A client's AI detector flagged my human-written marketing copy. What can I send them as proof?

Treat it the way a photo desk treats a disputed picture: you do not argue about what the image looks like, you show where it came from. The proof is the provenance trail, not another score.

For text that means the document's own history. Google Docs keeps version history; it shows the file growing over your four days, with timestamps. Add the brief, the call notes, the two revision emails. Lay them out in date order on one page. A marketing manager understands a timeline. She does not understand a classifier, and neither does her boss.

From the other side of the table, Lena, what you describe is the evidence an integrity office accepts when a student is referred: work that visibly developed over time, consistency with the person's other writing, and the ability to discuss the choices made in it. It is also, in effect, what your client has initiated: an informal misconduct process, without a standard of proof, a right of reply, or an understanding of the instrument.

I would supply exactly what Marco lists, with one addition: a short cover note stating in two sentences what the attached material shows and what a detector score does and does not establish. Offer a fifteen-minute call in which you walk her through two or three deliberate choices in the copy and why you made them. In my experience that conversation ends most referrals faster than any document.

On why this happened, since it may help you write the cover note. Marketing copy is written to be predictable. Short declarative sentences, parallel structure, the product name repeated, a call to action in the same place every time. That is the craft. It is also a large part of what text classifiers score: how closely the next word matches what a language model would have chosen. Good copy and model output look alike on that axis, for different reasons.

The three scores you got are the tell. If the tools were measuring a property of the text, they would agree. They are each measuring their own training distribution, and your copy sits in a region where those disagree. The thread on why detectors give different scores covers the mechanics if she wants them.

Thinking about this as a dispute rather than a technical question, because that is what it is. Your client has made a public-ish claim inside her company and now needs a way to reverse it without looking foolish to her boss. Give her that exit. A one-page PDF she can forward unchanged, with the timeline Marco describes on top, Maya's quoted sentence on detectors underneath, and your offer of a call at the bottom. Make it something she can send upward and say "I checked, here is the evidence".

If the engagement runs through a freelance marketplace, read its payment terms before anything else. On the platform I work for, a detector score alone is not grounds to withhold payment for delivered work, and the dispute team will say so in writing if asked. Yours may be similar. That letter is sometimes the whole argument.

One more thing you could add, Lena. Run the copy through the same detector she used, make a handful of small edits, and run it again. If the score drops a lot, that shows it was human writing all along, because a real AI text would stay high.

Teacher, not a copywriter, but the pattern is identical. When our indicator went on, the students who walked out of the meeting were the ones who could open a document and scroll back through a week of drafts. The ones who could not were the kids who wrote the whole thing in one sitting in a notes app and pasted it in. Same work, different evidence.

So whatever tool you write in, write in something that keeps history. It is a terrible reason to choose software. It is also, right now, the reason.

@Priya N. I would leave that out, and not only because it is extra work. A score that moves a lot after small edits shows the score is fragile, nothing more. It would move just as much for model output; that is the entire premise of the paraphrasing tools in the humanizer thread. So the demonstration proves nothing about who wrote it, and it hands the client a reason to distrust any future score in either direction. Fragility cuts both ways, and here it cuts against the point Lena wants to make.

@Maya Lindqvist recording the version history tonight. Brief, call notes, two emails, one paragraph, your quoted line. Nothing else. The contract clause goes into every new agreement from Monday. Should have been there two years ago.