Anthropic adds mandatory watermarks to Claude text for EU compliance

Compliance with EU regulations

Anthropic is updating its Claude models to include machine-readable watermarks on all generated text. This change is driven by the company's commitment to the EU AI Act, specifically Article 50, which requires clear labeling of AI-generated content. As reported by Notebookcheck, the company signed a code of practice to meet these transparency standards.

The rollout began with models released after August 2 and is now expanding to older versions. Specifically, Claude Mythos 5, Opus 4.7, and Sonnet 4.6 have been updated to include these marks. Anthropic cited a December deadline for completing this retrofit across its entire model lineup.

The watermark applies globally and across all of Anthropic's products. This includes the standard Claude apps, Claude Code, Cowork, and the API. Because the marking happens at the model level, it also appears in text generated through third-party platforms like AWS, Google Cloud, or Microsoft Foundry.

How the invisible mark works

The watermark functions by subtly adjusting the statistical patterns of word choices. While the changes are designed to be unnoticeable to human readers, they create a signature that machines can identify. Anthropic reportedly uses a variant of Google's SynthID technology, which is the same system used for AI images in Apple's iOS 27.2.

Anthropic has stated that these marks do not contain any personal data. The system does not store information about the user, their specific organization, or the content of their private conversations with the AI. The goal is purely to identify the source of the text rather than to track the individual using the tool.

The mark is designed to travel with the text even when it is copied and pasted into other documents. It may persist through some minor editing, but it is not permanent in every scenario. If a user heavily rewrites a passage, translates it into another language, or keeps the output very short, the watermark may become undetectable.

Who can see the watermark

Currently, the general public cannot easily verify these watermarks. Detection capabilities are currently in a private preview phase and are restricted to specific groups. These include regulators, law enforcement, media outlets, fact-checkers, and educational organizations.

To gain access to the detection tools, these organizations must submit an application for review. This restricted access ensures that the tools are used by those with specific legal obligations or public interest roles. It prevents the general public from constantly scanning every piece of text for a digital signature.

Anthropic also notes that a detected watermark does not serve as definitive proof of authorship. For example, if a human asks Claude to proofread a draft or translate a document, the final output will still carry the mark. The system identifies that the AI was involved in the process, but it does not distinguish between original generation and assisted editing.

Technical limitations and loopholes

There are specific technical conditions under which the watermark might not be reliably applied. On the Sonnet 4.6 and Opus 4.6 models, the mark is only consistent when certain API settings remain at their default values. These settings include temperature, top_p, and top_k, which control the randomness and diversity of the AI's responses.

Developers who use the API and intentionally change these parameters may receive text that the watermark system cannot reliably detect. However, these specific settings are not available to users of the standard Claude apps, meaning the watermark is more consistent for everyday consumers. This creates a distinction between how developers and general users interact with the model.

Anthropic's approach differs from some of its competitors in the industry. For instance, OpenAI has only enabled its watermark by default for users in the EU who use ChatGPT or Codex. In contrast, Anthropic's watermark is mandatory for all users and cannot be opted out of by API customers. This reflects a more aggressive stance on meeting the requirements of the EU AI Act.

Visual media and future steps

While the current news focuses on text, Anthropic is also applying standards to visual media. For PNG and JPEG files, the company adds Content Credentials based on the C2PA standard. This is a different type of metadata that allows anyone to verify the origin of an image using the free Claude Content Checker.

The move toward mandatory watermarking suggests a broader trend of accountability in the AI industry. By integrating these marks at the model level, Anthropic ensures that the labels are baked into the output regardless of which cloud provider or interface the user chooses to use.

Moving forward, the company will continue to expand these marks to all remaining models in its catalog. As the EU AI Act moves toward full enforcement, these technical safeguards will likely become the standard for how AI companies handle content generation. The focus remains on balancing user privacy with the need for machine-readable transparency.