Does removing AI watermarks make text undetectable?
No. Detectors judge writing patterns, not bytes. Statistical watermarks like Google's SynthID Text live in word choice, where no character cleaner can reach them. Anything promising "undetectable" by stripping hidden characters is selling you something that does not work.
There is a booming category of tools that promise to make AI text undetectable by scrubbing hidden watermarks out of it. Since text cleaning is a thing we are building, you might expect the same pitch here.
You are not going to get it, because the pitch is false. Here is the honest version.
Detectors do not look for hidden characters
AI-text detectors judge writing. They model statistical patterns: how predictable each next word is, how uniform sentence rhythm is, how vocabulary distributes. That is the fingerprint of how models write, not of what invisible characters ride along. Strip every zero-width space and narrow no-break space from a generated essay and its word-choice statistics have not moved a millimetre. A detector that flagged it before will flag it after.
The deliberate watermarking schemes in research, and in production systems like Google's SynthID Text, live in that same statistical layer. They bias which ordinary words get chosen, invisibly to readers, detectable only with the right key. Character cleaning cannot remove a signal that was never in the characters, and no character cleaner touches it. That includes anything we ship.
Meanwhile the detectors themselves are unreliable in both directions. They flag human writing, with non-native English speakers hit hardest, and they miss edited machine writing. Building anything on "will it pass a detector" is building on sand.
So why clean AI text at all?
Because the invisible layer causes real problems that have nothing to do with detection.
- It breaks things. Invisible characters break search, dedupe, validation and code. A narrow no-break space in a config file is a bug you cannot see.
- It reads as pasted. Em dashes, curly quotes and ellipsis characters make text look machine-drafted to human readers. That is a social problem rather than a detector problem, and cleaning genuinely fixes that one.
- It is not yours. Text you publish under your name should contain characters you chose. Typography you did not pick is someone else's accent in your mouth.
- Privacy. Some invisible characters exist specifically to fingerprint text, to mark which copy leaked. Removing characters whose only job is tracking you is basic hygiene, the same reflex as stripping tracking parameters from links.
Every one of those is about quality, correctness and privacy for text you have the right to use, and none of them requires fooling anyone.
The line we draw
If your goal is to pass off machine writing as your own where that is cheating, a graded essay or an application that forbids it, a character cleaner will not save you, and we would not help if it could. Detectors judge patterns. Only rewriting in your own voice changes patterns.
If your goal is what most people actually want, AI-assisted drafts that paste clean, read as deliberate, do not crash parsers and do not carry trackers, then cleaning is exactly the right tool. It should happen automatically on your own machine rather than through a stranger's web form.
Honest tools should make honest claims. This one will clean your clipboard. It will not launder your prose, and nothing does.
Text cleaning is coming next to tdy.
tdy is a Mac menu-bar app that already tidies every link you copy, with nothing uploaded. Invisible characters and AI typography are the next upgrade. One email when it ships, nothing else.
Join the launch list