Anthropic's artificial intelligence, known as Claude, has taken a step forward in content transparency. The company has announced that all models released from August 2nd onward will incorporate an invisible markup system in the text they generate. This measure, which will be implemented globally, aims to align with European Union requirements regarding the identification of synthetic content.
The main objective is for any text produced by Claude to be identifiable as such, without the user noticing any difference in quality or readability. The watermark is not visible to the naked eye, but remains embedded in the text even when it is copied and pasted on other platforms. This initiative responds to the Code of Good Practice on Transparency of AI-Generated Content, signed by more than 190 organizations, and to the EU AI Law, which, since August, has required the watermarking of this type of content.
Anthropic's decision is no coincidence. European regulations require AI companies to label synthetic content so users can distinguish it from human-created content. Article 50.2 of the AI ​​Act states that providers must ensure their systems' outputs are labeled in a machine-readable format. Anthropic, along with other major tech companies like OpenAI, Google, Meta, and Microsoft, has signed this transparency code, although companies like Elon Musk's xAI have not yet joined.
The company explained that the watermarking is done in two ways: for images and graphics, metadata based on the C2PA (Coalition for Content Provenance and Authenticity) standard is used; for text, an invisible watermark is employed that is integrated into the generation process itself. This watermark does not alter the meaning or quality of the responses and is transmitted along with the text when it is copied, exported, or shared via an API.
The system is based on the sequence of tokens generated by the model. Instead of adding hidden characters or visible metadata, the watermark is encoded in the choice of words when several equally valid options are available. For example, given the phrase "the sky was…", the model might choose "cloudy" or "gray". With the watermark, this choice is not random, but rather depends on a cryptographic key and the preceding words. Thus, whoever possesses the key can statistically determine whether the text was generated by Claude.
This method, already used in other systems like Google's SynthID, adds neither cost nor latency to the process. Watermark detection is not foolproof : if the text is edited, paraphrased, translated, or mixed with other content, the watermark may disappear. Furthermore, the signal is weaker in very short fragments or in code. Anthropic acknowledges that the absence of a watermark does not mean the content was not generated by AI, and that the presence of a watermark is not conclusive regarding actual authorship.
The company has specified that the watermark applies to all models released since August 2nd, and that earlier models will receive the update gradually. The measure is global, not just for Europe , and affects all Claude products, including the API, Claude Code, Claude Cowork, and third-party platforms such as AWS infrastructure , Google Cloud, and Microsoft Foundry. However, public detection is not yet available: Anthropic has announced that it will release a tool in a few weeks so that users and third parties can check if a text contains the watermark.
Another significant limitation is that the watermark doesn't identify the user or the original conversation. It also fails to distinguish whether the text was created from scratch or if Claude simply helped correct, translate, or summarize existing content. This has drawn criticism, as European regulations explicitly exclude standard editing tasks, yet Anthropic applies the watermark to all uses. Despite these shortcomings, the measure represents progress in the fight against disinformation and the misuse of AI.
Anthropic's decision is part of a broader industry movement. OpenAI, Google, Meta, Microsoft, and Mistral have also signed the Transparency Code and are implementing similar systems. However, the lack of a unified standard and technical limitations mean that detection is not perfect. For example, taking a screenshot of an image can remove C2PA metadata, and completely rewriting text can erase the watermark.
In educational and professional settings, these tagging tools could be useful for identifying AI-generated work, but experts caution that they are not a definitive solution. Anthropic itself emphasizes that the tag is not proof of authorship, but rather an indication of probability. As the technology evolves, we will likely see more robust systems, but for now, transparency remains a challenge.
In short, the addition of invisible watermarks in Claude represents an important step toward complying with European regulations and providing users with more information about the origin of content. Although the system has limitations, it is a significant step in the right direction. The tech community will be watching closely the detection tool that Anthropic promises to launch soon, as well as how other companies adapt to this new reality.