Future Claude models will generate text that contains a watermark.
This feature complies with the EU AI Act, and will help to determine the likelihood that Claude was involved in writing the text.
All artificial intelligence (AI) vendors will have to provide a similar feature under the terms of the act.
Claude’s text watermark is a version of the SynthID-Text approach published by Google DeepMind in a Nature paper in 2024.
Anthropic says it won’t have any practical impact on the quality or content of Claude’s outputs, and won’t use any more tokens or add cost to users.
In addition, the difference between watermarked and un-watermarked text will not be distinguishable to readers; nothing will be added to the text; and there are no hidden characters.
Claude’s watermarking carries no identifying information and can’t be traced to a specific person, organisation or chat.
Watermarking uses low-stakes choices in text generation to leave a pattern in Claude’s responses. Instead of using an arbitrary random number generator to pick the next word, watermarking uses the key and a few words that come before to settle what word the model should pick.
According to Anthropic, the words that Claude picks are random, as they are now, but the sequence can be checked to see if it’s consistent with the choices Claude would make if it was using the key. If it is, one can assign a probability that the text was generated by Claude.
The company says that internal testing demonstrates no impact of watermarking on the content, level of creativity or readability of Claude’s text. In the SynthID-Text paper, which introduced the technique uses, Google DeepMind tested the impact by serving a model that used watermarking to a portion of their Gemini traffic and comparing thumbs-up and thumbs-down ratings. They found no statistically significant differences from the unwatermarked model. And in a controlled study, human raters comparing watermarked and unwatermarked answers side-by-side saw no difference in quality.
The watermark won’t be applied to code, as this requires exact output rather than choosing from a number of options.