AI watermarks in text
Remove an AI watermark from text: the statistical one, not the characters
A watermark in AI text today is rarely a symbol you can find and delete. It is statistical, placed in the words the model chose as it wrote.
Two marks wear the same name
-
1. Invisible characters and file metadata
Zero-width spaces, unusual spacing, provenance records left inside a file. They are findable, so they are removable: a free cleaner counts them, deletes them, and the count is yours to check. If this is the whole problem, it is solved, and it is free.
-
2. The statistical mark, which lives in word choice
Anthropic has been marking Claude output since 2 August 2026, in text, code and files. It is placed as the text is generated, in the sequence of words the model chose. There is nothing to search for and nothing to strip; only re-expressing the wording reaches it.
What a character cleaner does to the second one, 26 August 2026
We ran a free cleaner over 180 marked texts on our bench and looked for the mark again. It was found in all 180 before the run and in all 180 after; on the longest texts the detector's confidence moved from 37.60 to 37.59.
Which one is in your draft
Start with where the text came from. Claude output carries the statistical mark from 2 August 2026 onward, whatever happens to the characters afterwards. Copying, pasting and reformatting change nothing about it.
Then look at the characters. Paste the draft into a plain text editor or any free cleaner. If the character count drops, that layer was real and it is now handled. If nothing changes, hidden characters were never the problem.
Do not let light editing settle it. The mark is spread across the whole text rather than sitting in a few phrases, so moving sentences around or trimming a draft leaves it where it was. Re-expressing the wording is what reaches it.
Every remover measured on one bench
We put 8 removers through the same marked sample, every run published, the latest on 12 September 2026. Take the marked sample and test any remover.
Questions people ask
Before you paste anything
Is it possible to remove an AI watermark from text?
The two marks have two different answers. Invisible characters and file metadata come out completely, in a second, and you can confirm that yourself. The statistical mark in word choice is reached only by re-expressing the wording. Any service that gives one answer for both is answering the easy half.
Does this work on ChatGPT or Gemini text?
Text from another assistant can be put to the checker today: take the marked sample, run it through the tool you have in mind and see what comes back.