Claude introduces an invisible marker in AI-generated text

Claude will gradually introduce an invisible marker embedded in its text, although users will not be able to see it. A detector will be able to use this marker to establish whether content was created or processed by Anthropic’s AI. The company has made the announcement after signing the European Union’s “Code of Practice on the Transparency of AI-Generated Content”.

Soon, text generated with Claude may be identifiable through an invisible marker. Its developer, Anthropic, has signed the European Union’s “Code of Practice on the Transparency of AI-Generated Content” in order to comply with Brussels’ new rules. One result of signing it is the introduction of this invisible marker: it will not be visible to users, but software will be able to determine whether a text was generated by AI.

Claude’s invisible watermark for AI-generated content

This transparency measure will initially apply to every AI model offered by Claude in the EU from 2 August. “Generated text will have embedded watermarks and generated files will include digitally signed provenance metadata where possible,” Anthropic explains. Older AI models may also receive this type of invisible marker at a later date. The company says it is also working on bringing support for this new feature to its earlier models.

The markers will be hidden in generated text worldwide, rather than only in the EU. In addition, every product will be affected where the model in use supports the feature. This includes Claude, Claude Code, Claude Cowork and Claude Tag, as well as Claude models used by third-party applications through the API. Software capable of detecting the invisible marker is already being developed. “We are also working to enable users and other third parties to detect Claude’s embedded watermarks and provenance metadata. Detection involves checking whether text or a file contains a supported Claude mark. If such a mark is detected, this indicates that the content may have been processed by Claude,” Anthropic says.

Invisible marker detection requires very careful interpretation

While introducing this technology, Anthropic acknowledges its limitations. “The presence of a detected mark indicates that content was processed by Claude, but is not fully conclusive proof,” the company warns. For instance, a text may contain Claude’s marker where its author used Anthropic’s AI to proofread, translate or even convert a file.

Likewise, the absence of this marker does not provide complete certainty that a text was not generated or processed by Claude. It could have been produced by an older Anthropic AI model that does not include its new invisible marker. Altering content may also remove the invisible marker. Anthropic further acknowledges that its signal is unreliable when a text is too short.

Other EU Code of Practice signatories

In any case, Anthropic is not the only AI laboratory to have signed the EU’s Code of Practice on AI. Other signatories include Google and the Meta group.