Text from Claude carries an invisible watermark. It reveals AI, not you.
Starting November 3, 2026, Anthropic watermarks text from older Claude models too. Newer models have done so since August. You won't see the watermark, and not a single character is added to the text. Here is how it works, why it exists, and what it means for your company, your website and your social media.
The watermark is in the choice of words, not in the characters
When a language model writes, it does not pick words in a fixed way. For each next word it has several options that fit almost equally well, and it picks one at random. That is why you never get exactly the same answer to the same question twice.
Anthropic replaced that randomness with a secret key. The key, together with the words written so far, decides which of the equal options the model picks. A single word reveals nothing. Over a longer text, though, a statistical pattern emerges that can be measured if you have the key.
With a person
Everyone has favorite expressions. You can't tell the author from one sentence, but from a full page you often can.
With AI
A watermarked model picks among equal words according to the key. One sentence tells you nothing, a longer text does.
No hidden character is added, and the token count and price stay the same. Anthropic says the watermark does not change the meaning, quality or readability of responses. Google marks text in a similar way with SynthID.
Where and since when Claude adds it
- Models released since August 2, 2026, such as Opus 5.5, Sonnet 5.5 and Haiku 5.5, have it from launch.
- From November 3, 2026, Opus 4.6, Opus 4.7, Sonnet 4.6 and Mythos 5 get it too.
- It applies in the Claude app, in Claude Code and through the API, including AWS, Google Cloud and Microsoft Foundry.
For Opus 4.6 and Sonnet 4.6, the watermark is added only with default settings. If a developer changes the parameters that control how random the answers are (temperature, top_p or top_k), the text may not carry it. For an everyday user this makes no difference. For a company that needs the watermark for its own compliance, it does.
Files that Claude creates, such as PNG and JPEG images, also get signed metadata based on the C2PA standard. Anyone can check those in the free Claude Content Checker.
Why they do it: the EU AI Act
Article 50 of the EU Artificial Intelligence Act (EU AI Act) requires AI providers to mark generated text, images, audio and video so that a computer can recognize them. Models placed on the market from August 2, 2026 must do so right away, older ones have until December 2, 2026.
Anthropic adds the watermark wherever Claude runs, not only in the European Union. Breaking Article 50 can lead to a fine of up to 15 million euros or 3% of worldwide annual turnover.
What the watermark reveals and what it doesn't
The watermark contains nothing about you, your company or your conversations. It only says that the text probably went through Claude.
- A detected watermark doesn't mean AI wrote the whole text. Claude may have only edited, translated or summarized it.
- A missing watermark doesn't mean a person wrote the text. Older models, other providers or a short text don't carry it.
- The watermark survives copy and paste. Heavy editing, translation, rewording or mixing with other text weakens or removes it.
Who can read it
Only someone with the key can check the watermark, which means Anthropic. Text detection is still in private testing. Access goes to organizations that EU law requires to verify content: regulators, law enforcement, media, fact-checkers, researchers, schools and companies with similar obligations. Others can apply through a form, and Anthropic plans to widen access over time.
Other AI tools, search engines and social networks can't read the watermark on their own, because they don't have the key.
What it means for companies
If your company uses Claude for emails, proposals or documents, you don't need to change anything. Quality and price stay the same.
If you publish AI text on matters of public interest in the EU, however, Article 50 requires you as the publisher to label it. This has applied since August 2, 2026. The exception is text that a person has genuinely reviewed and for which someone holds editorial responsibility. A quick glance before publishing does not count as a review.
“A city office publishes a notice written by AI that nobody edited. It should state that AI created it. If a clerk reads it, corrects it and signs it, it is a text with a responsible author.”
If you build your own product on top of Claude and offer it to customers, you may count as a provider under the law yourself. In that case the marking obligation applies to you directly. This is not legal advice, so check your case with a lawyer.
What it means for your website and social media
Nothing changes for the reader. The watermark is invisible and common tools can't read it.
Something else changes: text you simply copy from AI carries a trace. Text to which you add your own experience, numbers and real examples is yours. The watermark fades with every edit. And above all, it is a better text.
Copied text
The same style as thousands of other posts, no experience of your own, and a Claude trace in every paragraph.
Text with your own work
AI helps with structure and wording. The facts, examples and opinion are yours.
Bottom line
The watermark is not a reason to stop writing with AI. It is a reason to stop publishing AI text without your own work.
Summary
| QUESTION | ANSWER |
|---|---|
| Since when | Newer models since 8/2/2026, older ones since 11/3/2026 |
| Where it is | In the choice of words, not in the characters |
| What it reveals | That the text went through Claude, nothing about you |
| Who can read it | Anthropic and selected organizations in private testing |
| What removes it | Heavy editing, translation, rewording, short text |
| What you should do | Nothing for everyday use. Review or label text published in the public interest |
