Watermarking AI text: Anthropic's plan to comply with the EU

What happened?

Anthropic, the company behind the Claude language model, has announced a proposal to incorporate a watermark into texts generated by its systems. The technique, inspired by the work of Google DeepMind (SynthID-Text), involves influencing the choice of 'inconsequential' words during the generation process, so that a statistical signature detectable with a digital key is introduced. This measure seeks to align with the transparency requirements of the EU AI Act, which requires AI providers to label synthetic content so that users can identify it.

Why does it matter?

Anthropic's initiative is relevant because it addresses one of the most pressing challenges of the generative AI era: the difficulty of distinguishing between human text and machine-produced text. Until now, attempts at watermarking text have been limited or easy to circumvent. Anthropic's proposal, while not infallible, represents a step toward transparency and accountability in the industry. Furthermore, it sets a precedent for the practical implementation of watermarking techniques in large-scale language models, which could influence how other companies approach regulatory compliance.

How does the technique work?

The method is based on the probabilistic nature of language models. During text generation, the model predicts the next word based on a probability distribution. Anthropic proposes subtly modifying that distribution to favor certain 'inconsequential' words (such as 'gray' instead of 'cold' in a meteorological context) that do not alter the meaning of the content. These modifications create a statistical pattern that can be detected using a secret key, thus allowing it to be identified if a text was generated by Claude.

According to the company, the impact on quality is nil: internal tests show no differences in content, creativity, or readability, and a study with human evaluators found no perceptible differences between texts with and without a watermark. However, the technique has limitations: it applies only to low-risk passages, not to factual texts or code, and can be removed with a complete rewrite.

Consequences for companies and users

For companies using generative AI, this watermark could become a de facto standard if Anthropic implements it in its models. This would imply that content generated by Claude would carry an invisible label that could be verified by third parties, adding a layer of traceability. For users, it means greater transparency: they could know if a text has been written by a machine, which is crucial in contexts such as journalism, academia, or marketing.

However, the real effectiveness will depend on widespread adoption and the ability of malicious actors to bypass it. As Anthropic itself admits, the watermark is semi-effective: a light edit does not remove it, but a total rewrite does. This limits its utility in scenarios where the text is substantially transformed.

Comparison with previous attempts

This is not the first attempt to watermark AI-generated text. OpenAI and others have explored similar techniques, but they have often been criticized for their fragility or for affecting quality. Anthropic's approach, based on inconsequential words, is more sophisticated and less intrusive than others, but it is still far from being a definitive solution. The history of digital rights management (DRM) offers a parallel: protection systems always find adversaries willing to break them, and the war between creators and evaders is continuous.

What readers should know

It is important to understand that this technology is not infallible nor does it claim to be. Anthropic presents it as a tool to promote transparency, not as a mechanism for absolute control. Readers should be aware that the watermark does not reveal personal information or identify the author; it only indicates that the text was generated by Claude at some point in its creation. Furthermore, the technique is still in the proposal phase and has not been implemented in production, so its real effectiveness remains to be seen.

Textual watermarking is a necessary step toward transparency in the AI era, but it is not a silver bullet: it requires a verification ecosystem and the collaboration of the entire industry.

Regulatory implications

The EU AI Act is the first comprehensive regulatory framework for AI, and its requirement to label synthetic content is a milestone. Anthropic's proposal could serve as a model for other companies seeking to comply with this regulation. However, the lack of standardization among providers could create confusion: if each company implements its own watermarking system, interoperability will be a challenge. Bodies such as the EU or ISO may have to intervene to define common standards.

The future of authenticity in the AI era

The battle for content authenticity is one of the great issues of our time. The watermark is just one piece of the puzzle; other technologies, such as cryptographic verification or the use of blockchain, could complement it. Anthropic's proposal is a significant advance, but also a reminder that a definitive solution does not yet exist. The industry must continue to innovate and collaborate to protect the integrity of information in a world where AI can imitate human writing with astonishing precision.