Anthropic Reveals Details of the New Watermark in Claude's Text

Anthropic's decision to add a digital watermark to the text generated by its chatbot Claude sparked a wave of debate and controversy among its users, to the extent that some announced they were canceling their subscriptions in protest. In an attempt to clarify the picture, the company published a detailed blog post answering the most prominent questions: How does this watermark work? Can it be hidden through editing? And what is its impact on programming code?

Why did Anthropic resort to watermarking?

The direct reason goes back to compliance with the requirements of the European «AI Act», specifically the transparency clause that obligates AI companies to adopt systems that allow the identification of machine-generated content and distinguish it from content written by humans. Anthropic confirmed that it is not alone in this approach, as other companies developing large models signed the same blog post and will apply their own watermarks soon.

How does the watermark actually work?

The company explains that the idea is based on exploiting the «low-importance choices» that the model makes while generating text. When there is more than one equivalent option to express the same idea — such as choosing between two synonyms to describe the weather — the model embeds a hidden pattern in its wording. This pattern cannot be noticed by an ordinary reader, but it is detectable using a dedicated digital key that decodes it.

Anthropic emphasized a fundamental point: the watermark has absolutely no effect on the quality of the output, as watermarked text does not differ from the reader's perspective from any other text. The company relied on the «SynthID-Text» technology developed by the Google DeepMind team in 2024, with a plan to launch a dedicated API to detect these watermarks.

Is it different from AI detection tools?

Anthropic clearly distinguished between its approach and traditional tools for detecting machine-generated content, such as those adopted by companies like «Pangram». These tools look for «stylistic markers» in the writing style that indicate the use of AI, while the watermark relies on a cryptographic key embedded within the text itself. According to the company, capturing stylistic patterns is fundamentally different from verifying the presence of an encrypted watermark.

Can it be hidden through editing?

This question preoccupied many users, and Anthropic's answer was realistic. Minor edits to the text will most likely not remove the watermark entirely, while a complete rewrite in which every word is replaced is enough to erase it completely. However, the company pointed to a logical paradox: if the text is completely rephrased, it is difficult in the first place to describe it as «AI-generated».

As for texts that Claude has merely proofread or edited, the appearance of the watermark depends on two factors:

• Text length: The greater the number of words, the higher the likelihood that the watermark is associated with it.
• Extent of editing: If the editing is light and the text retains most of the human author's words, the system will not find enough words to attach the watermark to.

What about programming code?

Anthropic explained that code will carry far less watermarking compared to regular text, because the model is obligated to produce code that works correctly, which narrows its margin for choosing between equivalent alternatives. Nevertheless, the watermark can be used in places that allow free choice between certain terms, such as annotations within the code. By its nature, its impact on the actual code will be negligible to the point of being almost insignificant.

Conclusion

This approach represents a step within a broader regulatory trend that seeks to enhance transparency in the world of generative AI. Despite the concern expressed by some users about the possibility of their use of the tool being «exposed» in their work or studies, Anthropic insists that the watermark is a tool for credibility, not surveillance, and does not detract from the quality of the content or the user experience. With other developers joining the same path, it seems that watermarks are on their way to becoming a standard norm in the AI industry.

✦ بقلم فريق دروب أيديا