TL;DR
- Since August 2026, Anthropic embeds an invisible watermark in text written by Claude. It survives copy-paste and reformatting, but not a complete rewrite.
- The watermark shows that Claude was involved, not that a person didn’t write the text. And light proofreading may leave too little trace to detect.
- Watermark detection is in private preview for eligible organizations. There’s no public tool that can check Claude’s text watermark yet.
- Most AI detectors are much cruder. Some flag a text just because it uses an em dash.
- The real question isn’t how to hide AI. It’s how to build trust in how you use it.
Since August, we know that Claude adds a watermark to the text we generate with it. The questions came immediately. Is this a problem? Should I worry? And if I need to, can I hide it?
I asked myself the same thing. I work on huge projects. The research and the volume of information I analyze are enormous, so of course I use Claude. It does it faster and more efficiently than I could. Does that mean that if someone finds a watermark in my final document, they’ll say it’s fake, that I didn’t do the work?
A close friend who works as an editor asked me a similar question. She was preparing her academic thesis and worried that her university used outdated AI detection methods. Her ideas were clearly her own and easy to verify as human, but a detector could still flag them as AI-generated, and her work would lose its value.
So let’s look at how this actually works, whether it should worry you, and what I think matters more.
How Claude’s watermark works
Anthropic announced the watermark in August 2026, as part of complying with the EU AI Act and the new EU Code of Practice on transparency of AI-generated content. It applies worldwide, across Claude, Claude Code, the API and Anthropic’s other products. No exceptions by region, no exceptions by plan.
The idea is elegant. When Claude writes, it constantly chooses between words that mean roughly the same thing, like “overcast” or “grey”. Normally that choice is random. With the watermark, a secret key slightly steers the pattern of those choices. In Anthropic’s words, “the words that Claude picks are still random, but now, one can check the sequence of words and see if it’s consistent with the choices Claude would make if it was using the key.”
The method is called SynthID-Text and comes from Google DeepMind’s research published in Nature in 2024. It’s related to older academic approaches, such as the “green list and red list” method, where the system counts how often a text picks words from a secret list.
The text doesn’t look any different. There’s no hidden label that someone can unlock. Only an algorithm with the secret key can count the choices and say: yes, Claude was here.
Does it survive copying and editing?
According to Anthropic:
- it survives copy-paste, reformatting and conversion to plain text;
- light editing keeps it, but a complete rewrite removes it;
- a translation produced by Claude also carries a watermark;
- it’s weaker in factual passages with few word choices, and almost absent in code;
- proofreading edits may be too sparse to detect.
Anthropic itself is clear that this is a step forward, not an unbeatable system.
Should you worry?
This is a deep, almost philosophical question. I won’t answer it for you, but here are a few things I think everyone should keep in mind.
The watermark shows involvement, not authorship. Anthropic says it can only indicate that “Claude was likely involved with the content at some point.” It can’t tell whether a human wrote the text, and it can’t tell creation from heavy editing.
AI learned from us. Language models are trained on human language. So what does it mean when a detector says I write “like AI”, when AI learned to write from people like me? It’s a bit like the thief shouting “stop, thief!”
Not being flagged doesn’t mean a human wrote it. If a detector doesn’t catch a text, that proves nothing either.
So, should you avoid AI detection? It depends. It depends on the context and on why it matters to you. For some people, it’s important to their image that everything they publish is 100% their own work. Others openly use AI and even encourage it. And in teams with a high level of trust, the fact that a document passed through AI doesn’t reduce the value of the work.
That’s why I believe the most valuable currency of the future will be trust. You trust me that my using AI tools doesn’t mean I signed a document without reading it.
There’s also the opposite risk, which I find even more interesting: AI slop inside companies. I make a document and tell AI to “make it sound smart”. I don’t read it. I send it to a colleague, who runs it through their own AI and passes it on. Nobody human checks it, and eventually something absurd reaches a decision. The problem there isn’t the watermark. It’s that nobody takes responsibility.
Who can detect it today?
Anthropic’s watermark detection is currently in private preview for eligible organizations, such as regulators, researchers, media and educators, as required under EU law. So no public tool can check Claude’s text watermark yet. Anything claiming otherwise is speculation.
That doesn’t mean AI detectors don’t work. They just use other methods: they look for patterns in the text. And many of them are surprisingly crude.
I know this personally. When I worked in media, my editor-in-chief taught me to use the em dash properly. I set up all my devices to turn a hyphen into an em dash automatically. Since ChatGPT came out, people think I use AI for every text I write.
Images are marked differently
For images and files, the method is different. It follows C2PA, an open standard for content provenance founded in 2021 by Adobe, Arm, BBC, Intel, Microsoft and Truepic. OpenAI and other AI companies later joined.
The idea is that every digital file carries a history, like a passport: where it came from and what it went through. Imagine you take a beautiful photo, edit it in Photoshop, and enter a competition that only accepts unedited shots. With this passport, the edit can be traced.
That’s also why photos you edit with ChatGPT and upload to Instagram or Facebook sometimes get an “AI info” label, even when the photo itself is real. Meta reads these industry signals, the C2PA and IPTC metadata that tools from OpenAI, Google, Microsoft, Adobe and others add to files.
Don’t write to beat detectors. Write like yourself.
There are guides online on how to get past AI detectors. I haven’t used any of them, and I wouldn’t recommend downloading files from strangers.
But the lists of patterns they describe are useful for a different reason: they’re exactly what makes text sound generic. Grandiose statements. The “it’s not this, it’s that” structure. The same rhythm in every paragraph. The same overused words.
Remove them because they make your writing worse, not to fool a detector. Put them in your instructions to AI as words and structures to avoid, and translate them into your own language if you write in it. Then add what no model has: your opinion, your examples, your experience.
What this means for us
Here’s my honest disclosure: I write with AI too. The research, the structure and many of the drafts for my content, including this article, are done with Claude, then reviewed, corrected and signed by me. The ideas, the opinions and the responsibility are mine.
I think that’s the only sustainable position. Instead of chasing every rumor on social media that “Claude is now banned” or “this is the one right way to use ChatGPT”, every organization should agree on its own rules: when AI use is fine, for what, with what limits, what never goes into AI, and who reviews the result.
Whatever tools come next, what matters is that we trust the person behind them.
About The Author
