Anthropic’s Claude to Embed Markers in AI-Generated Content

Anthropic is introducing machine-readable markers for content generated by its Claude AI models as part of its efforts to meet transparency requirements under the European Union’s AI regulations.

The company has signed the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content. As a result, Claude models launched in the European Union from August 2 onward will include machine-readable identification signals when generating content.

One of the most notable developments is Anthropic’s approach to identifying AI-generated text. Unlike images, where AI-generated content can sometimes be identified through visible or embedded metadata, text can easily be copied and presented as human-written.

Anthropic says Claude-generated text will contain an imperceptible watermark designed to remain associated with the content even when it is copied, pasted or lightly modified.

For supported files, the system will use another form of provenance tracking. Claude-generated SVG, PNG and JPG files can include digitally signed provenance metadata indicating their AI-generated origin.

However, Anthropic cautions that these signals will not necessarily prove that Claude originally created every part of a piece of content. For example, AI may have been used to process or modify material that was originally produced by a person.

The company also acknowledges that detection may not always work. Extensive editing, very short pieces of text or the removal of metadata can result in false negatives.

The markings are expected to apply to outputs from supported Claude models across Anthropic’s ecosystem, including its API platform, Claude, Claude Code, Claude Cowork and Claude Tag. Anthropic said support for individual marking methods may vary depending on the platform and feature.

The company also confirmed that supported Claude models will retain these embedded markings when accessed through cloud platforms such as AWS, Google Cloud and Microsoft Foundry.

Anthropic said it plans to release additional information about how users, researchers and other third parties can identify and verify Claude’s content markers.

The move comes as AI-generated material becomes increasingly common online, making provenance and transparency mechanisms more important for distinguishing AI-assisted content from entirely human-created work.