Does ChatGPT Add Watermarks to Text?
Not every unusual character is an AI watermark, and deleting Unicode cannot guarantee removal of a model-level statistical watermark.
People use the word watermark for several different things. That creates confusion when a text cleaner claims to remove all of them with one click.
Invisible characters are not proof of authorship
Copied text can include zero-width spaces, non-breaking spaces, soft hyphens, byte-order marks and directional controls. These characters may come from websites, PDF extraction, document editors, localisation systems or deliberate obfuscation. Their presence does not prove that a language model wrote the text.
Statistical watermarks work differently
A model-level text watermark can influence token choices during generation. The signal is distributed across word selection rather than stored as a hidden character. Replacing punctuation or deleting zero-width spaces does not reliably remove that kind of signal.
What a cleaner can truthfully do
A Unicode cleaner can reveal and remove invisible code points, normalise spaces and make copied text more predictable. It can fix real formatting and validation problems. It should not promise detector bypass or claim that every hidden character is an AI marker.
Use the Unicode Text Inspector to see the exact code points in a sample before deciding what to remove.