The Claude text watermark is an invisible pattern in the words a supported model chooses. It helps Anthropic estimate whether Claude processed a passage, and it applies worldwide, including India. It does not identify the person who used Claude.
For readers, the important limit is detection: a mark can indicate Claude was involved, but it cannot establish who wrote the ideas. Anthropic offers text detection to eligible organisations in a private preview. The public file checker handles a different kind of mark.
Which Claude models support text watermarks?
Anthropic maintains a model-by-model support table for text watermarks and file credentials. Check the text-watermark column for your exact model; support for file credentials alone does not mean its text is watermarked.
How the Claude text watermark works
Whenever a language model writes, it picks each next word from a range of likely options. Often several words would be equally good: “start” or “begin”, “quick” or “fast”. Normally a random number generator settles those low-stakes choices.
With watermarking on, Claude still makes those choices randomly, but the randomness comes from a secret key combined with the few words that came before. Anyone holding the key can then check a passage and ask how consistent its word choices are with what Claude would pick using that key. Across hundreds of choices, that gives a probability that the text came from Claude. One choice proves nothing; the signal comes from the pattern.
Anthropic says its method builds on SynthID-Text, which Google DeepMind published in Nature in 2024 and which goes back to a 2022 proposal by computer scientist Scott Aaronson. Google has open-sourced SynthID for text and uses it in its Gemini models. A few points follow from the design:
- Nothing is hidden in the characters. There are no invisible Unicode characters, markup or extra tokens. The words themselves are the watermark, so it travels with copy and paste.
- No personal data. Anthropic says nothing in the watermark or its key lets anyone recover information about the user, their organisation or their chats. It marks text as Claude’s, not yours.
- No cost or speed penalty. It produces no extra tokens, so API prices don’t change, and Anthropic says the impact on speed is negligible.
- Every surface, every region. The watermark is applied at the model level, so it appears whether you use the Claude apps, the Claude Platform API, Claude Code, Claude Cowork or Claude Tag. Supported models accessed through AWS, Google Cloud and Microsoft Foundry carry it too.
- No ownership change. Anthropic says watermarking doesn’t change who owns the output or what rights you have to use it.
Text watermarks vs Content Credentials
Claude uses two separate marks. Text gets the statistical watermark above. Files Claude creates, such as PNG or JPEG images, get Content Credentials: signed metadata that follows the C2PA standard and records that Claude processed the file. Nothing inside the file’s pixels changes. Content Credentials are added only where Claude can generate files, which means the Claude apps and the Claude Platform API, including Claude Platform on AWS and Claude in Microsoft Foundry. That includes documents and decks if you use Claude to make docs and slides, where the file format supports it.
Metadata is easy to lose. Converting a file, re-saving it or taking a screenshot can strip it. The same weakness affects C2PA labels on AI images in general, which is why the visual checks in our guide on how to spot AI-generated images still matter.
When the watermark is strong, and when it isn’t
The watermark can only live in words Claude chose freely. The more Claude writes, the more choices it makes and the stronger the signal. Anthropic’s own explanations give a clear picture:
| Type of output | Watermark strength | Why |
|---|---|---|
| Long essays, emails, articles written by Claude | Strong | Hundreds of free word choices |
| Translations by Claude | Present | Claude chooses every word of the translation |
| Short replies | Weak or none | Too few words for a reliable signal |
| Factual passages | Sparse | Few words can change without becoming wrong |
| Light proofreading of your text | Weak or none | Almost all the words are still yours |
| Code | Generally less | Code must be exact; comments and arbitrary names can carry it |
| Claude text lightly edited by you | Probably still present | Most of Claude’s choices survive |
| Claude text completely rewritten | Removed | None of Claude’s word choices remain |
Detection also can’t tell the difference between “Claude wrote this” and “Claude heavily edited this”. Anthropic’s help page spells out the limits. A detected mark means the text may have been processed by Claude, not that Claude came up with the ideas. And the absence of a mark doesn’t prove a human wrote it: the text may come from an older model, be heavily paraphrased, be very short, or come from another AI service altogether.
Who can check for Claude’s watermark
Not you, for now. Text watermark detection is in private preview through an API, for organisations eligible under EU law. Anthropic lists regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations and EU civil society groups, plus enterprises that need to verify watermarking for their own compliance with the AI Act. Organisations can register interest through the Claude Watermark Detector Access Request Form. Anthropic says it plans to widen access but hasn’t given a timeline, and it hasn’t said whether organisations outside the EU qualify.
For files, there’s a free public tool. The Claude Content Checker reads Content Credentials in images, video and audio, including JPG, PNG, WEBP, HEIC, AVIF, MP4, MOV, WAV, MP3 and FLAC, up to 100 MB. The check runs in your browser and the file isn’t uploaded. It doesn’t check text.
How to check a file Claude made
- Open claude.com/check-content in any browser.
- Drop in the original file, not a screenshot or a re-saved copy.
- If a Claude-issued Content Credential is found, the file was processed by Claude. If nothing is found, the file may still be AI-made; the metadata may simply have been stripped.
Why Indian users get it too
Anthropic says it applies watermarking globally because it does not yet have a durable way to restrict it by region. Supported output therefore carries the mark in India too. The reason is Anthropic’s rollout policy.
What it means for students, writers and developers
Students
Essays and assignments drafted or translated by a supported Claude model now carry a mark that an eligible educational organisation could, in principle, detect. Grammar fixes to your own writing leave very little signal. The honest approach is the simplest: follow your institution’s AI policy and declare AI help where it’s required. The same applies to the free AI tools for students and writers that run on other models, which may carry their own marks.
Developers
- You can’t disable it. There’s no setting for Claude API customers to turn the watermark off.
- Your own obligations are separate. If you deploy Claude in your product for EU users, Anthropic says you should independently assess what Article 50 requires of you.
Limitations and open questions
- No public accuracy numbers. Anthropic hasn’t published detection or false positive rates.
- You can’t verify your own text. People whose writing carries the mark can’t check it themselves until detection opens up.
- It isn’t an AI detector. AI detectors such as Pangram look for stylistic patterns and can flag human writing. The watermark only identifies Claude’s own pattern, and only with Anthropic’s key.
- Older models are the gap. Until every model is covered, text from unmarked models looks the same to the detector as human writing.
FAQ
Can I see the watermark in Claude’s text?
No. There are no hidden characters to find. The mark is a statistical pattern in which words were chosen, and it can only be detected with Anthropic’s key.
Does the watermark identify me?
Anthropic says no. It carries no information about the user, organisation or conversation, only that the text came from Claude.
Can I turn off the Claude text watermark?
Anthropic does not provide a user setting to disable watermarking on supported Claude models.
Does it apply in India?
Yes. Anthropic applies the watermark to supported models worldwide, including India.
Will editing Claude’s text remove the watermark?
Light editing probably won’t remove it completely. A complete rewrite, where every word is replaced, will. Short passages may not carry a detectable mark at all.
Does it cost more or slow Claude down?
No. It adds no tokens, so prices don’t change, and Anthropic says the effect on speed is negligible.
Can my college check whether I used Claude?
Only if it has access to Anthropic’s detection API, which is currently a private preview for organisations eligible under EU law, including educational organisations. Even then, a detected mark shows Claude processed the text, not who came up with the ideas.
- SourceClaude Content Checker
