A claim is anything a reader could call you on
Before you can verify claims, you have to find them all — and "read it carefully" is not a finding method. You read your own copy the way you wrote it: nodding along. The claim your eye slides over is precisely the one the model invented, because the model's whole talent is producing sentences your eye slides over. So extraction goes mechanical.
What counts as checkable? The test: could a reader, holding evidence, prove this sentence wrong? "The pack that skips the parts you skip" — unfalsifiable, pure voice, no row needed. "Weighs 890 grams" — a scale settles it. Checkable. Which gives you three families to hunt:
- Numbers. Weights, prices, thresholds, star ratings, durations. Every digit in copy is a claim. The regex
\d+(?:\.\d+)?catches integers and decimals — 890, 75, 4.8 — without dragging trailing punctuation along. - Absolutes. "Free", "guaranteed", "every", "always." Small words, enormous promises. "Free shipping" is a checkable claim about a policy document that has conditions your copy just erased.
- Superlatives. "Lightest", "fastest", "first." A superlative is a claim about every competitor at once — the most expensive kind of sentence to be wrong in, because the competitor it's wrong about is the one who'll notice.
Run the editor. One line of re.findall on the lowercased copy, four claims out: 890, free, 75, guaranteed. That's the checklist for the verification table — each one becomes a row, each row needs a source, and the copy waits until they're all closed.
Is a keyword-and-number regex the whole art of claim extraction? No — it will miss "backed by our warranty" and flag a harmless "28" in a product name, and you'll tune it per client the way you tune the banned list. Doesn't matter. The bar it has to clear isn't perfection; it's your tired eyes on deadline, and it clears that bar every single run. The regex never nods along.
UK journalist surveys show fact-checking among the least common AI uses — 12%, against 22% for research. Practitioners trust models least exactly where the stakes are highest, which is correct instinct. Notice the pipeline here respects it: the model gets no vote in this lesson. Extraction is a regex, verification is a lookup against documents, and both would work exactly the same if the copy had been typed by a human who happened to be sloppy.