Claude has started quietly watermarking what it touches. Is your content affected? Right now there is no way to find out, and that includes me.
Anthropic confirmed on August 14, 2026 that Claude embeds an invisible statistical watermark in its text output. The trigger is the EU AI Act’s transparency rules, which took effect on August 2. The mark is applied globally, not only to EU traffic.
Read the wording carefully, because most coverage did not. Anthropic says “future Claude models will generate text that contains a watermark,” and that models launched before August 2 sit in a transition period: “we’re working to add watermarking for those models as well.” So this is a rollout, not a switch that was thrown. Whether the specific model behind your pipeline is marked today is not something Anthropic has published, and not something you can test.
A watermark removal tool is already past 15,000 GitHub stars. It does not remove the watermark. More on that below, because it is the part everyone got wrong.
What a statistical watermark actually is
Forget hidden characters. Nothing is added to the text.
When a model picks the next word, there are usually several choices that mean the same thing. Fast, quick, rapid. A statistical watermark biases that pick: among candidates that are all fine, it slightly favours the ones matching a pseudo-random pattern derived from a secret key plus the words already written.
Anthropic’s own description is the clearest one I have read: “the watermark only changes the source of the randomness used to pick among words.”
The text is the watermark. That is why it survives a copy-paste through a plain text editor, which kills the zero-width character tricks people used to worry about. And it is why you cannot find it by looking. There is no string to search for. Every individual word is a word a human might have picked.
The good
It is honest engineering, and it solves a real problem.
The mark is invisible, costs the reader nothing, and does not degrade the writing. It rides on choices that were arbitrary anyway. Anthropic uses the SynthID-Text approach that Google DeepMind published in Nature in 2024 and open sourced, rather than inventing a private scheme nobody can inspect.
For files rather than text, Claude attaches signed C2PA provenance metadata, the same open standard camera makers use. Different mechanism, same goal, and that one is visible if you go looking.
No user information is embedded. The mark says “Claude touched this.” It does not say who asked.
The bad
Three limits, and the third is the one that will bite content people.
Nobody outside Anthropic can detect it. Detection needs the key. Anthropic says a detection API is coming “soon” with no date attached. Third-party AI detectors cannot see it at all; tools like Pangram use unrelated statistical methods on text patterns. So today the watermark exists, and its only reader is the company that wrote it.
It proves processing, not authorship. Anthropic is explicit: it “cannot distinguish ‘Claude wrote this’ from ‘Claude heavily edited this.’” Run your own paragraph through Claude for a grammar pass and the output carries the mark. A positive result is not evidence of fraud, or of a policy breach, or that a human did not write the substance.
It is weakest exactly where your money is. Detection “performs poorly on small samples,” and the signal is thinner on factual writing because there are fewer word choices that do not damage accuracy. Product descriptions, spec pages, pricing tables, short meta descriptions: the commercial pages are the ones carrying the least signal.
What people will read into a hit: that no human wrote this, that it is untrue, that a policy was broken, or that you can tell which account produced it.
What it actually proves: that Claude was involved in producing this text at some point. Nothing beyond that.
The ugly
Here is where it gets messy, and the mess is instructive.
Google’s own staff got it wrong for two weeks. On August 5 a Google representative told developers on the official AI forum that Gemini API text output was not SynthID watermarked. On August 19, staff corrected it: “text generated via the Gemini API IS actually SynthID-watermarked!” If the vendor cannot answer this reliably, treat every confident blog post on the subject with suspicion, including the ones I read while checking this. This is the same failure I keep hitting from the other direction: documentation stated confidently is not evidence, and the fix is always to go and check the primary source.
The removal tool does not remove it. watermarks-remover appeared on August 11, three days before Anthropic published the technical detail, and is past 15,000 stars. Read its own README and it classifies statistical text watermarks as “Layer B,” handled by “Agent rewrite.” That is not stripping. That is paying a second model to paraphrase your text until the pattern is gone, which is the looks-right-but-isn’t pattern with a star count attached.
Which is exactly what Anthropic said would work, and what it costs: “Light editing probably won’t remove the watermark completely; a complete rewrite where every word is replaced will.”
The compliant get marked and nobody else does. ChatGPT still ships no text watermark. OpenAI marks supported images with both C2PA and SynthID, and supported audio with SynthID alone, but text, the highest-volume output and the one that moves most freely between containers, is unmarked. A rule that binds Claude and Gemini and skips the largest text generator is not a provenance regime. It is a tax on the two vendors who complied.
Where the three actually stand
| Text watermark | Files / media | Can you detect it? | |
|---|---|---|---|
| Claude | Rolling out (SynthID-Text). New models at launch, older ones in transition | C2PA signed metadata | No. API “coming soon”, no date |
| Gemini | Yes, since 2024 (SynthID-Text) | SynthID + C2PA | Method is open source; detecting Google’s output still needs Google’s key |
| ChatGPT | No | SynthID + C2PA on images and voice | Not applicable for text |
One nuance worth holding onto, because most coverage flattens it. Google open sourcing SynthID-Text does not mean you can check Gemini’s output. The published detector ships with a demo model for testing the workflow. Detecting real Gemini text still needs Google’s key. The algorithm is public and the key is not, which is the only arrangement that works: publish the key and anyone can strip the mark.
How much mark you are actually carrying
It depends on one thing: how many of the words Claude picked.
Run your own paragraph through for a grammar pass and the mark can only live in the handful of corrections it made, which Anthropic says “might be too few to register.” Ask for a first draft, a summary, an outline rewritten into prose, and every word is a choice Claude made. That is the dense end. Anthropic names translation as the clearest case, “because in this case every word is chosen by Claude,” but the principle covers anything generated rather than edited.
So the question is not “did I use AI.” It is “whose word choices are these,” and most content sits somewhere in the middle.
I did nothing about my own pages, and the reasoning holds for anyone running content through Claude. The only removal method is a full rewrite. That means destroying tone and terminology work you paid for, to erase a mark no third party can currently read, to solve a problem nobody can currently observe. If a detector ships and it ever matters, removal becomes a tool you run at that point, with something real to verify against. Doing it now buys nothing and costs the work.
Related: AI Trusts Your Docs. That’s the Problem. · Don’t Assume AI Fixes Things Properly · Why Claude Code Ignores Your CLAUDE.md
Sources
- Anthropic, How Claude’s text watermarking works, August 14, 2026
- TechCrunch, Anthropic says it will watermark text generated by its AI models, August 11, 2026
- Google DeepMind, synthid-text (open source, Nature 2024)
- Google AI Developers Forum, Does Gemini API text output carry SynthID watermarking? (staff correction, August 19, 2026)
- Search Engine Journal, Anthropic Reveals What The Watermark Is And How It Can Be Defeated
- guillaumemeyer/watermarks-remover (MIT, created August 11, 2026)