Anthropic has started embedding machine-readable watermarks in text generated by a new generation of Claude models. This approach was first used for the relevant model that was launched in the EU after 2 August 2026, and will then be extended to the world, covering Claude, API, Claude Code, and many cloud-based cooperation channels.
Synchronize implementation with EU norms
Anthropic discloses this arrangement in the supporting document. The background is that companies have signed the Code of Conduct on Transparency accompanying the EU AI Act. According to the company, the watermarks are directly embedded in the text itself and the user ' s naked eye cannot be identified and will not change the meaning, readability or surface quality of the text.
According to its description, such marks would remain with the text after being copied, pasted and partially edited. In addition to text-level watermarks, the document will be accompanied by a second layer of signature metadata based on the C2PA open standard for recording the source of content and subsequent modifications.
Test methods are not yet available
Anthropic currently does not specify how the watermarks will be achieved. The information available shows only that watermarks are valid at the model level and are part of the text birthmark rather than a separate external metadata tool. Test documents and identification tools for the validation of watermarks have not been published to date.
On this basis, it is assumed that this may be a statistical feature option, i.e., that the model has a weak but detectable bias in its choice of wording. Such thinking is close to Google 's Syndid Text, but it is still not possible to identify specific mechanisms and thresholds until Anthropic releases its detectors.
Open-source project tried to remove
Shortly after the Anthropic announcement, several removal projects for Claude watermarks have appeared on GitHub. One of the projects will first clean up the invisible Un-currency distribution of de characters, and then rewrite the text with a non-Claude model to disrupt the possible word distribution feature.
Another larger open source project claims to be able to remove Claude text tags, as well as C2PA and SynthID type signals from PNG, JPEG, SVG, PDF, DOCX, etc. According to the relevant developers, the statistical text watermarks are not reliable enough to support the source of the content and may lead users to use the second model to clean the text.
The report also mentions that in March this year Anthropic removed a hidden trace mark from Claude Code. At that time, the researchers stated that the undisclosed Unnet issue de tagging related to the location of some users and the proxy use of information, which also gave rise to greater privacy discussions in the watermarking programme.
Additional information:The report mentions that a United States bill called COPIED Act is also promoting the standardization of AI content watermarks. Anthropic still has no indication of when an external test tool will be released.