Anthropic adds invisible watermarks to all Claude models

Compliance with EU regulations

Anthropic has begun a rollout to add machine-readable watermarks to its entire suite of Claude models. This initiative is a direct response to the EU AI Act, specifically Article 50, which requires clear marking of AI-generated content. The company signed a code of practice to meet these legal requirements.

The rollout is happening in waves to ensure all models are covered. While newer models already had these features, Anthropic is now retrofitting older versions like Claude Mythos 5, Opus 4.7, and Sonnet 4.6. The company aims to have the retrofit completed by a December deadline.

Because these marks are applied at the model level, they are universal. The watermarks will appear in text generated through various platforms, including the Claude apps, Claude Code, Cowork, and the API. This also means the marks will show up for users on Amazon Bedrock, Google Cloud, and Microsoft Foundry.

How the invisible mark works

The technology behind the watermark involves subtle adjustments to the way Claude selects words. By slightly altering word choices, the system creates a statistical pattern that identifies the text as AI-generated. These changes are designed to be completely invisible to the human eye during normal reading.

Anthropic reportedly uses a variant of SynthID, which is the same technology Google uses for AI images in iOS 27.2. This method allows the mark to travel with the text even when it is copied and pasted into other applications. The mark may also persist through some basic editing of the content.

However, the watermark is not indestructible. If a passage is heavily rewritten, translated, or is very short, the mark may no longer be detectable. Anthropic also clarifies that the watermark does not prove authorship; for example, text that Claude only proofread or translated will still carry the mark.

Who can see the watermark

For the general public, the watermark is currently invisible and there is no way to check it. Anthropic has placed the detection tools in a private preview. This means only specific groups can access the technology to verify the marks on content.

Access to these detection tools is restricted to regulators, law enforcement, media outlets, and fact-checkers. It is also available to researchers, educational organizations, and companies that have specific legal obligations to monitor AI content. These groups must submit an application to gain access to the detection tools.

For visual media, Anthropic uses a different approach for PNG and JPEG files. These files include Content Credentials based on the C2PA standard. Unlike the text watermarks, these credentials can be read by anyone using the free Claude Content Checker tool provided by the company.

Technical limitations for developers

While the watermark is mandatory for the Claude apps, there is a technical loophole for developers using the API. On certain models like Sonnet 4.6 and Opus 4.6, the watermark is only reliably applied when specific settings are left at their default values. These settings include temperature, top_p, and top_k.

If a developer manually changes any of these parameters through the API, the resulting text may not contain a detectable watermark. This creates a distinction between the standard user experience and the customizable environment provided to software engineers. The Claude apps do not offer these settings to the general public.

This approach differs from how OpenAI handles similar requirements. According to reports from The Decoder, OpenAI has had watermarks enabled by default for ChatGPT and Codex users in the EU since October. However, their API customers must manually choose to turn the watermark feature on themselves.

Privacy and data security

Anthropic has provided specific assurances regarding what information is contained within the watermark. The company states that the mark does not include any personal data. It does not contain information about the specific user, their organization, or the content of their private conversations with Claude.

The primary goal of the watermark is identification of the source rather than the collection of metadata. By keeping the mark at the model level, Anthropic aims to satisfy the EU's transparency requirements without compromising the privacy of the individuals interacting with the AI. This balance is central to their strategy for international compliance.

As the rollout continues, the company will likely face scrutiny over the effectiveness of these marks. Because the marks can be stripped by heavy editing or specific API configurations, the ultimate success of this policy depends on how well regulators can identify the most common types of AI-generated content in the wild.