Anthropic commits to invisible watermarks in text and images from Claude

Anthropic announced it will embed invisible watermarks in text and images produced by its Claude models, a step aimed at meeting the transparency requirements of the European AI Act. The watermark will be applied at the model level, so any product or interface built on Claude—from the API through Claude Code to Claude Cowork—will automatically generate marked content.

What the European law requires

The AI Act came into force on 2 August, with a four-month grace period for products already on the market. According to the law, Anthropic will have new models watermark content from day one, while support for existing models is still under development. The company has not set a deadline for completing the retroactive rollout.

How it works in practice

Two different mechanisms are used. For images, Anthropic adopts the C2PA standard, a provenance metadata format also adopted by Adobe, OpenAI and Google, which will be embedded in supported file types. For text, a separate method called “invisible watermark” is inserted directly into the string generated by the model, without changing meaning, quality or readability. Anthropic says the watermark survives copy-paste and may survive certain edits.

The watermark operates at the model level, meaning it will be present regardless of the platform used to access Claude, including access via AWS, Google Cloud or Microsoft Foundry. The company is also building tools that will let users and third parties detect the marks, and promises to publish detailed technical documentation later.

Detection and known issues

There are caveats. C2PA metadata is known to be easy to strip, sometimes inadvertently when uploading to online platforms, and it is unclear how robust Anthropic’s text-watermark solution is. The company itself says the systems are “far from being fault-tolerant,” and any content lacking identifiable marks may still originate from a generative model.

Meanwhile, communities such as AO3 fan-fiction readers have already built primitive detection systems for Claude-generated content, but the new standards are expected to provide far broader coverage, provided they actually work.