Claude text watermarks: detector limited to vetted groups

Nine Claude models now put an invisible watermark into the text they write, and some of them were released before the EU’s August 2 cutoff. That comes from the model table in Anthropic’s support article on how Claude marks AI-generated content, as it stood on October 23. Anthropic says the marks fulfil its commitments under the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content, which it signed both as a provider of generative AI models and as a provider of generative AI systems.

The scope has grown since August. An instructor on O’Reilly’s live training platform quotes Anthropic’s August announcement as saying that “Future Claude models will generate text that contains a watermark.” The support page now goes further: “Consistent with our commitments under the Code, Anthropic is adding watermarks to outputs from models released before August 2, 2026.” Fable 5 and Opus 4.8, both in service by early July, show check marks in all three columns: text watermarks on Anthropic’s own surfaces, text watermarks on AWS, Google Cloud and Microsoft Foundry, and Content Credentials on files. The other models with all three are Fable 5.1, Mythos 5.1, Opus 5.5, Opus 5, Sonnet 5.5, Sonnet 5 and Haiku 5.5.

Fable 5, Opus 4.8 and Sonnet 5 also carry an asterisk in the cloud partner column, with a footnote: “Rollout on Amazon Bedrock will complete by October 23.” Seven older models get a single check mark each: Mythos 5, Opus 4.7, Opus 4.6, Opus 4.5, Sonnet 4.6, Sonnet 4.5 and Haiku 4.5. None of them appear on the O’Reilly instructor’s list of models that watermark their text. So for those seven the single mark is most likely the file credential. Search Engine Journal reports that Claude models launched in the EU on or after August 2 support marking at launch. Earlier models fall under a transition period, and Anthropic is working to add marking to them as well.

The text watermark works by changing where the model gets its randomness. Claude writes one token at a time, and when several candidate words mean much the same thing, a random number usually picks one. With watermarking, as Anthropic’s announcement explains, that choice is made by a secret key combined with the few words before it. Anyone holding the key can check whether a sequence of words matches Claude’s choices and estimate how likely it is that Claude wrote the text. Anthropic says the method never pushes Claude toward a word it wouldn’t have considered anyway. It adds no hidden characters or extra tokens, costs nothing extra, and carries nothing that could identify a person, an organization or a chat.

Because the mark is applied at the model level, it comes with the text from Claude Desktop, Claude Code and the API alike. It stays in the text through copy and paste, and it may survive some editing. Files work differently. When Claude generates a supported type such as a PNG or JPEG, it attaches signed provenance metadata, called a Content Credential, that follows the open C2PA standard.

Checking for these marks is uneven. Anyone can test a file with the free Claude Content Checker. Detecting text watermarks is still in private preview, open only to organizations eligible under EU law: regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations and EU civil society groups. Enterprises that have to verify watermarking for their own compliance with the Act can get access too, and others can register interest through an access request form. The O’Reilly instructor wrote that “anyone who has an AI watermark detector can easily determine if that content was AI generated”. For now, that group of people is small.

A positive result also proves less than it might seem. A detected mark means the content may have been generated or processed by Claude, not that Claude wrote it. Proofreading, translation, summarizing and file conversion can all leave a mark even when the ideas and the original text came from someone else. A mark can also be missing: the text may come from an older model, it may have been heavily edited, or it may be too short to give a clear signal.

Anthropic isn’t the only lab doing this. Search Engine Journal says the European Commission counted about 190 signatories by the end of July. Google, Meta, Microsoft, Mistral and OpenAI joined Anthropic in the provider section, which covers machine-readable marking and detection. OpenAI’s name stands out. The same outlet’s author previously covered OpenAI’s decision in September 2024 to drop its own watermarking plans, after a company survey found almost 30% of ChatGPT users would use it less with watermarking added.

Teams that build Claude into their own products may have their own obligations. Search Engine Journal points to two exemptions in Article 50. Systems that only assist with standard editing can fall outside the marking requirement if they don’t substantially change the input or its meaning. Published text that has gone through qualifying human review can be exempt from the disclosure requirement when someone holds editorial responsibility for it. Anthropic leaves that call to its customers: “If you deploy Claude in your own product, you should independently assess what Article 50 requires of your products and services.”