Anthropic introduces an invisible watermark system that allows it to be determined later that the texts created by Claude are the product of artificial intelligence. Adopting the European Union’s code of practice, which aims to increase transparency in content produced by artificial intelligence, the company places a mark in the texts created by supported Claude models that is not noticed by humans but can be detected by machines. This watermark is added directly during the output creation process without changing the meaning, quality or readability of the text. Moreover, copying the text and pasting it to another platform does not remove the mark, and the watermark can be preserved even in limited edits. However, the system may not be as effective with very short texts or with extensive rearrangement of content.
According to information provided by Anthropic, all new Claude models available in the European Union as of August 2 automatically apply the text watermark. The fact that the marking process is carried out directly by the model ensures that it works independently of the product or infrastructure used. Therefore, eligible outputs from Claude API, Claude Code, Cowork or other Claude-based services are subjected to the same mechanism. Similarly, using the model via AWS, Google Cloud or Microsoft Foundry does not change the watermark application. In addition, Anthropic plans to expand support for older models introduced before August 2 and implement the system on a global scale, not just limited to the European Union.
Claude text watermark can also be preserved in edited content
The main difference between text watermarks and visible signs used in images is that they do not appear to the reader as a logo, symbol or description. As Claude produces the text, he incorporates a pattern into the output that detection tools can recognize. According to Anthropic, this method continues to work when the text is copied elsewhere and is somewhat robust to minor changes. Despite this, the system is not an absolute detection method; In short printouts, there may not be enough signal, and in heavily rewritten texts, the signal may be lost. The company is also working on detection tools that will identify content watermarked by Claude, and states that these will be available later.
One of the striking aspects of the application is that the watermark is not limited to the content that Claude wrote from scratch. When a user has Claude translate, summarize, format, or simply correct existing text for grammar, the resulting output will also be watermarked. This approach covers a wider area than the exception in the relevant implementing rules of the European Union. These rules do not require marking in cases where artificial intelligence serves the purpose of standard text editing or does not substantially change the content and meaning provided by the user. Instead of applying this distinction, Anthropic prefers to mark all supported texts originating from Claude. Therefore, detecting the watermark does not necessarily mean that the ideas or original content in the text in question were created by Claude.
The method on the text side also differs from the application planned for visuals. Anthropic plans to use C2PA-based metadata that indicates content origin in JPG, PNG and SVG files, as classic visible watermarks added to the image can be easily removed with cropping or editing tools. The open C2PA standard, developed by the Coalition for Content Provenance and Authenticity, aims to carry verifiable information about the source of digital content and the transactions made on it. Thus, instead of placing a permanent mark on the image, information about the origin of the file can be presented through verifiable metadata. Since it is not possible to trust file metadata in texts, the mark must be placed directly in the properties of the produced text.
Anthropic’s approach offers a technical solution to transparency debates at a time when AI content is increasingly taking up space on the internet. The benefit of this for users is that an additional signal of whether a text is Claude output can be obtained when appropriate detection tools are available. On the other hand, the application of watermark in auxiliary processes such as editing and translation requires careful interpretation of the detection result; The sign alone does not prove that the content was written entirely by artificial intelligence. For AI companies, such techniques could help determine the extent to which training data is mixed with AI-generated content in the future. The practical value of the system will become clearer as we see how reliably Anthropic’s detection tools will work and how robust the watermark will remain against different forms of editing.
Join Channel