How can you tell if ad copy was written by AI?

Recognisable openers such as 'Elevate your' and 'Look no further', the 'not just X, it's Y' construction, three adjectives in a row, and body copy with no contractions at all.

Can AI detect its own writing?

Not reliably. Asking a language model whether text sounds machine-written tends to produce agreement rather than judgement. Counted signals — specific phrases, contractions, repeated openers — are more dependable than asking.

How to stop your ad copy reading like AI wrote it

A feed is a place where people are already looking for a reason to scroll past. "Elevate your workflow with our cutting-edge solution" hands them one in six words.

This matters more in an ad than almost anywhere else. A blog post that reads as generated is just dull. An ad that reads as generated is skipped before it is read, and you are charged for the impression either way.

The tells, in order of how badly they cost you

Opener phrases. "Elevate your", "Unlock the", "Look no further", "In today's fast-paced". That isn't weak writing, it's a signature — and readers have learned it.
The not-just-X-it's-Y shape. "It's not just a CRM, it's a growth engine." The single most recognisable generated-copy construction there is.
Three adjectives in a row. "Fast, simple, and powerful." The rhythm gives it away before the meaning lands.
No contractions anywhere. Body copy with none reads as a press release. Nobody speaks that way, and formal register in a feed reads as corporate rather than authoritative.
Every cell opening on the same word. As a set this reads as generated even when no single line does.

Why you cannot ask a model to check this

Asking a language model whether text sounds like a language model is asking the defendant to sit on the jury. It'll agree the copy is excellent, because agreeing is what it does. Every check worth running here is a count: does this phrase appear, are there contractions, do all four headlines start with the same word. Counts don't flatter.

What we found in our own output

Running that check across 38 generated campaigns: 27 came back clean, 7 read stiffly, and 4 were recognisably generated. Roughly a quarter needed a rewrite before they were worth paying to show anyone. We publish it because a tool claiming a perfect record on its own output isn't telling you about its checking — it's telling you about its marketing.

What this does not tell you

Whether the copy is any good. It finds machine register, not weak positioning, and a page of plain human sentences can still be a bad ad.