Anthropic Plans Claude Watermarks As EU AI Act Duties Start

The Verge reported that Anthropic plans to mark Claude-generated text and images with machine-readable signals as AI transparency duties under the European Union's AI Act begin to reshape product design.

The change is not immediate for every Claude service.

Existing products that were already available before the new obligations took effect on August 2 have a four-month grace period, while new Claude models are expected to carry the marking system from launch.

Anthropic's support material creates two lanes for provenance.

Claude-generated or Claude-processed image files will use C2PA metadata when the file type supports it, putting Anthropic alongside other companies that have adopted the content-authenticity standard.

Text generated by Claude will instead receive an imperceptible watermark that is built into the output itself.

The text approach matters because most AI provenance work has centred on images, audio and video files that can carry metadata.

Anthropic says the text marker should remain with copied text and may survive some later edits, which would make the signal less dependent on a platform preserving attached file data.

The rollout is also broad by product surface.

Supported Claude models are slated to apply the marks across the Claude Platform API, consumer Claude products, Claude Code, Claude Cowork and Claude Tag.

Anthropic also plans to apply text watermarks when its models are accessed through AWS, Google Cloud or Microsoft Foundry.

Detection remains the weaker side of the plan.

Anthropic is working on tools for users and outside parties to identify the embedded signals, but detailed technical documentation has not yet been released.

The Verge noted that existing C2PA inspection tools are available elsewhere, though compatibility with Claude-generated material remains unclear.

Reliability is another unresolved constraint.

C2PA metadata can be removed when files are edited, compressed or uploaded through services that strip provenance data.

Text watermarking faces a different challenge: the marker must be strong enough to persist through ordinary reuse without visibly changing the quality or readability of a response.

For platforms, publishers and software vendors, the operational question is whether Claude's marks become useful enough to integrate into moderation, labelling or audit workflows.

A machine-readable signal could support automated checks, but false negatives would still leave organisations unable to treat the absence of a mark as proof that content was human-made.

The next evidence point is technical rather than policy language.

Anthropic has committed to publishing more detail on detection, and the usefulness of the system will depend on how easily third parties can test, verify and preserve those signals across normal web distribution.