How to remove the AI watermark from text

AI writing tools increasingly stamp their output with a watermark so it can be identified as machine-generated. This guide explains the two kinds of AI text watermark — statistical ones like Google’s SynthID and Claude’s, and invisible-character ones — and exactly how to remove both while keeping your meaning. You can do all of it free with Unmark.

The two kinds of AI text watermark

Removing one does nothing about the other, so it helps to know which you’re dealing with.

1. Statistical watermarks (SynthID, Claude)

A statistical, or token-level, watermark biases which words the model chooses, creating a pattern spread across the whole passage that a matching detector can measure. Google’s SynthID-Text pioneered the approach, and in 2026 Anthropic announced that future Claude models will watermark their output with a SynthID-style method — applied by default, with no opt-out, to help comply with the EU AI Act. Because the signal lives in the word choices rather than any visible mark, deleting a sentence or swapping a few words rarely clears it. The reliable fix is to rewrite the text enough that the token sequence no longer matches.

2. Invisible-character watermarks

The second kind hides invisible Unicode between the letters — zero-width spaces and joiners, byte-order marks, bidirectional controls, tag characters and look-alike glyphs. They show as nothing on screen but survive copy-paste, so they can fingerprint or track a document. Removing them is deterministic: scan the text and strip those code points (while keeping load-bearing ones like emoji sequences), which leaves the readable text identical.

How to remove an AI watermark (step by step)

1. Paste your AI text. Copy the output from ChatGPT, Claude, Gemini, Copilot or anywhere else into the box on the Unmark home page. It scans for invisible characters instantly, in your browser.
2. Choose a strength and remove the watermark. A light setting paraphrases; a heavier one rewrites more aggressively. Unmark rewrites the wording to disrupt the statistical/SynthID watermark and strips invisible characters in the same step.
3. Copy the clean text. The result reads the same to a human but no longer carries the machine fingerprint. Re-read it to confirm it kept your meaning.

Does removing the watermark make text undetectable?

Rewriting substantially degrades statistical watermarks and AI-detector signals, and the more you rewrite, the more it degrades. But be realistic: no tool can guarantee a text is permanently undetectable against every current and future detector. Treat watermark removal as best-effort. For invisible characters the story is simpler — they are removed deterministically and completely.

Is Claude’s new watermark removable?

Yes. Anthropic’s watermark is a statistical one, and Anthropic itself notes that light editing may leave it intact while a complete rewrite removes it — which is exactly the rewrite Unmark performs. The same approach applies to SynthID and any other token-level scheme.

FAQ