Anthropic says its AI watermark will be applied globally to meet EU regulations
Anthropic announced that it will embed a watermark in Claude-generated text worldwide to comply with the European Union's AI Act, explaining how the technology works and its limitations.
Anthropic said it will embed a watermark in the output of its Claude models worldwide to meet the requirements of the European Union's AI Act. The company explained that the watermark will be added to new models at launch and later to older versions, because it does not yet have a reliable way to limit the feature to specific regions.
The watermark is based on a technique described in a 2024 Google DeepMind paper. It relies on subtle patterns in word-choice decisions made by the model, which can be detected with a secret "key" that Anthropic plans to release via an API for users and third parties. The company said the watermark does not involve hidden characters or visible changes and will not be noticeable to readers.
Anthropic noted that the key will only indicate the likelihood that Claude contributed to a text, not that any AI was involved. Longer passages are easier for the key to assess, while factual or code outputs may receive fewer marks to avoid compromising accuracy. The firm clarified that the watermark does not affect users' ownership rights to the content.
"Light editing probably won't remove the watermark completely; a complete rewrite where every word is replaced will," the blog post states.
Anthropic said it has not observed an increase in subscription cancellations since announcing the watermark, and it expects other AI providers to adopt similar measures.