Remove AI watermarks from your text

How AI text watermarking actually works

A text watermark is not a tag attached to your document. It is a pattern hidden in the word choices themselves. When a watermarked model writes, it splits its vocabulary at each step into a pseudorandomly chosen “preferred” group and the rest, then leans slightly towards the preferred group. Do that across a few hundred words and the bias becomes measurable. Anyone holding the seed can run the text back through the same procedure and check whether the preferred words appear more often than chance allows.

The important consequence is that the signal has no life of its own. It exists only in that exact sequence of words. Change the sequence and there is nothing left to detect, which is why rewriting works and why find-and-replace tricks do not.

This is not hypothetical. Anthropic has published details of watermarking in Claude, and Google has marked text from the Gemini app since 2024. If you are drafting with either, the output can carry a mark that survives copy-and-paste.

Why a different model is the point

Rewriting Claude text with Claude leaves you roughly where you started, because the same model draws from the same preferences. Send it to a model from a different provider and that no longer holds. The second model picks its own words with no knowledge of the first model’s seed, so the pattern does not survive the trip. Your meaning does.

This is why the tool asks which model wrote your text. It is not bookkeeping. It decides where your rewrite gets routed, and getting it wrong is the one way to end up with output that still carries the original fingerprint.

Light edits do not clear a watermark

Swapping a few synonyms feels like it should be enough. It is not. Watermarking schemes are built to survive exactly that kind of tampering, because a scheme that broke under light editing would be useless to the people deploying it. If most of your words are unchanged, most of the statistical bias is unchanged too.

That leaves a genuine trade-off rather than a free lunch. Staying close to your original and clearing the signal pull in opposite directions. Light strength is there for cases where the wording matters more than the watermark; if removal is what you are after, medium or heavy is the honest setting.

What this means for SEO

It is worth being straight about the ranking question. Google does not penalise content for being AI-generated, and it has said so plainly: what matters is whether a page is helpful, original and worth someone’s time. The March 2024 update went after scaled content abuse, meaning mass-produced pages with nothing to offer, not provenance. There is no public evidence that any search engine reads AI provider watermarks at all.

So the reason to rewrite is not to dodge a penalty that does not exist. It is that raw AI drafts tend to read the same way: even sentence lengths, stock connectives like “moreover” and “furthermore”, tidy three-item lists, words like “robust” and “seamless” doing no real work. Readers notice, even when they could not name what they are noticing, and they leave. That is what costs you rankings.

Use this as a first pass, then do the part no tool can do for you: add your own data, your own examples, the opinion you actually hold. The rewrite gets you prose that does not announce itself as machine output. What makes the page worth ranking is still yours to supply.