Claude's Words Now Carry a Hidden Signature, and the EU Is Why
From 3 November, Anthropic extends its invisible text watermark to four older Claude models to meet the EU AI Act. Two of them have a settings loophole.
Starting 3 November, more of what Claude writes will carry a hidden signature. Anthropic is extending its invisible text watermark to four older Claude models, so text from most current Claude versions can now be traced back to the model. The move is driven by the EU AI Act, and it shows that AI labelling is becoming the default, not an option.
The four models are Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6, according to an email Anthropic sent API customers, as reported by The Decoder and Mashable. Newer models such as Claude Opus 5.5 already carry the mark. The rollout on Amazon Bedrock may take a few extra days.
How the watermark works
Anthropic describes the mark as "an imperceptible pattern in word choice." It is woven into the text as the model writes, so it travels with the words when they are copied and pasted. Anthropic says it adds no extra characters, does not change pricing or response formats, barely slows output and contains no information about the user or their conversations. Those are company claims that have not been independently tested.
Because the mark lives inside the model, it applies wherever Claude runs, including AWS, Google Cloud and Microsoft Foundry, and customers cannot turn it off. Files Claude creates, such as images, get a separate signed label using the C2PA Content Credentials standard.
The Decoder reports that the method is a variant of Google DeepMind's SynthID text technique. Not everyone accepts Anthropic's claim that quality is untouched. Blogger John Gruber argues that choosing words by a watermark key instead of best fit must cost some precision.
The loophole
There is a catch on two of the four models. On Opus 4.6 and Sonnet 4.6, the watermark applies only when developers leave three sampling settings, temperature, top_p and top_k, at their defaults. Change any one and the mark is skipped. Anthropic's own email warns: "If you rely on the watermark for your own compliance, keep these settings at their defaults." The email does not mention that limit for Mythos 5 or Opus 4.7.
Detection is also limited. Anthropic's watermark checker is in private preview for regulators, media, researchers, educators and companies with their own compliance duties. Anthropic itself says a detected mark is a signal, not proof. Short passages, heavy edits, paraphrasing or translation can all hide the mark.
Why the EU is forcing this
Article 50 of the EU AI Act requires providers of systems that generate text, images, audio or video to mark the output in a machine-readable format so it can be detected as artificially generated. Anthropic has signed the EU's Code of Practice on the issue, and The Decoder reports the company points to a December deadline for older models.
Rivals are moving too. SpaceXSI, the name xAI now uses in its developer documentation, says Grok API text from seven models will be watermarked by default from 7 November, though team admins can opt out per model. The Decoder reports that OpenAI's system ships switched off for API customers.
For privacy, the trade is new. The mark does not identify you, Anthropic says, but it does mean your Claude-written text can be flagged as AI by anyone with detector access. If you work with clients who have banned AI tools, that matters.
Photo: "Typing on a laptop" by Simon Hattinga Verschure, via Unsplash and Wikimedia Commons, CC0. Graded and overlaid by PSI.
Sources
- How Claude marks AI-generated content (Anthropic)
- Anthropic's watermark rollout leaves a loophole on two older Claude models (The Decoder)
- Anthropic expands text watermarking to older Claude versions beginning November 3 (Mashable)
- Article 50: Transparency Obligations for Providers and Deployers of Certain AI Systems (EU AI Act)
- Output Watermarking (SpaceXSI Docs)