Anthropic further discloses how Claude text watermarks work. This arrangement relates to the EU Code of Conduct on Transparency under the AI Act and requires AI companies to establish mechanisms for identifying AI-generated content.
Watermark embedded in word selection
The company explained in its latest blog that Claude would embed a specific model on some of the low-risk options. When describing weather, for example, choose between multiple words that have a similar meaning. Such models are difficult for ordinary readers to detect, but can be detected by the party holding the corresponding key.
Anthropic indicates that text watermarks will not affect Claude output quality. For the reader, there is no significant difference between a watermark and a watermark-free response.
SynthID-Text will be used for testing
According to the company, Claude used Google DeepMind ' s Syndrid-Text programme, which was introduced in 2024, and plans to introduce watermark testing API. At the same time, Anthropic emphasizes that this is different from the way some AI tests companies judge whether or not to use AI by analysing writing habits and sentence features.
The former is to check for pre-embedded marks, while the latter is to be inferred from text style. The rationale for both approaches is not the same.
The watermark in the code scene is weaker.
As to the user's concern as to “whether or not it can be identified after rewriting”, Anthropic argues that the slight editing rate cannot completely remove the watermark, but if the text is completely rewritten and almost every word replaced, the watermark may disappear.
The company also mentioned whether watermarks could be detected, depending on the length of the text and the extent of the changes, if a text was only made through a Claude proofreading or a small monetation. If the changes were light, the vast majority of the content would still be done by human authors, and there would be very few water stamps.
Watermarks are expected to be less intense than normal text in code-generated scenarios. The reason for this is that the model requires the output of run-off codes and a smaller amount of free-replaceable expression space. However, in positions such as code notes, watermarks are still available and have little effect on actual codes.
Additional information:Anthropic indicates that Claude will not be the only chat robot to use text watermarks, and that other large model developers who sign the same code of conduct for transparency will deploy their programmes.