An AI detector flagged my draft 97% machine. The rewrite scored 0%.

Paste something straight out of ChatGPT into an email, and half your readers can tell. Not because the facts are wrong. Because of how it reads.

The word "delve" where nobody says delve. Tidy lists of exactly three things. Every sentence the same comfortable length as the last one. The long dash almost nobody types by hand. And the characters you cannot even see, invisible marks that ride along in the text and quietly flag it as machine-made.

So I built Ghostwriter, a free, open-source tool that strips those tells. Then I measured it instead of guessing. I ran a raw AI blog draft through ZeroGPT, one of the popular detectors. It came back 97% machine-written. I rewrote the same content with the tool and ran it again: 0%. The detector called it human.

This is also why the "humanize my text" buttons never really work. The tells sit at different depths. Hidden characters and long dashes are trivial to delete. The words and the rhythm take rewrite rules. But the deepest tell is not a word at all. It is how the whole thing is built: whether it spells out its own point at the end, whether every thread ties up too neatly, whether it names a real example or waves at "a well-known case."

That shape gets baked in while you draft. So by the time you hit a button to fix it, the structure is already there, and the button is just painting over cracks. The fix has to happen while you write, not after.

One caveat, because this topic attracts snake oil. Detectors are noisy, especially on short posts, and no tool turns bad writing into good writing. Ghostwriter pulls off the machine fingerprint. It does not hand you something worth saying. That part is still yours.

The full write-up has all three layers, the detector runs across three different Claude models, and how the thing installs into Claude and Copilot by copying a single folder: read it here

So, honest question. What was the first AI tell you learned to spot, the one you cannot un-see now?