How we work

We read federal bills with AI, end to end, and explain what they do in plain English. Here is exactly how, and where the limits are.

How we pick which bills get a full read

Congress produces far more bill text than anyone can deep-read, so we triage. A fast, inexpensive AI pass scores every bill we ingest on two things: how much it matters (money spent, people affected, powers granted) and how hard it is for a regular person to read (length, cross-references, last-minute changes). Bills that score high on both get the full treatment: a larger AI model reads every page, section by section, and writes the analysis you see here. Bills below the line still get a page with their official status and sponsor, marked honestly as not yet deep-analyzed, with a link to the full text on Congress.gov.

How we flag a buried provision

A buried provision is language that most people would miss but would care about if they saw it. When we deep-read a bill we look for seven specific patterns:

• Vague words, big money. Loose language like “such sums as may be necessary” attached to large spending or broad authority.
• Contradicts the title. A section that works against what the bill's own name and stated purpose promise.
• Imported from a failed bill. Language that already failed to pass on its own and was tucked into a bigger, must-pass bill.
• Quietly consequential. A short, dry-sounding section that changes who has power, who gets paid, or who is exempt.
• Constitutional question. A provision that courts could plausibly strike down or that shifts power between branches of government.
• Sovereignty transfer. Decision-making authority handed from elected officials to bodies voters cannot vote out.
• Private or foreign power. Government powers, immunities, sensitive data, or critical infrastructure handed to private companies or foreign entities. We apply this the same way no matter which company or country benefits.

Every flag names the section it came from and explains, in a sentence or two, why we flagged it. You can expand the exact bill text behind any flag and judge it yourself.

How we keep it party-neutral

We never say a bill is good or bad, and we never say who should win. We say what a provision does, who it affects, and what the tradeoff is. After the AI writes an analysis, an automated check scans every sentence against a list of loaded political words and framings; any hit blocks publication until the sentence is rewritten neutrally. A sponsor's party appears as plain fact, the way a phone book would list it, and never as a lens.

What we cite, and what we do not claim

Every analysis links to the official bill text on Congress.gov, with the date we retrieved it. Voting records come from Congress.gov, house.gov, and senate.gov, and are shown as official public record, separate from our analysis. We do not predict whether a bill will pass, and we do not take editorial positions. AI can make mistakes, which is why every claim is anchored to text you can check.

When we get it wrong

We fix it, we date the fix, and we keep the old version visible. A corrected bill page carries a banner for 30 days saying what changed, and the note stays in the page's version history after that. Our full policy is on the corrections page. If you spot an error, tell us; the “Report an error” link in the footer of every page goes straight to us.