Anthropic has started adding an invisible watermark to text written by Claude. It arrived in August 2026, it applies everywhere in the world, and there is no way to switch it off.
The announcement caused a good deal of concern, and much of what circulated in the days afterwards was inaccurate, including claims that documents are now traceable back to individual users and that Google will demote any content carrying a mark.
In this guide, we explain what the Claude watermark is, how it works, where it will show up in your day-to-day work, what it means legally for businesses publishing content, and whether other AI platforms are likely to follow.
Key takeaways
- The Claude watermark is an invisible pattern in the words Claude chooses, not a hidden character or a piece of metadata.
- It applies to Claude models released from 2 August 2026 onwards, worldwide.
- It carries no information about you, your company or your prompts.
- Only Anthropic can currently read it. No search engine can.
- The change was made to comply with Article 50 of the EU AI Act.
- Businesses publishing AI-assisted content have a separate legal duty that Anthropic’s watermark does not cover.
- Google, Meta, Microsoft, Mistral and OpenAI signed the same agreement, so similar marking is expected across the market.
What is the Claude content watermark?
The Claude content watermark is an invisible statistical pattern woven into the text Claude generates. It signals that a Claude model was involved in producing or processing the content. It cannot be seen by a reader, it contains no identifying information, and it can only be detected by someone holding Anthropic’s secret key.
Anthropic describes it as marking at the model level, which means the pattern is created as the text is written rather than added afterwards. Nothing is inserted into the text, and there are no invisible characters hiding between the words.
Files are handled differently, because metadata can be attached to a file in a way it cannot be attached to plain text. When Claude produces an image or a graphic file such as a .png, .jpg or .svg, it attaches a content credential, which is a small, cryptographically signed note in the file’s metadata recording that the file was made or processed with Claude, using an open industry standard called C2PA.
| Text watermark | File content credential | |
|---|---|---|
| What it marks | The words themselves | The file’s metadata |
| Visible to readers | No | No |
| Survives copy and paste | Yes | Not applicable |
| Survives format conversion | Yes | Usually lost |
| Who can read it | Anthropic, for now | Any C2PA-aware tool |
How does the Claude watermark work?
Claude writes one word at a time, and at many points in a sentence there are two or more words that would work equally well, which is where the watermark operates. Instead of settling that choice with a random number, Claude settles it using a secret key and the words that came just before.
Anthropic’s own example is the phrase “The weather today was cold and…”. The next word is very unlikely to be “sugary”, but it could reasonably be “overcast” or “grey”. Either word leaves the meaning of the sentence broadly intact.
The choice remains effectively random, but the source of that randomness changes, which makes the resulting pattern reproducible by anyone holding the key.
Detection works much like testing a slightly weighted coin, where four flips tell you nothing about whether the coin is fair and four hundred make the bias obvious. Longer passages therefore produce far more reliable results than short ones.
Two technical points are worth knowing:
- The watermark is applied at the point where words are selected, not inside the model itself. As Sebastian Raschka’s technical breakdown explains, no retraining was required to add it.
- Detection does not require rerunning the model, because the method scores the finished text directly, which keeps the cost of checking low.
Anthropic reports no impact on speed or price, since the process produces no extra words.
Where will the Claude watermark show up?
The watermark appears in longer pieces of original text written by Claude and travels with that text when it is copied and pasted elsewhere. It shows up weakly or not at all in short passages, factual writing, code and lightly proofread human work. Nobody reading the content will notice anything.
Content that carries a strong mark
- Blog posts, articles and reports drafted by Claude from a brief
- Translations produced by Claude, since every word is chosen by the model
- Long-form copy that has been formatted or exported but not rewritten
Content that carries a weak mark or none at all
- Text under roughly 150 words, which falls outside the requirement entirely
- Highly factual passages where only one word is correct
- Code, because most of it has to be exact, although comments within code can pick up a mark
- Human writing that Claude has only proofread for grammar and spelling
Converting a document from Word to PDF preserves the words and therefore preserves the mark, whereas genuinely rewriting several paragraphs weakens the signal considerably. Content that has had the least work done to it is consequently the most likely to be detected.
Anything already published is affected by timing, because Claude models released before 2 August 2026 have until 2 December 2026 to add marking, so material produced earlier in the year may carry nothing at all.
What can a Claude watermark actually prove?
A detected watermark shows only that Claude was likely involved with a piece of text at some point. It cannot tell anyone whether Claude wrote the content or merely edited it, cannot identify the person or organisation that used Claude, and cannot report what percentage of a document came from the model.
Anthropic has set out these limits in its own documentation, explaining that detection produces a probability rather than a verdict, and that the probability becomes less reliable as passages get shorter.
What the watermark shows:
- A Claude model probably generated or processed this text
- Confidence increases with the length of the passage
What the watermark does not show:
- Who used Claude, or which company they work for
- Whether the content was written or only edited
- What proportion of the document is AI-generated
- Whether a different AI tool was used
The absence of a mark proves nothing either, because older models, short passages, heavy rewriting and other AI tools all produce unmarked text, so a clean result is not evidence that a human wrote something.
Why has Anthropic added a watermark?
Anthropic added the watermark to comply with Article 50 of the EU AI Act, which took effect on 2 August 2026. The law requires providers of generative AI systems to mark their outputs in a machine-readable format so the content can be identified as artificially generated.
Alongside the law, the European Commission published a voluntary Code of Practice on Transparency of AI-generated Content. Signing it is optional, but the underlying legal obligation is not. Around 190 organisations had signed the Code by 31 July 2026, including Anthropic, Google, Meta, Microsoft, Mistral and OpenAI.
The penalties are substantial, which explains the speed of adoption across the industry. Fines for breaching Article 50 reach EUR 15 million or 3% of total worldwide annual turnover, whichever figure is higher.
Three dates set the timetable:
- 2 August 2026 – transparency obligations take effect. Content published before this date does not need labelling retrospectively.
- 2 December 2026 – deadline for marking AI systems that were already on the market, including older Claude models.
- 2 February 2027 – deadline for providers to make their detection tools work with each other, so text can be checked against multiple watermarks at once.
Anthropic chose to apply the mark globally rather than only in Europe, and has said this is because it does not yet have a durable way to limit marking by region. That decision is what turned an EU compliance requirement into a worldwide change, and it is why businesses in the UK, the US and elsewhere are affected without having any European operation.
What does the watermark mean for businesses publishing content?
Article 50(4) of the EU AI Act places a separate duty on the organisation publishing content rather than on the AI company that built the model. Where AI-generated text is published to inform the public on matters of public interest, the publisher must disclose that it was AI-generated, unless the text has been through genuine human review by someone with responsibility for it.
The European Commission has confirmed that the provider’s marking does not discharge the publisher’s duty. Claude adding a watermark does nothing for a business’s own obligation to label content where labelling is required.
What counts as a matter of public interest
The Commission interprets this broadly, covering any subject open to public debate. Topics inside the scope include:
- public health and consumer safety
- environmental matters
- financial and economic developments
- scientific and cultural developments
Product descriptions, AI-generated fiction and a chatbot reply seen only by the person who asked for it fall outside it.
A great deal of everyday marketing content sits comfortably outside this, although a financial services brand explaining a rate change, a healthcare business writing about a condition, or an energy supplier commenting on environmental policy sits closer to the line than the subject matter alone suggests, because the test can also turn on the publisher’s purpose.
What counts as human review
The exception for human review is stricter than the wording first suggests, because the Commission has stated that spell-checking or a cursory sign-off is not sufficient, that the reviewer needs relevant competence in the subject, and that any substantive AI edit made after the review removes the exception entirely.
A defensible record contains four things:
- A named reviewer with the competence to review that subject.
- A date, tied to a specific version of the content.
- Evidence that the review was substantive rather than a proofread.
- Confirmation that no AI edit was applied after sign-off.
Organisations that already record who reviewed what have very little to change here, which is the practical case for keeping people in the loop at the point of editorial judgement rather than only at the tidying-up stage.
Does the Claude watermark affect SEO or AI visibility?
No search engine can currently read Anthropic’s watermark, because detection requires Anthropic’s secret key and the public detection tool has not launched yet. Each AI company holds its own key, so Google cannot read Claude’s mark and Anthropic cannot read Gemini’s. Claims that watermarks now act as a Google ranking filter are not supported by anything published.
Several SEO articles in 2026 have argued otherwise, in one case naming an algorithm update that does not exist, and the mechanism they describe is not currently available to search engines. Provenance work at the major platforms has concentrated on images, video and audio, where the marking sits in the file and is straightforward to read.
What the available research found
One study has looked at this directly, and its findings are worth reading alongside its limitations. First Page Sage tracked 1,682 pieces of content across 139 websites in four B2B industries between 2 and 24 August 2026.
| Metric | Content made without AI | Content made with AI |
|---|---|---|
| Average Google position | 6 | 11 |
| AI citation rate | 12% | 7% |
The study reports three limitations of its own:
- content quality was not controlled for
- rankings were measured within three days of publication
- the findings show correlation rather than cause
A reasonable interpretation is that this measures the quality gap between carefully reviewed work and lightly supervised model output, rather than any watermark-reading ability at Google. That reading fits with how AI systems retrieve and compare sources, where a model gathering several sources has little reason to cite the fourth page saying the same thing as the first three.
The February 2027 interoperability deadline matters here, because once detection tools are required to work together, checking text against multiple watermarks becomes cheap and routine, and platforms beyond search engines gain access to it. Building generative engine optimisation around original substance rather than volume remains the sensible position either way.
Will other AI platforms add watermarks?
Similar marking is expected across the major AI platforms, because Google, Meta, Microsoft, Mistral and OpenAI all signed the same EU Code of Practice and carry the same Article 50 obligation. The requirement applies to the AI system rather than to any one company, so watermarking is becoming a standard property of generated content rather than an Anthropic policy.
Progress varies by company and by content type:
- Google has been running SynthID text watermarking in Gemini for some time, alongside marking for images, video and audio.
- OpenAI has documented provenance signals more fully for images and audio than for text, and has stated an intention to extend them to text as standards mature.
- Anthropic has published more detail about its text watermarking method than its competitors, which is partly why the change attracted so much attention.
Three developments look likely over the next eighteen months.
Public detection tools are arriving, because Anthropic has said it will offer a watermark detection service and the Code of Practice requires providers to make a detection mechanism available, in principle free of charge, with unrestricted access for regulators, journalists, fact-checkers and researchers.
Detection becomes interoperable from February 2027, allowing a single check across several providers rather than running each one separately.
Labelling will start appearing in products, since the European Commission has finalised three official icons covering fully AI-generated content, AI-modified content, and a basic icon with a second layer of detail, which publishers can use where visual disclosure is possible.
One gap remains open, because forensic detection of content whose mark has been stripped is still optional under the Code, since the available tools do not yet meet the required standard.
What do you need to do about the Claude watermark?
Most businesses need to do three things before December 2026: explain the change internally, record how AI-assisted content is reviewed, and set a written rule against removing marks. Changing AI provider is not one of them, since every major provider signed the same agreement.
What does not need to change:
- Your choice of AI tools
- Any content published before 2 August 2026, which does not need retrospective labelling
- Ordinary use of AI for research, summarising, proofreading or internal documents
What is worth doing:
- Explain it internally in one short paragraph. Much of the worry circulating in businesses concerns confidentiality and client data, which the watermark does not touch, so correcting it early saves a great deal of time.
- Record your review process. Named reviewer, date, version, and evidence that the review was substantive. This is the highest-value action on the list, because it satisfies the Article 50(4) exception.
- Put a written rule against removing marks in your content policy. Article 50 prohibits deliberately removing AI markings, which turns an ordinary editing decision into a deliberate act.
- Add content credentials to your own photography and graphics. Original media benefits from carrying provenance information as detection tooling becomes more common.
- Ask suppliers what they use. Agencies and freelancers producing content under your brand sit inside your obligations rather than theirs, so provenance questions belong in onboarding and contracts.
Suppliers are frequently left out of this process, because a freelancer who drafts in Claude and edits lightly produces content that carries a mark, publishes under someone else’s brand, and forms part of that brand’s compliance position. Adding the question to onboarding takes an afternoon, and it belongs alongside the rest of an enterprise content governance framework.
Final thoughts
The Claude content watermark is a weak signal attached to a strong regulatory direction. It says a model handled the text, it says nothing about who used it, and for now only Anthropic can read it at all.
The change that matters for businesses is the transparency law behind the mark, which places a disclosure duty on publishers that no AI company can satisfy on their behalf. Organisations with a documented editorial review process are largely covered already, and those publishing unreviewed AI output at volume were carrying that risk long before any of this became law.
Detection will become better, interoperable and more widely available over the next eighteen months, so understanding how your content performs in AI search will tell you far more about your visibility than any watermark ever will.
Frequently asked questions
Can you remove the Claude watermark?
The watermark can be removed to some extent, because Anthropic states that light editing probably will not remove it completely, while a full rewrite replacing every word will. Removing a mark deliberately is prohibited under Article 50 of the EU AI Act, which turns an ordinary editing decision into a deliberate act with intent attached.
Does the Claude watermark identify me or my company?
The watermark carries no identifying information about any user or organisation. Anthropic states that nothing in the mark or its key would allow anyone to recover information about a user, their organisation or their conversations with Claude.
Does the Claude watermark apply outside the EU?
The watermark applies worldwide, because marking is done at the model level and travels with the model wherever Claude is offered. Anthropic has said it is applying watermarking globally because it does not yet have a durable way to limit the mark by region.
Will Google penalise watermarked content?
There is no evidence that Google does this or currently can, because detection requires Anthropic’s secret key, which Google does not hold. Provenance features at the major search platforms have focused on images, video and audio rather than article text.
Does watermarking make Claude’s writing worse?
Anthropic says there is no practical impact, citing a study in which Google DeepMind analysed approximately 20 million watermarked and unwatermarked Gemini responses and found the thumbs-up rate differed by 0.01% and the thumbs-down rate by 0.02%, both statistically insignificant. Critics argue that rating buttons measure whether an answer was useful rather than whether the best word was chosen, so for most business content the difference is negligible.
Does the watermark apply if Claude only proofread my writing?
Proofreading rarely produces a detectable mark, because the watermark attaches only to words Claude chooses and a light proofread leaves very little for it to attach to. Anthropic notes that whether it registers depends on the length of the text and how heavily Claude edited it.
Does watermarking affect code?
Code carries very little marking, because an exact output is usually required and the watermark is not applied where only one answer is correct. Anthropic notes that it can appear in arbitrary choices such as comments within code, with negligible effect on the code itself.
Does the watermark change who owns Claude’s output?
Ownership and authorship are unaffected. Anthropic states that a watermark says nothing about who owns content and does not change a user’s rights under its terms, since it only helps test whether Claude produced or processed the text.
