Reporting regimes diverge
FRONTIER, the AI Incident Reporting Act, the AI Flaw Reporting bill, the Secure AI Development Act, and the intelligence authorization use different triggers, clocks, recipients, and mandatory or voluntary structures.
77
Measures and legislative vehicles tracked
—
Core AI safety and security measures
—
Enacted laws and provisions
—
States in curated coverage
Legislative landscape
The inventory is organized by the principal legislative mechanism, with secondary tags preserving cross-cutting functions such as CBRNE, incident reporting, preemption, and model security.
Analytical taxonomy—not an official congressional classification.
Measures by primary legislative type
Outlook and final status
77 entriesTracking scope
3 layersSearchable inventory
Search full text, filter by legislative type and committee, compare measures, and open any row for the policy substance and committee dynamics.
| Measure | Title | Type | Sponsor | Status | Committee | Outlook / result |
|---|
No measures match the current filters.
State coverage and comparison
A curated safety/security view. Highlighted states have at least one officially verified record; select a state to filter the inventory.
Discovery via Open States · publication only after official-source review
Coverage map
Tile map is schematic. Gray states have not yet entered the curated starter set.
State comparison
| State | Tracked | Enacted | Core | Latest action |
|---|
What to watch
The bill count overstates policy coherence. The central challenge is reconciling overlapping and sometimes incompatible architectures across committees and vehicles.
Evaluation is split across theories
Pending bills alternate between mandatory developer assurance, secure pre-release testing, federal evaluation capacity, federal-use testbeds, and voluntary documentation standards—with no clear reciprocity rule.
Preemption remains the fault line
FRONTIER’s targeted model differs sharply from the failed broad state-law moratorium. Any new amendment should be checked line by line for coverage of developers, deployers, generally applicable law, civil rights, privacy, minors, and procurement.
Model theft has a lifecycle gap
Prevention, detection, incident reporting, investigation, sanctions, and export-control response are spread across separate bills and agencies rather than joined into one statutory pathway.
CBRNE coverage is not end-to-end
Congress is addressing synthesis screening, model evaluation, terrorism assessment, and military incidents, but no single measure connects capability findings to controlled access, screening, escalation, and response.
Funding decides implementation
NIST, CAISI, CISA, NSA, Defense, and intelligence-community duties depend on appropriations. Authorization without people, compute, secure facilities, and evaluation funding may remain aspirational.
Methodology & maintenance
Official legislative sources provide the facts. The tracker adds clearly labeled analysis, cites the evidence behind each record, and keeps a public history of changes.
Inclusion and verification
- Includes material safety/security duties, not every bill that mentions AI.
- Open States supplies discovery candidates only.
- Every published state record is checked against an official legislative or executive source.
- Read the formal inclusion criteria.
Intended operating cadence
- The site owner aims to run an automated status check each Friday.
- During major markups and floor debates, the owner may check key committees and must-pass bills more often.
- Automated updates are reviewed before publication, especially when they change analytical labels or conclusions.
- The owner aims to run broader completeness reviews monthly and at the end of each legislative session.
Where the information comes from
- Official facts: bill identity, sponsors, actions, committee referrals, votes, and text versions come from government sources.
- Tracker analysis: scope, legislative type, context, and outlook are editorial judgments made for this project.
- Update history: GitHub records what changed, when it changed, and the sources used.
Open data
Download the current generated dataset, its full validation schema, or a spreadsheet-friendly CSV. All three are rebuilt from the same canonical records.