AI Governance Legislation Tracker
Monitoring state & federal AI legislation across the United States, for policymakers, researchers, and the public.
Methodology
Every bill in this tracker is placed on a two-axis taxonomy: what it regulates, and how it regulates. Classification is done by a language model and checked against trained human coders. This page documents the labelling scheme and how that checking is carried out.
Codebook v0.2 · Accurate as of October 18, 2026
The corpus
Bills are pulled from LegiScan across all 50 states and Congress, then filtered to those touching artificial intelligence or automated decision systems. Coverage below 100% means the remaining bills have not yet been through the classifier, not that they were judged irrelevant.
The two axes
A single-axis scheme forces a choice between describing a bill's subject and describing its legal instrument. Separating them lets a bill be both, which is usually the truth: a hiring-algorithm bill is Employment/HR on one axis and Bias Audits on the other. Bills carry as many labels per axis as apply.
How validation works
Classifier output is not taken at face value. A stratified sample of bills is assigned to trained human reviewers who code each bill independently, without seeing the model's answer. A share of the sample is deliberately assigned to two reviewers so their agreement can be measured. Where two reviewers disagree, an adjudicator picks the canonical labels, and that reconciled coding becomes the gold standard the classifier is scored against.
- Stratified sample drawn across states so no jurisdiction dominates.
- Reviewers code independently against a written codebook.
- A portion is double-coded; Cohen's κ measures agreement.
- Disagreements are adjudicated into a single reconciled coding.
- The classifier is scored against that reconciled gold standard.
Every submission is stamped with the codebook version in force at the time, so a later revision does not silently contaminate earlier codings.
Limitations
- Coverage depends on LegiScan. Bills it does not index, or indexes late, are absent here.
- Classification runs on bill title, summary and text. A bill whose AI provisions sit in an amendment not reflected in the indexed text can be missed.
- Validation measures agreement on bills that were sampled. It does not measure how many AI-related bills the upstream filter failed to surface in the first place.
- Labels are multi-label by design, so counts across an axis sum to more than the number of bills. Percentages should be read as share of bills carrying a label, not a partition of the corpus.