AI Text Watermarks Are Here: What Anthropic’s Launch Means

AI text watermarks spent years as an announcement without a product. That ended on 2 August 2026: right on the enforcement date of the transparency obligations in Article 50 of the EU AI Act, Anthropic began embedding an invisible watermark into the output of its new Claude models. Google has been marking Gemini output with SynthID since 2024, so Anthropic is the second major provider to follow, and the first to frame it explicitly as an answer to European regulation. It sounds like a story for providers, yet it concerns every company that produces text with AI. What gets marked, what a detection actually proves and which follow-up work lands on your desk is what this piece sorts out.

How the text watermark works

According to Anthropic, the method builds on SynthID-Text, the approach Google DeepMind developed and runs in Gemini itself. When a model generates an answer, every token is picked with a share of randomness. That randomness is shifted so that a statistical pattern accumulates across many tokens. Readers cannot see it, style and quality stay untouched, and there is no metadata attachment that could get lost in copying. The pattern lives in the text itself and travels with it, whether into a CMS, an email or a Word document. For file outputs, Anthropic adds signed provenance metadata.

Marking applies to the output of models launched on or after 2 August 2026. Older models, including Opus 5 released in July, are supposed to follow; Anthropic’s announcement names no date. The marking is active on every surface Claude is available on: the API, the web interface, Claude Code, Cowork and the cloud platforms of AWS, Google and Microsoft.

What a detection proves, and what it does not

The most important point first: no publicly usable detector exists yet. Anthropic has announced tooling for users and third parties, but none has shipped so far. Today nobody outside Anthropic can test a text for the watermark, us included.

Even with a detector, the evidence stays limited. A positive result means: this text may have been processed by Claude. It does not mean Claude wrote the text, and certainly not who commissioned it. Light editing often survives, according to Anthropic, while heavy revision, translation or blending with other material makes detection unreliable, and a full rewrite removes it. Anyone who publishes bilingually and edits thoroughly, as we do, produces exactly the kind of text where detection is weakest.

Two sobering conclusions follow. As a plagiarism or AI detector the watermark is not fit for purpose, it breaks too easily. As a free pass it is just as useless: what nobody can test today may be testable tomorrow. Any process that quietly relies on AI use staying invisible is built on sand.

The provider met its duty, the responsibility stays with you

Legally, the launch slots in cleanly. Article 50(2) of the AI Act obliges providers of generative systems to mark outputs as artificially generated in a machine-readable form. That duty sits with Anthropic, OpenAI and their peers, not with the company using the tools. For images, machine-readable marking via C2PA manifests has been established practice among the large providers for a while; text was the open flank. With this launch, Claude closes it.

For companies deploying AI, their own obligations do not change. Whoever publishes AI-generated text that informs the public on matters of public interest still needs either a visible label or the exemption under Article 50(4): a named person reviews the text in substance and carries editorial responsibility. A watermark does not replace that review, it does not even document it. We record reviewer and review date for every blog article directly in the content repository, and this launch changes nothing about that process. The 72-day roadmap for the labeling obligation we published in May remains valid; the gap it described in text marking has simply been closed by the provider since.

Three items for your own governance

First, add launch dates to the model inventory. Only output from models released on or after 2 August 2026 carries the mark. To know which of your content is watermarked, you need to know which model produced it and when. One line per tool in the AI inventory does the job.

Second, review every process that silently assumes non-detectability. Texts submitted by agencies, job applications, expert reports: once detectors ship, Claude involvement in them may become visible. The right response is not a ban but an honest rule for where AI assistance is fine and where it must be disclosed.

Third, documented human review remains the evidence that holds. The exemption in paragraph 4 is tied to editorial responsibility, and that needs a record: who reviewed, and when. It belongs in the same step as publishing, not in an after-the-fact reconstruction.

Conclusion

The watermark makes AI text neither forbidden nor reportable, it makes it traceable in principle. If you know which models are in use, build no process on invisibility and document your editorial reviews, you have already absorbed this launch. For everyone else it is a good moment to set up AI governance properly. For working out which obligations the AI Act triggers in your specific setup and how to keep the evidence pragmatic, our AI consulting for the mid-market helps hands-on.

Frequently Asked Questions

What is the Claude watermark?

An invisible statistical pattern that Anthropic has been embedding into the text output of new Claude models since 2 August 2026. It builds on the SynthID-Text method, shifts token selection minimally and survives copying. The trigger is the machine-readable marking duty in Article 50 of the EU AI Act.

Can you check whether a text carries an AI watermark?

Not at the moment. Anthropic has announced detection tooling for users and third parties but has not released any. Even a future match only shows that Claude may have been involved in a text, not authorship or who commissioned it. Heavy editing and translation additionally make detection unreliable.

Does the watermark satisfy the labeling obligation for companies?

No. Machine-readable marking is a provider duty under Article 50(2) and sits with Anthropic. Companies publishing AI text on matters of public interest still need either a visible label or the documented editorial review by a named person under Article 50(4) of the AI Act.

Which Claude models are affected by the watermark?

Every model Anthropic has released since 2 August 2026 marks its output from day one. Older models, including Opus 5 from July 2026, are supposed to be retrofitted; Anthropic has not committed to a timeline for that yet.