Methodology

Our analysis integrates multiple authoritative data sources to provide a comprehensive view of AI's impact on occupations. We continuously monitor new research to update our assessments.

• Massenkoff & McCrory (2026) - Labor market impacts of AI: observed exposure metric from real Claude usage dataPrimary
• Eloundou et al. (2023) - GPTs are GPTs: theoretical task exposure framework (beta scores 0, 0.5, 1)
• Brynjolfsson et al. (2025) - Canaries in the Coal Mine: observed employment effects using ADP payroll microdata
• U.S. Bureau of Labor Statistics (2024) - Employment Projections 2024-2034 with occupation-level growth rates
• O*NET SOC Classification System - standardized occupation and task taxonomy used across all assessments

We use four primary metrics to quantify AI's impact on each occupation. Each metric captures a different dimension of how AI interacts with work tasks.

Overall Exposure

A combined metric that synthesizes theoretical and observed exposure data to provide a single summary score (0-100) of how much an occupation's tasks overlap with current AI capabilities.

Theoretical Exposure

Measures what AI could potentially automate based on academic research and capability assessments. Derived primarily from Eloundou et al. (2023) beta task exposure scores, which rate each task's susceptibility to LLM automation.

Observed Exposure

Measures what AI actually does in practice based on real-world usage data. Derived from Anthropic's analysis of millions of Claude conversations mapped to O*NET occupational tasks, reflecting actual adoption patterns.

Automation Risk

The probability of significant job displacement within the assessed time frame. Combines exposure metrics with employment trend data, wage levels, and task substitutability to estimate the likelihood of workforce reduction.

We decompose each occupation into individual tasks and assess the degree to which current and near-future AI systems can automate each task. This task-level approach provides more nuanced insights than whole-occupation estimates.

Task-Level Decomposition

Each occupation is broken down into its constituent tasks using O*NET's Detailed Work Activities (DWAs). We assess each task independently rather than making blanket judgments about entire occupations.

Beta Score Methodology

Following Eloundou et al. (2023), each task receives a beta score: 0 (no exposure), 0.5 (partial exposure with human oversight), or 1 (full exposure to AI automation). These scores are aggregated to produce occupation-level metrics.

Time Series Construction (2023-2028)

We construct time series data using actual measurements for 2023-2025 and estimated projections for 2026-2028. Actual data is clearly distinguished from estimates in all visualizations using solid vs. dashed lines.

Projection Methodology

Forward-looking estimates (2026-2028) are based on observed trend rates, announced AI capability improvements, and BLS employment projections. All projected values are clearly marked as estimates.

Transparency about our limitations is essential to responsible analysis. Users should consider these factors when interpreting our data.

Sample Sizes

Our observed exposure data is based on millions of Claude conversations, providing robust statistical coverage for high-usage occupations. However, less common occupations may have smaller sample sizes and wider confidence intervals.

Geographic & Occupational Coverage

Currently covering 55 occupations with plans to expand to 200+. Data primarily reflects the U.S. labor market and English-language AI interactions, which may not fully represent global patterns.

Update Frequency

Core metrics are updated when new research publications or data releases become available. BLS projections are updated annually. Observed exposure data is updated as new analyses are published by Anthropic.

Theoretical vs. Observed Gap

There is often a significant gap between theoretical exposure (what AI could do) and observed exposure (what AI actually does). Adoption barriers, regulatory constraints, and organizational inertia mean actual AI impact typically lags behind technical capabilities.

We maintain a transparent record of major data updates and methodology changes.

All data points are linked to their original sources. We provide full citation information for transparency and independent verification. See the complete reference list below.