Anthropic says text watermarking scheme relies on inconsequential words

In an effort to adhere to the EU AI Act, Anthropic has introduced a text watermarking scheme for its Claude model. This technique aims to subtly modify the AI's word choices in a way that can be detected as AI-generated. Traditional watermarks, like those on currency or official documents, are patterns or images, but in the digital realm, they can encompass various methods for adding an identifier to electronic data.

Anthropic's approach revolves around altering inconsequential word choices made by its language models, a concept demonstrated in Google DeepMind's SynthID-Text paper. Large language models generate text by predicting the next word in a sequence. For instance, when prompted with "The weather today was cold and…", Claude might suggest "cold" or "gray", but is less likely to choose "sugary".

However, Claude Opus 4.8 exhibited an overzealous response: "...crisp, the kind of cold that nips at your fingertips and turns your breath to little clouds...". To generate the watermark, Anthropic encourages deviations from the predicted word with a different source of randomness, detectable with a digital key. When asked about the impact on meaning, Anthropic stated that internal testing revealed no difference in the content, creativity, or readability of Claude's text post-watermarking.

However, they emphasized that this technique should not be applied to consequential or factual passages, as the algorithm cannot replace method names or crucial details. While some may view this as performative compliance, Anthropic maintains that the watermark adds negligible speed and cost to the model.