Anthropic has revealed additional details about the watermarking technology it is preparing to use with texts produced by its AI robot Claude, explaining how the technology works, whether it can be removed from modified texts, and its potential impact on code.

This step comes after the company announced earlier this week that it would start watermarking the texts generated by the Claude models as part of its commitment to the EU AI Act and its associated transparency rules, which require systems to help identify AI-generated content.

How does the watermark work in Claude?

Anthropic explained that the watermarking technology relies on exploiting what it described as "low-importance choices" during text generation. When Claude chooses between words with similar meanings, such as using "cloudy" instead of "gray" to describe the weather, the model can create a specific pattern within the text that is not noticeable to the reader but becomes detectable using a decryption key or key containing the necessary information to recognize the watermark. The company confirms that this process will not affect the quality of the generated text; the watermarked text will appear almost identical to ordinary unremarkable text for readers.

Anthropic uses SynthID-Text technology

The company confirmed that it will rely on the SynthID-Text technology presented by Google DeepMind in 2024. Anthropic also plans to launch a custom API for watermark detection, which could allow various entities to verify texts and determine if they were generated by Claude. This technology differs from AI-written text detection tools that search for distinctive linguistic patterns within the writing, such as repeated structures or specific styles. According to Anthropic, analyzing these linguistic features fundamentally differs from searching for a coded watermark embedded within the text itself.

Can the watermark be removed by paraphrasing the text?

Yes, but it depends on the extent of the modification. Anthropic says that simple modifications to the text are unlikely to completely remove the watermark. However, rewriting the entire text, replacing nearly every word, can eliminate the watermark. The company notes an important point here: in the case of a complete rewrite, one might question whether it is appropriate to describe the new version as AI-generated text since all the words have been paraphrased.

What happens if Claude is only used for proofreading?

If Claude is used to conduct a linguistic review or make limited edits on text written by a human, the watermark's ability to appear will depend on the length of the text and the extent of the modifications made by Claude. If the changes are minimal, most words will still be from the human author, meaning there won't be much text for the watermark to link to. In other words, passing human-written text through Claude for proofreading doesn’t automatically mean that the entire text will carry a strong, detectable watermark.

Anthropic says the watermark would be weaker in code compared to regular text because writing code doesn't always give the model freedom of choice between many different words or phrasings; in many cases, the model must produce code that works correctly, which reduces the options available for inserting a watermark without affecting functionality. However, this doesn’t mean there will be no watermark at all. In parts where there are random or multiple choices within the code, such as comments and the selection of certain terms, a watermark can be used. Nonetheless, Anthropic emphasizes that the impact on actual code would be very small.

Claude isn't the only one who will use watermarks

Anthropic also indicated that the use of watermarks won’t be limited to Claude, saying that other major AI model developers have signed on to the same set of practices related to transparency and are expected to implement their own systems for watermarking content produced by their models. This means that over time, watermarks could become one of the key tools used by AI companies to help distinguish between text written by humans and that generated by AI models.

Why is the move controversial?

Anthropic’s plan has sparked controversy among some Claude users since it was announced, with some seeing the watermark as a way for them to be tracked or their use of AI detected. Others, however, consider its presence necessary to increase transparency and prevent the misleading use of text generated by AI. The move comes at a time when it is becoming increasingly difficult to distinguish between human-written text and that produced by AI models, prompting tech companies to seek more reliable technical means to verify the source of content.