Why your draft keeps getting flagged as AI, and how to fix it

An author came to us after a publishing platform flagged her manuscript for the third time. The ideas were hers, built over a decade of coaching. She had used AI in the drafting and revised every chapter. The flags kept coming, and each one carried the same sting: a machine reading her book and concluding nobody was home.

Start with what the platforms actually say. Amazon’s Kindle Direct Publishing, the largest self-publishing platform, splits the world in two. Content the tool created is AI-generated and must be disclosed when you publish, even if you edited it afterward. Content you created and then refined with tools is AI-assisted and needs no disclosure (KDP content guidelines). The burden of the distinction is on you.

Meanwhile, the detectors doing the flagging are less reliable than their confidence suggests. OpenAI built its own AI text classifier and withdrew it within months, saying plainly that its accuracy was too low (OpenAI, 2023). When the company that makes the generator cannot reliably detect its output, treat every detector verdict as an opinion.

Disclosure and detection are separate problems. The real question is who did the creating, and whether the page shows evidence of that person.

Because the tells are real, and they are not vocabulary. Experienced readers catch them without software.

Uniform rhythm. Sentences of similar length and identical shape, paragraph after paragraph. Real writers interrupt themselves. Like that.
Hedged symmetry. On one hand and on the other, everywhere. People who have lived their material take sides.
Frictionless generality. Grammatical, agreeable, unoccupied. No specific mornings, no named failures, no numbers that could only come from one person’s ledger.

None of that is fixed by a thesaurus. So skip the popular dead ends: detector-chasing rewrite loops that sand off whatever personality survived, humanizer tools whose entire output is defined by the machine you are escaping, and the panic rewrite from scratch that burns sound ideas because the prose failed.

Our rescue process does not paraphrase machine prose, because paraphrased machine prose is disguised machine prose. We go back to the source, which is you. Hours of recorded interviews on the book’s material. A documented voice model built from the transcripts: your rhythm, your vocabulary, your point of view, written down as rules a drafting pass can follow and an editor can check. Then the book is rebuilt chapter by chapter, your specifics restored, with a logic audit hunting the unsupported claims AI-heavy drafts accumulate.

The durable fix for detection is not evasion. It is presence.

The result reads like you because at every level that matters it is you, and in our experience that is also what quiets the flags. Detectors are imperfect, but they are imperfect in a revealing direction: what they are trying to measure is the absence of a person.

If you have a flagged manuscript, or one that no longer sounds like you, the book underneath is almost always recoverable. We have rebuilt worse. Start a Conversation.

Sources

Amazon Kindle Direct Publishing, content guidelines. The platform distinguishes AI-generated content, which must be disclosed at publication even if edited, from AI-assisted content, which does not require disclosure. https://kdp.amazon.com/en_US/help/topic/G200672390
OpenAI, AI text classifier announcement, 2023. OpenAI withdrew its own AI-written-text classifier months after launch, citing a low rate of accuracy, a caution worth applying to every detection verdict. https://openai.com/blog/new-ai-classifier-for-indicating-ai-written-text