SUBSCRIBE

How we verify

BSKiller is a nightly fact-check of the AI claims everyone is already seeing. Every claim gets one of four rulings with the receipts attached, and every claimant gets a permanent track record. This page is the full method, because a methodology you can't audit is just branding.

THE RECORD SO FAR: 160 claims · 810 receipts · nothing unsourced

The four rulings

Every checked claim ends in exactly one of these. No composite scores, no vibes:

What we check (the coverage law)

We fact-check the claims people are actually seeing, not the claims we happen to find. A trend desk reads the biggest AI amplifier newsletters and the loudest forums every day; any claim echoed by two or more independent amplifiers becomes a mandatory candidate for that night's issue: we stamp it, or the issue notes say why not. A fact-checker that misses the claim everyone saw is not a fact-checker.

How a claim becomes a ruling

  1. Extracted, not summarized. We quote the assertion the claimant actually made, attributed and dated, with a link to where they made it.
  2. Receipts, plural, including a hostile one. Independent sources across separate domains, and at least one chosen to argue against our ruling. Every source URL is checked live the day we publish and archived so a receipt cannot quietly rot. Every quoted line is matched to its source word for word. Rulings without receipts don't ship.
  3. The named mechanism. Where a claim misleads, we name HOW: the specific trick (reference-class swap, denominator hiding, self-marked benchmark, and so on), because "misleading" without a mechanism is just an opinion with a scowl.
  4. An adversarial gate, then a second, independent one. Every story passes a red-team pass and a separate automated review by two independent reviewers before it can ship. A factual flag from a single reviewer is re-run to confirm it isn't noise; a story that keeps failing is quarantined: publicly absent, never quietly published.
  5. Falsifiable, with a stated flip condition. Where a ruling depends on something that could change, the claim page says what evidence would flip it, and Promise Watch tracks the claims with deadlines until they resolve.

Track records: the memory

Rulings accumulate. Every claimant (company, executive, pundit) has adossier: their claims over time and how each ruling landed. That is the instrument: before you trust a word, check the speaker's record. Free readers get 5 track-record lookups a month;Pro is unlimited.

Corrections are loud

Fixed claims keep their history; we never silently edit a ruling. If you can show a ruling wrong, it changes, and the change is public: email[email protected]with the claim ID and your evidence. If you're right, the correction ships with credit.

The BS Index (retired 2026-09-05)

Earlier editions carried a "BS Index," a weighted average that scored each verdict tier and averaged the issue. We retired it. The weights were editorial choices, not measurements, and averaging them over a handful of stories produced a number with more precision than meaning. A site that exists to call out invented precision should not publish its own. The per-claim rulings and the named spin mechanisms are the instrument; no composite score sits on top of them.

Built to be quoted

Every ruling ships twice: once as the page you're reading, and once as machine-readable ClaimReviewdata embedded in the page, the same structured format Google's fact-check results use, so a search engine or an AI assistant can quote our rulingwith the rating, the date, and attribution back to us, instead of guessing. If a machine repeats one of our calls, it repeats the receipts too.