professional coding contexts, in automated pipelines, in first-party API integrations. But 25% of programming tasks remain outside observed AI use, and those uncovered tasks tend to be the hardest: novel architecture decisions, security reasoning, understanding legacy systems, managing stakeholder tradeoffs. Concentrating expertise there increasingly separates engineers who are hard to replace from those who are not.
For a global view of which roles are growing across all sectors, the WEF Jobs Report 2025 analysis maps the same occupational shifts from the employer survey side.
If you lead an engineering or product team navigating AI exposure risk, Pooya Golchian's AI team consulting includes task exposure assessments and skills transition planning.
What to Watch Next
The authors designed this paper for ongoing updates. The Anthropic Economic Index is a living dataset, and observed exposure will shift as AI models improve and as organizations integrate AI more deeply into workflows.
Three signals are worth tracking:
- The closing rate of the theory-observed gap. Computer and Math went from 94% theoretical to 33% observed. When and how quickly that 33% climbs toward 50%, 60%, or beyond is the key leading indicator of actual disruption.
- Young worker entry rates by field. The 14% hiring slowdown for 22-25 year olds in exposed occupations is the most actionable early signal. If it persists or broadens to older cohorts, the displacement story becomes substantially more concrete.
- BLS projection revisions. The current -0.6pp (percentage points) per 10pp rule is calibrated on data through 2025. New model generations may steepen this curve.
The signal for mass unemployment has not arrived. The signal for structural change in who gets hired, into which roles, and at what career stage has already begun.
Citation
If you reference this research in your own work:
@online{massenkoffmccrory2026labor,
author = {Maxim Massenkoff and Peter McCrory},
title = {Labor market impacts of AI: A new measure and early evidence},
date = {2026-03-05},
year = {2026},
url = {https://www.anthropic.com/research/labor-market-impacts},
}Massenkoff, M., & McCrory, P. (2026, March 5). Labor market impacts of AI: A new measure and early evidence. Anthropic. https://www.anthropic.com/research/labor-market-impacts