Forty-one of seventy-seven
Run the editor. Those are real numbers. CNET published 77 machine-generated finance explainers under the byline "CNET Money Staff." After it came out, CNET's own audit issued corrections on 41 of them — a 53% correction rate — with factual errors in basic financial math and, per Futurism's reporting, plagiarized phrasing in the mix. Around the same time, Sports Illustrated's operator was caught running product reviews under AI-generated author personas supplied by a content vendor; the content was pulled and the vendor relationship ended.
Note what the failure wasn't. It wasn't using a model to draft — half the industry does that, openly. The failure was structural: generated copy went to publication with no claim-by-claim check between the model and the reader, and no honest name accountable for each piece. The error rate was always going to surface. The only question was whether it surfaced in a pre-publication table or in the press.
The rule that follows
Journalism wrote the fix down decades before the models arrived. The SPJ code: take responsibility for the accuracy of your work; verify information before releasing it; use original sources whenever possible. And per coverage of AP's standards, AI output is "treated as unvetted source material" — which is the single most useful sentence about generated copy ever written. Not banned. Not trusted. Unvetted — the same status as a tip from a stranger. You can use it; you must check it first.
You write product copy, not news. Doesn't matter — the mechanism transfers whole. Your Drift 28 copy says 890 grams, free shipping over $75, rated 4.8 stars, guaranteed for life. Every one of those is a checkable claim, and the model that drafted the copy is capable of having invented any of them, with identical confidence either way. If the warranty is actually ten years, "guaranteed for life" isn't a voice problem. It's a refund dispute, a review-site pile-on, and depending on jurisdiction and luck, a legal letter — addressed to your client, forwarded to you.
The pipeline this lesson builds
Three moving parts, drilled in order: extract every checkable claim from the copy, mechanically, so none hide. Verify each against a source table — a real document, never the model. Gate the ship decision on every row being closed, and when something wrong has already shipped, log the correction so the same hole never opens twice. CNET rebuilt this process in public, under a news cycle. You get to build it this afternoon, for free.