You can usually feel it now. You read a couple of lines and something is off before you can name it. Nobody actually says "delve," but there it sits. The points show up in neat little threes. Every sentence the same polite length as the one before, no rush, no pause. A long dash you would never type by hand. And under all of that, characters you cannot even see: invisible marks that hitch a ride in copied text and quietly rat it out as machine-made.
I got tired of spotting it, so I built something to strip it. Ghostwriter. Free, open source.
Then, instead of trusting my own eye, I checked. A raw AI draft scored 97% machine on ZeroGPT. Same words run back through the tool, scanned again: zero. The detector called it human and moved on.
The part that caught me off guard was that you cannot patch this at the end. The cheap tells, the dashes and the hidden characters, those wipe off in a second. Vocabulary and rhythm are rewrite rules. But the tell that really sticks is not a word. It is the shape of the thing. Whether the last line dutifully explains the point you already made. Whether it names an actual book or waves at "a classic example" and hopes you nod along. And that shape gets decided while you draft, which is the whole reason those one-click "humanize" buttons flop. Click one and the bones are already wrong under the polish.
One warning, because this corner of the internet runs thick with snake oil. Detectors get jumpy on short text, and no tool turns a dull paragraph into a good one. Ghostwriter takes the machine fingerprint off. Whether there is anything worth reading underneath is still your problem, not the tool's.
The long version walks through all three layers, the runs across a few Claude models, and how the whole thing drops into Claude and Copilot as one folder you copy across: read it here
Real question, not a rhetorical one: which tell ruined AI text for you? The one you now catch in every newsletter that hits your inbox.
Top comments (0)