U.S. State AI Legislation Corpus: Classified Bills, Code, and Validation Annotations

This deposit contains the dataset, classification system, and validation materials accompanying:

Bhargava, H. K., & Soliman, M. (in preparation). The Structure of U.S. State AI Policy: Evidence from a Multi-Dimensional Bill Classification.

Contents:

DATASET
– Final_Database_10.csv: 2,571 U.S. state AI bills introduced between 2019 and 2026, sourced from the National Conference of State Legislatures, enriched with sponsor metadata via LegiScan, and classified along a seven-element taxonomy by a multi-agent LLM system.

VALIDATION ANNOTATIONS  
– annotator_alinda.csv, annotator_ayan.csv, annotator_rakshita.csv: Independent human annotations on a stratified sample of 91 bills (83 AI bills plus 8 non-AI controls), used to validate the classifier against trained human judgment.

CLASSIFICATION SYSTEM
– Bill_Analyst_Agent source code: Multi-agent classification pipeline built on LangGraph and LangChain, including the keyword pre-filter, seven dimension analyst agents with scoring rubrics, and the judge agent revision loop. Reproduces the analyses in the paper from raw bill text.

DOCUMENTATION
– Database_Creation_doc.md: Complete documentation of data construction, schema, classification methodology, and validation procedures.

The seven taxonomy dimensions:
1. Product Safety, Risk, & Accountability
2. AI Inputs & IP
3. Sectoral Use & Application  
4. AI Markets & Competition
5. Institutional Framework & Processes
6. AI Advancement and Development
7. Existential & Societal Risk

Requirements (to reproduce classification): Python 3.10+, OpenAI API key, LegiScan API key.

Code is also available at: https://github.com/marksoliman3/Bill_Analyst_Agent (release tag v1.0.1-rp-submission).