Written by Ankush Gupta

A press release sat in our pipeline because a number on a dashboard would not come down. The copy was accurate. The client had already approved the quote inside it. Our own workflow kept pushing it back because an AI-detection score said a machine had written it. One of our writers reworked the same paragraphs again and again. What finally moved the score was not a better argument. It was shorter sentences.

We built that loop at FameNinja ourselves, in n8n, because clients ask about detection scores and we would rather answer the question before it gets asked. Building it is also how we found out what the thing actually reads.

The score responds to surface, not to thinking

Watch a rewrite loop run enough times and it stops being mysterious. Sentence length changes the number. So does punctuation density. Swap a formal connector for a plain one and it moves again, while the argument inside the paragraph sits exactly where it was, right or wrong, untouched. We have had drafts where the only difference between a failing pass and a clearing pass was rhythm.

That is a useful thing to know about your own copy. It is a bad thing to put a hiring decision on top of.

A rewrite loop trains people to write for the detector

Our writers are not naive. Not long after the loop went live, they had worked out the tells and started writing to them from the first draft. Pass rates climbed. The work did not get better. We had quietly rewarded a skill with no relationship to the job, which is an interesting thing to notice in a content pipeline and an expensive thing to miss in a hiring funnel.

Put the same filter in front of applications and it behaves the same way. It does not find the people who can think. What it finds is the people who already know how to dress text so that it reads as human, and that is an afternoon of practice for anyone who has been shown how.

There is a limit to what I can claim here. We run this loop on marketing copy, not on job applications, and I have not tested a detector against an applicant pool. What I can report is what the tool does to the people who write into it every day, and it is the same class of tool now being sold into hiring stacks.

The applications it removes are not the ones you expect

Our team sits across several time zones and most of us are working in a second or third language. The drafts that came back flagged were rarely the careless ones. They were the ones where somebody had worked hardest to be understood: even sentences, plain vocabulary, no slang, nothing regional, every ambiguity sanded off. Careful writing and machine writing look alike from the outside.

Someone who learned English in a classroom writes in that register by default. So does someone who has been coached to sound professional, and someone who is anxious about how they will be read by a stranger. Screen on the score and those are the applications that leave first. That is not a hiring standard. It is a tax on the people who tried hardest.

We ask what they would change, and why

So we stopped asking who wrote it.

What goes out now is a short piece of copy with something genuinely wrong in it, sent along with the brief it was supposedly written against. We want a revision back, plus a few lines on what they changed and why. It is deliberately small work. We also tell candidates they can use any tool they like, including a model, and we mean it rather than testing whether they will admit it.

Then we read the note instead of the prose. Did they find the thing that was actually broken, or tidy the sentences around it? Did they catch that the brief and the copy were asking for two different outcomes? On a call afterwards we pick one of their changes and ask them to defend it. That conversation is short and it has told us more than any score ever has.

Someone who used a model and can account for every choice in the output is doing the job. Someone who sent back clean prose and cannot say why a single line of it is there is not, and no detector was ever going to separate those two people for us.

The detector still runs in our PR pipeline daily. It is a checkpoint on copy about to ship under a client’s name, in a market where editors ask the question. It flags a draft. A person decides.

That is the whole distinction, and it is worth holding onto. Run the tool on work, where a wrong call costs you a rewrite. Keep it off people, where a wrong call costs someone a job they would have been good at, and nobody ever finds out that it did.

Author Bio:
Ankush Gupta is a Fractional CMO at FameNinja, where he works on online reputation management, digital PR and marketing automation.