Anthropic announced last week that it would be adding watermarks to all text generated using its Claude series of models, but has since clarified that they will not be recognisable by regular readers of the content.
The change has been effective since 2 August in order to comply with the EU AI act. On Friday, Anthropic released more details about how its chosen watermarking, which will allow machines to predict with greater accuracy whether text was written using Claude models.
The company first stressed that the chosen watermark has no practical impact on the models’ output, and that the difference between watermarked and unwatermarked text is not distinguishable to readers.
It also assured users that the technique will not increase cost or token usage for models, and will not provide information that allows text to be traced back to a particular individual, organisation, or chat.
Its functionality comes from altering the fundamental basis of Claude’s weightings. In broad terms, a large language model such as Claude generates text by predicting the most likely next word in a sequence. Anthropic used the example of the sentence “The weather today was cold and …”.
In this case, there are only a handful of words that are likely to come next, such as “overcast” or “grey”. For a user, there is little difference between the two words, so Claude chooses a word based on a random number.
The AI giant’s chosen watermarking technique is a version of the SynthID-Text approach
pioneered by Google DeepMind in 2024. This adds a watermark to generated text by replacing the random number with a key of sorts. This means that anyone with access to the key would be able to detect certain choices made over a long enough piece of text as being likely made by Claude.
There are, however, clear limitations with this low-interference approach. In shorter text, there may not be enough words that fit a watermarked pattern for it to be detectible. Similarly, there are likely to be fewer watermarks in technical text or code, as there are fewer “equally good” next words.
Even for text where there are watermarks, applying the key can only tell a user how likely it is that Claude was involved in some part of the text generation process, rather than what its role was. This means a key cannot provide certainty on whether a given text was generated by Claude, even in part.
For accessing the key, Anthropic has announced that it will launch a watermark detection API “soon”.
