We're building both a detector and a humanizer, so it's fair to ask what stops us from training the detector to wave our own rewrites through. We test Humanize the way we test any other writer.
How we test it
Detect is never trained or adjusted with Humanize's output in mind. When we measure Detect, we run Humanize's rewrites through it alongside text from other writers and other models, and we publish how often they get flagged.
Why we bother
If Detect let our rewrites through, its results on everyone else's writing would be worth less, yours included. You'd have no way of telling whether "No AI signals found" meant there were none, or only that the text came from us.
What Humanize is for
Humanize comes after Detect. It rewrites the passages Detect flagged so they read like you again, and it removes the hidden characters Detect found. How often its rewrites still get flagged is the number we'll publish.
