Guides › What changed

AI text watermarks, explained: what Claude's mark can and can't tell anyone

Updated Oct 31, 2026 · Level: Beginner · Facts from Anthropic's own pages, unless marked otherwise.

What's reported for Nov 3

Anthropic began marking AI-generated text to comply with the EU AI Act. Its rules on marking AI content have applied since Aug 2, 2026. Claude models launched since then, such as Opus 5.5, carry the watermark from day one. ^3,1

Anthropic is now adding the same watermark to older models. According to an email Anthropic sent API customers (organizations that use these models), published by consultant Pasquale Pillitteri on Oct 28 and reported by MACCUS, from Nov 3, 2026 the watermark applies to Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 on the Claude Platform (Anthropic's API). The email says Amazon Bedrock may take a few extra days. ^2,4

Other people who say they got the email have posted about it too. One, developer Alessio Ragni, quoted its subject line: "[No Action Required:] Text watermarking begins November 3, 2026 on some models." He also says that for Opus 4.6 and Sonnet 4.6 it applies only when the model's sampling settings are left at their defaults. ⁵

The email reportedly went only to organizations that use these models through Anthropic's API, so if you only use the Claude app or website, you probably didn't get it.

Not confirmed by Anthropic's own pages yet. As of Oct 31, we couldn't find the Nov 3 date anywhere on Anthropic's own pages. Its help page (last updated Oct 28) still doesn't show a text watermark for these four models, only Content Credentials for files. It says Anthropic is "working to add marking support" to models released before Aug 2, 2026, without a date. Anthropic's news post, its release notes and its official X accounts don't mention Nov 3 either. So treat the date as reported, not confirmed. We'll check again on Nov 3 and update this page. ^1,3

Anthropic applies the watermark worldwide, not only in the EU. It says it doesn't yet have a reliable way to limit it by region. ¹

How the watermark works, in plain words

When Claude writes, it picks one word at a time. Often several words would work equally well: "grey" or "overcast," for example. Normally a random number settles the choice.

With watermarking, that choice is settled by a secret key and the few words before it instead. The text reads the same, but anyone with the key can check whether the word choices fit the pattern. Anthropic says its method is a version of SynthID-Text, a technique Google DeepMind published in 2024. ³

So "invisible watermark" is a bit misleading. Nothing is hidden in the text. There's no secret code to strip out. The mark is in which words were chosen.

What Anthropic says about it^3,1

• Readers can't tell the difference. Anthropic says it found no effect on quality, creativity or readability in its own testing.
• It costs nothing extra. No extra tokens and no change in price.
• It goes wherever the text goes. It survives copy and paste, and may survive some editing. Light edits probably won't remove it; rewriting every word will.
• It's weaker in some text. There's less of it in facts, code and light proofreading, where there are fewer equally good word choices. Very short passages may be too short to check.
• Translations are marked, because Claude chose every word.

Can it be traced back to you?

No, according to Anthropic. The watermark and its key carry nothing about you, your organization or your chats. At most, a check can say how likely it is that Claude was involved in writing a piece of text. ¹

"Claude is going to start snitching"? Not quite

One widely shared post this week put it that way. It's an understandable worry, but it doesn't match how this works:

• Claude isn't reporting anything. The watermark isn't a message sent anywhere. It's a pattern in the text that only shows up if someone runs a check.
• It can't name anyone. There's no user, account or chat information in it. ¹
• Ordinary people can't run the check. Anthropic's text detector is in private preview. It's open to organizations that EU law says should have it: regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations and EU civil society groups. Businesses that need it for their own EU AI Act compliance can also apply. Anthropic says it plans to expand access over time. ¹
• A result isn't a verdict. Anthropic says a detected mark is "not fully conclusive." The text may have been written by a person and only proofread, translated or summarized by Claude, or edited afterwards. Anthropic also says a mark can't tell "Claude wrote this" apart from "Claude heavily edited this." ¹
• No mark doesn't prove a person wrote it. The text may be too short, heavily rewritten, from a model that wasn't watermarked yet, or from another AI.

Is it only Claude?

No. Anthropic says other major AI providers signed the same EU Code of Practice and are adding their own watermarks. Each uses its own key, so Claude's check can't tell whether text came from another AI. ¹

What about images and files?

Different system. When Claude makes a supported file, such as a PNG or JPEG, it adds a signed note in the file's details (a "Content Credential," using the open C2PA standard) saying Claude made it. Anyone can check those with Anthropic's free Claude Content Checker. Taking a screenshot or converting the file can remove that note. ¹

What this means for you

• Using Claude to help you write? Nothing changes in what you see or pay. If you share AI-written text where that matters (school, work, a client), the simplest thing is to follow their rules on AI use.
• Worried someone will "catch" you? The watermark is a likelihood, not proof, and it carries no personal details.
• Building with Claude for EU users? Anthropic says you should check what Article 50 of the EU AI Act requires of your own product. ¹

Anthropic's pages don't mention a way to turn the watermark off.

Common questions

4 quick questions, answered

When does it start?

It already applies to Claude models launched since Aug 2, 2026, like Opus 5.5. Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 are reported to follow on Nov 3, 2026, according to an email Anthropic sent API customers. We couldn't find that date on Anthropic's own pages yet.

Will my answers get worse?

Anthropic says its testing found no effect on quality, and Google DeepMind's study of the same technique found no difference in user ratings.

Can I check my own text?

Not yet for text, unless your organization qualifies for the private preview. Files can be checked with the free Claude Content Checker.

Does it change who owns the text?

No. Anthropic says the watermark doesn't affect ownership or your rights under its terms.

Related: Claude plans and prices · Claude's Max and Ultra plans, and what "verified" means · Get the most out of Claude · Using AI safely

New to AI?

AI crash course: start here

What AI chatbots are, how to ask, and how to stay safe. About 10 minutes.

Should I pay for AI?

What free plans really give you, what a paid plan adds, and who needs one.

Which AI model should I use?

Pick the right model and thinking level in 60 seconds.