A reader sent us a draft last week. “It’s 97 percent AI,” the detector said, “and I wrote half of it myself.” We believed her, because that is how detectors work: they measure the probability shape of text, and humans who edit AI drafts end up with text that walks like the model.

The three tells are consistent across tools. Rhythm: models write sentences of similar length, marching in even time. Vagueness: claims float without numbers, examples or dates. Vocabulary: the same twelve transitions show up everywhere, and “delve” heads the list.

The fix is not a humanizer tool, which swaps synonyms and leaves the shape untouched. The fix is the editing pass in our humanizing guide, and it takes ten minutes. Break the rhythm first: split one long sentence into three words, then let the next run long. Add one specific number per paragraph, from your own data or experience. Delete the borrowed phrases, all of them.

Then read it aloud. This is not superstition; detectors key on statistical smoothness, and human stumbling is statistical noise. When you stumble, you rewrite, and the rewrite is where your voice enters the text. The draft we fixed for that reader scored 11 percent after the pass, and it read better than her original human-written half, which is the actual goal. Detectors are smoke alarms, not judges. The writing was always the point.