Making Sense of AI Policy With Computational Methods
Lawmakers have introduced a flood of AI and automated decision-making bills, and many borrow language from each other. That creates hidden interactions that are hard to see bill by bill. A new analysis of 1,804 state and federal proposals from 2023 through April 2025 shows how carefully chosen computational tools can surface those patterns and stress-test individual bills before they become law.
The sample was built with curated AI-related keywords and includes a wide range of proposals. About 30% create task forces, and many target specific sectors like health care. The key move: match the tool to the question, instead of forcing AI into every step.
What the data shows across bills
Topic modeling surfaced clear themes. Nearly 500 state bills focus on generative AI, with at least one bill in every state and more than 50 in New York alone. Many target deepfakes in elections and explicit synthetic imagery, often via watermarking or disclosure requirements.
Task forces are a second major thread. There are over 400 state bills and more than 100 congressional proposals creating AI-related task forces, signaling broad interest but mixed depth. That aligns with the overall count where roughly a third of the sample is task-force-only.
Policy diffusion is visible in the text. Lawmakers frequently reuse language from model bills or high-profile states, a dynamic sometimes called the "California effect." One Lawyers' Committee model bill seeded the AI Civil Rights Act of 2024 and shares substantial language with 11 other bills across Congress and states including Illinois, Massachusetts, New York, and Washington. A model bill linked to a large HR firm showed up in six states, plus a near match in Oklahoma.
Looking inside a single bill
Definitions decide what a bill actually does. They set scope, name accountable entities, and ripple across the text. Visualizing those definitions as a graph helps reveal hidden dependencies and weak spots.
Cycles in definitions can create ambiguity and loopholes. The Fair Credit Reporting Act contains a cycle between "consumer report" and "consumer reporting agency," a pair that has driven years of debate. Detecting cycles early gives drafters options to clarify terms before passage.
Graph methods can also flag the most central terms. In the American Privacy Rights Act of 2024, "sensitive covered data" sits at the center of the network. That kind of signal tells policy teams where precision matters most.
Methods you can apply now
- Corpus curation: Define inclusion criteria (keywords, sectors, jurisdictions). Track bill types (e.g., task force vs. substantive requirements) to avoid overgeneralizing.
- Trend tracking: Use topic modeling to map themes over time and by state. Pair with counts to spot spikes in areas like deepfakes or content disclosure.
- Policy diffusion: Compare texts using n-gram overlap and embeddings. Cluster similar bills and check against known model bills to trace lineage.
- Definition graphs: Extract defined terms and references to build a directed graph. Detect cycles, compute centrality, and flag terms that need tighter language.
- Review loop: Treat computational flags as triage. Pair findings with legal and subject-matter review, then track revisions so you can measure improvement.
Recommendations to strengthen this work
Standardize legislative data. Common file formats, consistent definition sections, and clear cross-references make analysis faster and more reliable. The United States Legislative Markup format is a strong anchor for this effort. See USLM documentation and the Congressional Data Coalition for community and practices.
Adopt a multilingual lens for regions under U.S. jurisdiction. English-only analysis will miss developments in places like Puerto Rico (Spanish) and Hawai'i (Hawaiian). Build language-specific models and involve native speakers and regional policy experts to ground the interpretation.
What this means for policy teams
Computational tools give you speed and coverage; expert review provides judgment. Use both. Track trends across states, spot model-bill influence, and pressure-test definitions with graphs and cycle checks.
The result is practical: fewer blind spots, clearer scope, and stronger bills. Start small, publish your method, and improve with each session of the legislature.
Your membership also unlocks:
