Can AI text watermarks be removed? What works and what does not
There are two very different things people mean by a text watermark. One can be stripped in seconds. The other cannot be removed by a cleaning tool. Here is which is which, with sources.
Nothing leaves this page. No upload, no account, no tracking.
Type 1: hidden characters
Some marks are invisible Unicode characters placed in the text. A scan finds them and a cleaner removes them. CtrAltRemove does this, shows each one in place, and decodes messages hidden in them.
Type 2: statistical watermarks
Other watermarks live in word choice. The model leans slightly toward certain words, and a detector measures the pattern. There is nothing to strip, because the text itself carries the pattern. Google documents limits of its own SynthID Text watermark (Google SynthID Text limitations), and academic work studies how rewriting affects detection (SynGuard, arXiv 2508.20228).
What OpenAI has said
OpenAI announced a word-choice watermark for ChatGPT and Codex in the EU, with its detector limited to approved researchers, and reported that swapping about a quarter of words in a 400-token passage cut detection sharply (OpenAI on EU text provenance). Rewriting weakens a statistical mark. No cleaner removes it.
What this tool will and will not do
It removes hidden characters and lists machine-sounding patterns for you to rewrite. It does not claim to defeat detectors and it does not give an AI score.
Common questions
Can a tool remove a SynthID text watermark?
No cleaning tool removes it. It is carried by word choice, not by hidden characters. Rewriting the text can weaken the pattern.
What can CtrAltRemove remove from text?
Zero-width and other invisible characters, look-alike spaces and Unicode tag characters.
Does it beat AI detectors?
No, and it does not claim to. It reports characters and patterns, with no AI score.
Is my text uploaded?
No. Everything runs in your browser.
Sources: Google SynthID Text limitations · SynGuard paper (arXiv 2508.20228) · OpenAI on EU text provenance
More: Remove zero-width spaces and other hidden characters from text · Remove C2PA Content Credentials and AI metadata from images · Check text from ChatGPT and other AI chats for hidden characters · Instagram and Facebook AI labels: what is in your image's metadata