Claude now signs the texts it generates: what changes for those who publish AI content
On August 11, 2026, Anthropic announced that Claude models launched from August 2 onward embed an invisible statistical watermark in generated text, worldwide, in implementation of the Code of Conduct under Article 50 of the AI Act. What this means for those who publish AI-generated content: which texts are marked and which are not, whether Google will penalize recognizable content, why 'how do I remove it' is the wrong question, and what to do in practice — explained by those who implemented transparency before the regulation.
Claude now signs the texts it generates. The right question is not «how do I remove it».
On August 11, 2026, Anthropic announced the invisible watermark in the texts of the new Claude models, worldwide. What is marked and what is not, what your SEO is really at risk of and what you should do — read by someone who implemented AI content transparency before it became mandatory.
All factual claims in this article cite Anthropic's official announcement and the press that covered it, with sources at the bottom of the page. The picture is evolving: we will update the article when the technical details on detection are released.
In summary
From 11 August 2026, Anthropic marks text generated by Claude models launched from 2 August onwards with an invisible statistical watermark, applied at model level and worldwide — not only in Europe. This is the implementation of the Code of Conduct on transparency under Article 50 of the AI Act, which Anthropic has signed.
- It applies to new models: those launched from 2 August 2026. Previous models do not (yet) mark: Anthropic states it is working to extend support.
- The watermark is in the text, not in the metadata: it survives copy-paste and «can persist through some edits»; it weakens with heavy editing, paraphrasing and translation, and very short texts produce an unreliable signal.
- No public detector exists yet: details on detection will come «in future technical documentation».
- For publishers, the substance does not change: the safeguard is not about hiding the origin, it is about governing content — documented human review, verified sources, declared transparency.
What exactly did Anthropic announce?
Two complementary mechanisms, declared on the official support page:
-
Statistical watermark in the text
An imperceptible signal woven into the text itself: it can't be seen, it doesn't change meaning or readability, and it travels with the text when copied and pasted elsewhere. It's not metadata that gets lost when saving the file: it lies in the choice of words.
-
C2PA provenance metadata
In parallel, metadata signed according to the open C2PA standard (Coalition for Content Provenance and Authenticity) — the same approach already used for generated images.
Three limits declared by Anthropic itself, worth more than any speculation. First: the watermark applies to models launched from August 2, 2026; for previous models, support «is in progress». Second: detectability weakens if the text is «heavily edited, paraphrased, translated, or mixed with other writing», and very short passages give an unreliable signal. Third: the mark proves that the text has passed through Claude, not who its author is — Anthropic acknowledges that people use Claude even just to correct, translate, or summarize their own texts, and that the mark «does not by itself confirm the full provenance of the content».
Will marked content be penalized by Google?
It is the first fear of anyone who publishes AI-generated content, and it is the wrong fear — for three verifiable reasons.
First: Google's position has not changed. Google has stated for years that it does not penalise AI content as such: it evaluates usefulness for people and targets mass production without value, regardless of the tool used. There is no announcement that watermarks will become a ranking factor.
Second: there is not even a public detector today. Anthropic has stated that details on detection mechanisms will be provided in future technical documentation. Fearing a penalty based on a signal that no one can yet read is premature by definition.
Third: our direct experience says the opposite. The content of one of our projects, produced with AraLith COS's governed AI pipeline and mandatory human review, ranks on the first page of Google and is cited as a source inside the AI Overview. If search engines wanted to penalise AI-sourced text as such, they would not cite ours among their sources. What search engines reward — and cite — is verifiable quality, not the absence of provenance.
Why «how do I remove it» is the wrong question
In the coming weeks, tools and guides will emerge to «clean» AI texts of their watermark. Before spending a single euro on them, it's worth understanding the legal architecture — because it dismantles the problem at its root.
Art. 50 of the AI Act distinguishes two obligations that are regularly confused. Paragraph 2 requires the provider of the AI system (in this case Anthropic) to mark outputs in a machine-readable format: the watermark is Anthropic's compliance obligation, not yours. Paragraph 4 concerns those who publish: AI-generated texts published to inform the public must be disclosed — unless the content has undergone human review or editorial oversight by a person who holds responsibility for it. We have analysed this article by article in the guide to Art. 50 of the AI Act.
The practical consequence: the editorial review exemption concerns your disclosure obligation, not the provider's watermark. The law does not ask you to remove it — and it is not in your interest to do so: your protection is not the absence of the mark, it is the evidence of human review (who approved it, when, what was changed). Moreover, for text that has been genuinely revised the problem tends to resolve itself: Anthropic itself states that substantial editing, rewriting, and integration with original text weaken the signal. Real editorial work produces that effect as a consequence, without the need for «cleaning» tools — and those who publish raw AI text passing it off as human-written have a quality problem before they even have a watermark problem.
We've always done the marking — on the right side
This news did not catch us off guard, because inside AraLith COS the transparency architecture has been in production since before the Code of Conduct came into force. Generated images already come with machine-readable metadata (DigitalSourceType «trainedAlgorithmicMedia», re-injected on variants as well); AI content carries a visible badge and a note in the frontend; and above all no AI modification to a published piece of content goes live on its own: it goes through a review queue with a field-by-field diff, and the approval records who reviewed it and when — exactly the evidence that triggers the editorial exemption under Art. 50, para. 4.
That is why we read Anthropic's announcement without anxiety: when content provenance is governed — tracked sources, documented review, declared transparency — one more watermark does not change your risk profile. It does change it for those who built on concealment.
What to do in practice, if you publish content produced with AI
Audit the models you use
The watershed date is August 2, 2026
Knowing which tools in your pipeline use models released after that date tells you which texts already arrive marked. And it applies to all providers: the AI Act Code of Conduct covers the entire industry, not just Anthropic.
Don't invest in «removal»
It's the wrong direction
The law doesn't require it, the tools that promise it work against the purpose of the regulation, and your real risk — quality — remains intact. Money spent hiding the origin yields more when spent on revision.
Document human review
It's your exemption, and your defence
Who approved it, when, what changed: this is the evidence that activates the editorial exemption under art. 50, para. 4 — and that distinguishes a governed editorial workflow from the publication of raw text.
Treat transparency as an asset
Not as an obligation to be minimised
Generative engines cite sources with clear provenance, verifiable data and clean structure. In a web filling up with anonymous synthetic text, declared provenance is a trust signal — for readers and for the AIs that choose whom to cite.
Frequently asked questions about Claude's watermark
No, if produced with models launched before August 2, 2026 — the watermark applies to models launched from that date onwards. Anthropic has stated that it is working to extend support to previous models, so the scope is set to expand.
No: Anthropic describes it as an imperceptible signal woven into the text, which does not change its meaning, quality, or readability. It is not a file metadata: it travels with the text when it is copied and pasted elsewhere.
Partially. It survives copy-pasting and «can persist through some edits», but Anthropic states that detectability weakens if the text is heavily modified, paraphrased, translated, or mixed with other writing — and that very short passages give an unreliable signal.
There is no evidence to say so: Google's stated position is to evaluate the usefulness of content, not the tool used to produce it, and there is not even a public watermark detector yet. Our direct experience points in the opposite direction: content produced with governed AI pipelines is cited within Google's AI Overview.
The question stems from a misunderstanding: the exemption under art. 50, c. 4 (human review with editorial responsibility) concerns your obligation to disclose the content, not the provider's marking, which is an obligation of Anthropic (c. 2). The law does not require removal, and your protection lies in documented evidence of the review — not in the absence of the mark.
Do your AI contents have governance or just a prompt?
Tracked sources, documented human review, declared transparency: this is how AI content holds up to the regulations coming — and becomes what search engines choose to cite. If you want to understand how to set it up in your workflow, let's talk.
Sources and methodological note
Primary sources (accessed on August 12, 2026): official Anthropic help centre page «How Claude marks AI-generated content» (support.claude.com), from which all quoted citations are taken; announcement covered by TechCrunch, Euronews and Forbes on August 11, 2026. For the regulatory framework: Reg. (EU) 2024/1689 (AI Act), art. 50 — our analysis with the cited articles is in the dedicated guide.
This article is the technical read for those who build and use content generation systems, not legal advice. The landscape is evolving rapidly: the public watermark detector is not yet available and technical details are expected in future documentation — we will update the article when they are released. Information accurate as of the publication date.
