The tell isn’t that it was AI-assisted, it’s that it reads like everyone else’s
Most brand style guides ban obviously robotic phrasing and call it solved. But the harder problem is subtler: a paragraph can be grammatically flawless, on-brief, factually correct, and still read as interchangeable with your competitor's paragraph, because both were smoothed by the same kind of assistant into the same safe, elevated, hedge-everything register. Readers don't consciously detect this. They just trust it less and remember it less, which shows up downstream as lower engagement and weaker recall, not as a complaint you can trace back to a sentence.
The fix isn't to ban AI assistance in your workflow, most teams can't and shouldn't. It's to check what comes out the other end before it goes live: does this still sound like a person at this company wrote it, with this company's actual opinions and actual level of directness, or does it sound like the assistant's default voice wearing your logo?
What NeverPrompted actually does with a draft
You paste in the post, the landing page copy, the email. Everything runs in your browser, nothing is uploaded anywhere, which matters when the draft is unpublished brand material. NeverPrompted checks for a statistical AI watermark if one happens to be present, and separately measures how far the draft's style sits from a reference corpus of real contemporary writing: sentence-length variance, vocabulary choices, structural patterns like uniform paragraph shapes or the same three connective phrases doing all the work.
Where it finds specific AI tells, stock transitions, em dashes doing the job a period should do, hedging that adds words without adding meaning, it proposes a rewrite and shows you a diff, so you accept or reject changes rather than getting a black-box replacement. Heavier settings change more of the original wording; lighter settings leave more of your draft intact and just clean up the obvious tells. Either way it's editing your writing, not generating new writing, so your actual claims and structure stay yours.
Where this fits in a content workflow
The natural place to run it is right before a draft leaves the writer's hands, whether the writer used AI assistance heavily, lightly, or not at all (assistant-flavoured phrasing creeps into human drafts too, from reading too much of it). Free gets you unlimited on-device checking plus a weekly budget of rewriting, which covers most individual contributors. Pro removes the weekly limit and adds a better on-device rewriting model, plus a dated PDF report if you need to show an editor or client what changed and why.
One honest limit worth stating plainly: if a piece of writing carries an undisclosed AI provenance watermark from another vendor, no tool, including this one, can guarantee defeating it, because nobody outside that vendor holds the key it was applied with. NeverPrompted checks for a watermark's statistical signature and reduces detectable AI-style evidence in the wording itself; it doesn't promise to beat a system it has no access to.
Wherever this page describes a result: a detected mark is not proof of authorship, and an absent mark is not proof of human authorship. NeverPrompted's on-device rewrite can reduce detectable evidence but cannot guarantee defeating a vendor's undisclosed watermark, on any tier.