README.md

Overview

This directory contains O*NET task mapping and automation vs. augmentation data from "Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations." The data and provided analysis are described below.

Data

• SOC_Structure.csv - Standard Occupational Classification (SOC) system hierarchy from the U.S. Department of Labor O*NET database
• automation_vs_augmentation.csv - Data on automation vs augmentation patterns, with columns:
  • interaction_type: Type of human-AI interaction (directive, feedback loop, task iteration, learning, validation)
  • pct: Percentage of conversations showing this interaction pattern Data obtained using Clio (Tamkin et al. 2024)
• bls_employment_may_2023.csv - Employment statistics from U.S. Bureau of Labor Statistics, May 2023
• onet_task_mappings.csv - Mappings between tasks and O*NET categories, with columns:
  • task_name: Task description
  • pct: Percentage of conversations involving this task Data obtained using Clio (Tamkin et al. 2024)
• onet_task_statements.csv - Task descriptions and metadata from the U.S. Department of Labor O*NET database
• wage_data.csv - Occupational wage data scraped from O*NET website using open source tools from https://github.com/adamkq/onet-dataviz

Analysis

The plots.ipynb notebook provides visualizations and analysis including:

Task Analysis

• Top tasks by percentage of conversations
• Task distribution across occupational categories
• Comparison with BLS employment data

Occupational Analysis

• Top occupations by conversation percentage
• Occupational category distributions
• Occupational category distributions compared to BLS employment data

Wage Analysis

• Occupational usage by wage

Automation vs Augmentation Analysis

• Distribution across interaction modes

Usage

To generate the analysis:

1. Ensure all data files are present in this directory
2. Open plots.ipynb in Jupyter
3. Run all cells to generate visualizations
4. Plots will be saved to the notebook and can be exported

The notebook uses pandas for data manipulation and seaborn/matplotlib for visualization. Example outputs are contained in the plots\ folder.

Data released under CC-BY, code released under MIT License

Contact

You can submit inquires to kunal@anthropic.com or atamkin@anthropic.com. We invite researchers to provide input on potential future data releases using this form.