AI content verification is a claim audit with a style pass attached.

Before you read for style, list what the copy asserts. Numbers, dates, quotations, superlatives, comparisons, awards, regulatory statements, promises about outcomes. That list is the verification plan, and it is usually shorter than people expect.

Generated drafts concentrate claims in the places that sound most authoritative: the opening statistic, the three-bullet proof block, the comparison table, the customer quote. Those are the paragraphs to read slowly.

Anything that cannot be given an inventory entry is either opinion, which should be marked as such, or decoration, which should be cut.

The test for a citation is not whether it looks credible. It is whether a reader can open it, see the same number, and know when it was true. A study named without a year, a survey with no sample size, a growth figure with no denominator, all fail that test.

Generated copy produces a specific failure worth watching for: the plausible citation. An organisation, a report title and a year, combined in a way that sounds researched and cannot be found. Absence of the source is not proof of invention, but it is proof that the claim cannot ship as written.

Where a claim is true but unsourced, the fix is usually to find the source or soften the claim. Rewriting the same number with more confidence is the one option that must be removed from the process.

Tell-list phrasing is not a moral failing, it is a signal that the paragraph was assembled rather than written. Stock transitions, inflated verbs, symmetrical triads and em dashes used as punctuation glue all show up in generated drafts.

The console measures a set of these, along with hedging density, absolute claims, repeated sentence openers and the variance in sentence length. Flat rhythm is a useful tell: human prose mixes a five-word sentence with a thirty-word one.

None of those numbers prove authorship, and they are not meant to. They point at paragraphs that need a human sentence, which is a different and more useful instruction.

A heading set is a contract with the reader: each heading promises a section that delivers it. Generated documents often break that contract by adding a heading to look thorough and then writing a paragraph that restates the title.

The structural checks catch mechanical breaks: skipped ranks, duplicate headings, no level-one heading at all, a long piece with no sections. What they cannot catch is a section that exists to satisfy a template, and that is an editorial call.

Say what the page is for in one line. If a section does not advance that, it is length rather than content.

A publish gate is smaller than a release gate and stricter about a few things. Block on unsourced statistics, absolute claims a legal team has not seen, and any placeholder or instruction left in the copy. Warn on style density and readability.

Then require a named reader for the pieces where being wrong is expensive: regulated industries, health, finance, legal, anything asserted on behalf of a company with a byline.

The record should say who read it and what they changed. A year later, when someone asks why a number appeared on the site, that record is the difference between an answer and a shrug.

The failure mode of AI editing is homogenisation: every sentence becomes medium-length, every claim gets a hedge, and the result reads like the median of the internet. Verifiability improves and the piece stops being worth publishing.

So verify the claims first, mechanically and without mercy, then edit for voice with the constraint that the verified facts cannot drift. Shorten one sentence to four words, let another run long, cut the transition that explains what the next paragraph will say.

That order matters. Editing before verifying produces a well-written lie, and verifying after editing means re-checking the numbers somebody moved while polishing.