Statistical detectors like GPTZero do not judge writing quality. They measure how predictable each word is given what came before (perplexity), and how much that predictability varies across the text (burstiness). AI text is smooth and even, so it is predictable, and so it is caught. A Korean government press release in English, with awkward grammar and padding, broke every style rule we had and still scored Human 100%.
One conclusion follows. Swapping vocabulary, mixing sentence lengths and fixing punctuation helps human readers but does not move the verdict. We tried eleven registers; all eleven came back AI 100%. Many services sell exactly that.
We pick human writing from before November 2022 whose subject is close to yours. We keep its sentence skeleton (function words, clause order, connectives, punctuation, paragraphing) and at least 80% of its words, and change only what the subject forces. No sentence of your draft survives. Where a donor sentence is simply true of your subject, it goes in as a cited quotation; a thesis is entitled to quote primary sources.
Genre decides the donor. Narrative from 2021 encyclopedia paragraphs, rules and data from official documents, argument and evaluation from briefings or essays. With the wrong genre, even a strictly kept skeleton came back AI 100%.
Long documents are built in 600–1,000-word sections, each scored, then joined. The document verdict is roughly a word-weighted average, so a weak section shows immediately. Under 90 we swap the donor and run once more.
No invisible Unicode characters, no deliberate typos, no homoglyph substitution. They break with the next detector update, a human reader spots them, and they are dishonest. We do not change facts, cite studies that do not exist, or invent experiences. Numbers we cannot verify are marked [verify].
The output reads human, but the prose is ordinary. The borrowed structure shows: a thesis sounds like an official document, an essay like a particular essayist. A slightly awkward phrase may remain, and that is part of why it passes. Discussion and conclusion chapters, being abstract argument, are the floor of the method at 80–89%. Check numbers and citations before you submit.
