Methodology

Sources

Everything here is built on authoritative, primary sources. We ingest directly from official records and established scholarly catalogs — no scraping of secondhand summaries — and every published item keeps its source link and date so you can go check the original yourself.

• Congress.gov — the Library of Congress’s official record of federal legislation, for U.S. AI bills and resolutions.
• LegiScan — state legislation normalized across all 50 states, for AI bills moving in the statehouses.
• The Federal Register — the U.S. government’s official journal, for executive orders and federal agency actions.
• arXiv — open-access preprints, for new AI and machine-learning research.
• OpenAlex — an open catalog of scholarly work, used to broaden research coverage beyond arXiv.
• A curated news allowlist — a hand-picked set of feeds from authoritative AI outlets and institutions, rather than an open firehose of headlines.

How we summarize

Every summary is written from the actual source text — the bill, the executive order, the paper, or the article itself — and not from a language model’s general memory. Each incoming item runs through the same pipeline:

• A first, lightweight pass (a fast Claude model, Haiku) screens the item for relevance — is this actually about AI’s measurable impacts? — and extracts its impact-area tags and, for legislation, its jurisdiction and status. Items that aren’t relevant are dropped before any further work.
• A second, stronger pass (Claude Sonnet) reads the full source text and writes the plain-English summary, grounded in that text.
• Both passes are constrained to a fixed, structured output shape that is re-validated in code before anything is written to the database — the model is held to the schema, not trusted to follow it.
• The cited claims in a summary are checked back against the source text, and each item is assigned a confidence score. These runs are processed as batched jobs for efficiency.

Low-confidence, unverified, or otherwise uncertain items are held back for closer human review rather than published automatically.

Neutrality and integrity

This dashboard aggregates and analyzes sourced evidence. It is not an advocacy site and does not take positions for or against AI; its job is to show what is actually happening, accurately. That commitment rests on a few hard rules:

• No unsourced claims. Every published item links to its primary source — the statute, the enrolled bill, or the Federal Register notice, published by the body that issued it — and carries the date it was published or last updated. Policy items are verified against that primary text before they are published, so the link you follow is the document the summary was written from.
• Impacts both ways. We track positive and negative effects alike — the goal is an accurate picture, not a flattering one.
• Policy never publishes on a model’s say-so. Every policy item is read against its primary text by a person before it is published. Research and news items publish automatically only when they clear a confidence bar and a citation check against their source; anything that falls short waits for a person.
• Primary-source verification for policy. Policy items are re-verified against the primary source — the bill or order itself — before they are marked published, and a verification badge reflects that check. Legislative status is treated as a re-verified time series, never a fact stored once, because the law changes: Colorado’s 2024 AI Act (SB 24-205) was repealed and re-enacted by SB 26-189 in 2026, and a tracker frozen on the old text would be authoritatively wrong.
• Named analysis, not verified accounts. Analyst annotations carry an editorial byline — a name and affiliation entered by an editor — as attribution for the analysis. These are editorial attributions, not logged-in, account-verified identities.

Corrections policy

When something we published turns out to be wrong, we log a correction rather than quietly editing it away. Each correction records the date and a plain-English summary of what was wrong and what changed. Corrections follow the same draft-then-publish discipline as everything else — only published corrections appear here. Every published correction is listed in the log below, and a correction tied to a specific tracked item also appears as a dated note on that item’s own page, so the record is visible where the error was.

Corrections log

No corrections have been issued yet. When we correct a published item, the correction will appear here — dated, with a plain summary of what changed — and on the corrected item’s own page.

Pulse

Pulse shows weekly counts of what this tracker ingested. It is not a survey of any sector, party, or public: it measures what entered this tracker, not the world.

• Weekly counts. A daily job counts items per ISO week (Monday, UTC) by the kind of source, which we call a pillar: academia, industry, government, press, labor, civil society. Research and news are counted in the week they were published; bills are counted in the week of their most recent status change and include bills we track but have not published. The current week is partial. Weeks with fewer than five items of a kind are not shown. Weeks older than eight weeks are frozen and no longer rewritten by the daily job.
• Coverage rule. Coverage begins, for each kind of content, with the first calendar month that holds at least ten published items of that kind. Earlier weeks show no data, not zero, and are excluded from every average. The boundary is computed from the data on every refresh, never typed in by hand.
• Markers. Marked weeks reflect a change in this tracker, not a change in the field. Each marker is a recorded event in the pipeline, such as a revision of the relevance-gate prompt, a bulk publish, or a source outage, listed with its date under the chart it marks.
• Observed and reconstructed policy weeks. Bill status changes are observed as they happen only from the date this tracker first recorded a status change. Weeks before that date are reconstructed from the bill histories of bills our search found later, not observed as they happened, and bills that died early in 2025 are under-counted. The boundary is the earliest recorded observation, read from the status history on each refresh, and the reconstructed weeks are hatched on the chart.
• Impact-area tags. Impact-area tags are assigned by a model (Claude Sonnet 4.6) at ingestion. The relevance-gate prompt was revised on 2026-07-22, which changed which items entered the tracker; shares before and after are not directly comparable.