AI Labor Market Impact Measurement

8 posts · Dec 2024 – Sep 2026

What this trail covers

Anthropic is developing novel frameworks and measures to quantify the labor market impacts of AI. This includes creating a new metric called 'observed exposure,' which combines theoretical LLM capabilities with real-world usage data, weighting automated and work-related uses more heavily. The goal is to provide a more predictive and reliable approach to understanding AI's economic disruption than previous methods, enabling early identification of vulnerable jobs and tracking the narrowing gap between human and AI capabilities. This post extends this by developing a robot exposure index based on current robot capabilities and cost-competitiveness, and backtesting its predictive power against historical wage and employment data.

All Anthropic trails

Sep 2026Can we predict the jobs robots will do?This post introduces a 'robot exposure index' to predict the impact of robotics on jobs. It details a methodology using Claude to assess current robot capabilities for performing job tasks across different environmental controls (E0-E3). The post quantifies that robots can perform 74% of physical tasks in the US, but are cost-competitive for only 0.3% of tasks. It also presents a historical backtest showing that jobs with higher robot exposure experienced greater wage and employment declines over 50 years, validating the predictive power of the exposure index.
Sep 2026Introducing the Anthropic Economic IndexThis post introduces the Anthropic Economic Index, a new initiative to understand AI's impact on labor markets. It details the methodology using the 'Clio' analysis tool to map anonymized Claude.ai conversations to O*NET occupational tasks. Key findings include the concentration of AI use in software development and technical writing, a higher rate of augmentation (57%) versus automation (43%), and a correlation between AI use and mid-to-high wage occupations. The post also discusses limitations of the study, such as distinguishing work vs. hobby use and the scope of data analyzed (Claude.ai Free/Pro). It announces the open-sourcing of the dataset and plans for longitudinal analysis.
Sep 2026Anthropic Economic Index: Insights from Claude 3.7 SonnetThis post details the analysis of Claude.ai usage data following the launch of Claude 3.7 Sonnet, focusing on its impact on labor markets. It introduces a new bottom-up taxonomy of 630 granular usage categories to complement the existing O*NET task mapping. The analysis highlights shifts in usage towards coding, education, and science, and identifies that the 'extended thinking' mode is predominantly used for technical and creative tasks. It also provides detailed breakdowns of augmentation vs. automation at task and occupation levels, noting specific patterns like 'task iteration' for copywriters/editors and 'directive' behavior for translators/interpreters. The post also discusses methodological changes, including filtering safety classifier-flagged conversations instead of occupational relevance, and using Claude 3.7 Sonnet for improved classification accuracy.
Mar 2026Labor market impacts of AI: A new measure and early evidenceThis post introduces a new measure of AI displacement risk called 'observed exposure,' which combines theoretical LLM capability (using Eloundou et al.'s β metric) with real-world usage data from Anthropic's Economic Index. It details the methodology for calculating observed exposure, emphasizing the weighting of automated and work-related uses. The post presents early evidence showing a negative correlation between observed exposure and projected job growth, and analyzes the demographic characteristics of highly exposed workers. It also investigates the impact of observed exposure on unemployment trends, finding no systematic increase but suggestive evidence of slowed hiring for younger workers in exposed occupations.
Nov 2025Estimating AI productivity gainsThis post details a methodology for estimating task-level time savings by using Claude to analyze real-world conversations from Claude.ai. It describes the process of generating time estimates with and without AI assistance, aggregating these estimates to O*NET occupations, and extrapolating them to estimate aggregate US labor productivity growth. The post also includes validation analyses, including self-consistency testing of Claude's estimates and external benchmarking against software development task completion times. The analysis suggests that Claude's estimates provide directional information comparable to human developers' own estimates, albeit with some compression and tendency towards overestimation. The findings indicate that current-generation AI models could significantly increase US labor productivity growth.
Advertisement
Apr 2025Anthropic Economic Index: AI's impact on software developmentThis post details an analysis of 500,000 coding-related interactions across Claude.ai and Claude Code. It quantifies the shift towards automation (79% on Claude Code vs. 49% on Claude.ai) and identifies user-facing app development (JavaScript, HTML, CSS) as a primary use case. The analysis also highlights startups as early adopters of Claude Code compared to enterprises. The findings are derived using a privacy-preserving analysis tool to categorize conversations by topic and automation/augmentation patterns, providing empirical data on AI's impact on software development workflows.
Dec 2024Clio: Privacy-preserving insights into real-world AI useThis post introduces Clio, an automated, privacy-preserving analysis tool for understanding real-world language model usage. Clio extracts facets from conversations, performs semantic clustering, generates descriptive titles and summaries, and builds hierarchies, all powered by Claude. It has revealed insights into user behavior, such as a significant emphasis on coding-related tasks (over 10% of conversations), educational uses (over 7%), and business strategy (nearly 6%). Clio also identified thousands of smaller, diverse conversation clusters. The tool supplements traditional safety measures by enabling bottom-up discovery of patterns, identifying coordinated misuse (e.g., SEO spam networks), and detecting emergent risks. It aids in refining Trust and Safety classifiers by highlighting false positives (e.g., resume advice, security programming questions) and false negatives (e.g., translation-based violations). Ethical considerations for Clio's development included addressing false positives, potential misuse, user privacy, and user trust, with safeguards like no automated enforcement actions, extensive validation, strict access controls, data minimization, and regular privacy audits.