Anthropic has revealed a previously unpublished detail: starting August 2, 2026, new Claude models launched in the European Union will support machine-readable watermarking of generated text from their release. This refers to future releases—no model has been released after that date yet, the latest currently being Claude Opus 5, released on July 24. This is stated in an article in the Help Center, updated on August 10, 2026. The rule will apply not only in the EU, but wherever Claude is offered—in the API, the Claude app, Claude Code, Claude Cowork, Claude Tag, as well as with partners like AWS, Google Cloud, and Microsoft Foundry.
Text and Files—Two Different Mechanisms
For text, Anthropic will use invisible watermarks that will be woven directly into the model's response itself, without changing its meaning, quality, or readability. Since the mark will become part of the text itself, it will be preserved during copy-paste and, partially, even after editing. This mechanism has no relation to the C2PA standard.
For files—for example, .svg, .png, and .jpg—a different, already operational mechanism applies: Claude attaches signed provenance metadata using the open C2PA (Coalition for Content Provenance and Authenticity) standard. Such a signature indicates that the file was processed by Claude and allows for the detection of tampering with the metadata after the fact. Support for this feature depends on the specific platform.
Old Models Without Text Watermarks
The rule regarding text concerns only new versions of Claude. Models released before August 2, 2026—meaning all currently available models, including Claude Opus 5—are in a transitional period and do not embed text watermarks. Anthropic says it is working on adding support for watermarking for them as well, and the documentation will be updated as readiness is achieved.
Cannot Check Text Yet
Anthropic states that it is working on separate tools that will allow users and third parties to detect text watermarks. The company promises to reveal details of the detection mechanism in technical documentation later—no such tool exists at the moment. Sites like c2pa.org or contentcredentials.org are not suitable for this: they check metadata in files, not text watermarks.
An important nuance: even when detection becomes available, the presence of a mark will not be definitive proof that the author of the text is specifically Claude, since the model is often used for editing, translating, or summarizing others' content. The absence of a mark will also prove nothing.
AI Opinion
From the perspective of macroeconomic regulation, the coincidence of dates does not appear accidental. The European Union, starting August 2, 2026, introduced mandatory requirements of Article 50 of the AI Act regarding the labeling of AI content, and the European Commission had already approved the final clarifications for this norm on July 20, providing for fines of up to 15 million euros or 3% of the company's global turnover for violating transparency. The launch of Claude's text watermark precisely on this day is, apparently, not a technical coincidence but a compliance reaction to a specific regulatory deadline.
The technical aspect that the article does not reveal is the resilience of such watermarks to paraphrasing and translation via another model. Industry experiments with text marks have shown that paraphrasing noticeably reduces detection accuracy, and translating text into another language often removes the mark completely. Whether Claude's invisible mark will remain detectable after processing by a third-party AI is a question whose answer only the company itself currently knows.
