Your name is on copy you didn't type

You're in the copy studio. What this chapter is: hear a client's voice, write on-voice variants, and close every checkable claim against a source before it ships. A voice pack is the written rules for how they sound. A claim check is the table of facts that must close. You do not need regex to start.

Here's the job now, whether anyone announced it or not. Muck Rack's State of AI in PR survey (564 practitioners, fielded December 2025) puts generative AI use at 76% — and the UK journalist numbers from the Reuters Institute (1,004 respondents) show 56% using AI professionally every week, with copy-editing among the top monthly tasks and one in six generating parts of articles at least monthly. The drafting is shared. The byline is not. When a number in the copy is wrong, nobody asks which model typed it. They ask who shipped it. That's you.

Two named failures hang over this chapter, and you should know them the way a pilot knows famous crashes. CNET published 77 machine-written finance stories under "CNET Money Staff" and ended up issuing corrections on 41 of them — a 53% correction rate, with plagiarized phrasing found in the mix. Sports Illustrated's operator ran product reviews under AI-generated author personas from a content vendor; when Futurism published the bylines, the content came down and the vendor contract died, and Arena Group denied the articles themselves were AI-written. CNET is the correction-rate story. SI is the byline story.

So this chapter builds the two systems a working copywriter actually needs: a voice that survives scale, and a fact gate that survives deadlines.

Why it matters

• Voice is your product, and "sounds like them" doesn't scale past one draft. A model can imitate a client for three sentences and drift by the twelfth variant. The fix isn't better vibes — it's measuring the client's real published copy and turning the measurements into rules a script can enforce on every variant of every batch.
• The verification gap is where the market is wide open. Most training drills prompting endlessly, while the guardrails limp behind the adoption. Muck Rack's survey of PR professionals — a neighboring beat, used here as a hook — has 76% using generative AI, and 51% now report a formal AI use-case policy at work, up from 21% in 2024. Those 51% are workplace policies. What almost nobody drills is the after: extracting every checkable claim from generated copy and closing each one against a source before it ships. Researchers who compared AI policies at 52 news organizations found them converging on exactly this — human review, disclosure, no publishable AI output unchecked. The policies exist. The procedure is what's missing. That procedure is this chapter.
• Corrections culture is a career asset. SPJ's code says it plainly: take responsibility for the accuracy of your work; verify information before releasing it; use original sources whenever possible. A copywriter with a corrections log and a claims table is more hireable in 2026, not less — because clients have read the CNET story too.

What AI specifically gets wrong

• It regresses to the mean of everyone's copy. Your client writes nine-word sentences and hasn't used an exclamation point since 2019. The model's training data loves "Unleash your adventure!" Left unchecked across a 12-variant batch, drift wins — not on variant one, on variant nine.
• It fills a claim-shaped hole with a claim-shaped number. "Rated 4.8 stars." Says who? Per coverage of AP's standards, AI output is treated as unvetted source material — that's the correct posture. The model is a draft source, never a fact source. Every number, superlative, and "free" in generated copy is unverified by definition until a human closes it against a document.
• It never flags what it invented. A human junior writer says "I wasn't sure about the warranty line." The model delivers the invented warranty line with the same confidence as the true weight. Confidence carries zero information; only the claims table does.
• It leaves no trail. When the correction request lands three weeks later, "which prompt, which pack version, which model" is unanswerable unless you logged it at ship time. Almost no training program teaches this. Your kill logs and batch records will.

What you'll be able to do at the end

• Fingerprint a client's voice from their real published samples — sentence length, contraction rate, exclamation count — and derive a versioned voice pack with provenance, not a taste memo.
• Run a variant batch through a voice gate: every survivor on-voice and inside its channel's character budget, every kill logged with a named reason you can defend to the client.
• Extract every checkable claim from a piece of generated copy, verify each against a source table, and hold the ship decision hostage until every row closes — soften or strike what won't.
• Write the correction entry when something wrong already ran: what ran, what was true, which gate gets stronger. The CNET drill, run before you need it.

The first step of this chapter has you measure a live product paragraph in the editor. Everything after that is the same kind of script — plain stdlib Python, no API calls, no credits. The scripts are small on purpose: the skill is the discipline, and the discipline transfers to whatever model drafts your copy next quarter.