Anthropic implements text watermarking in Claude to comply with EU AI Act

Future Claude models will generate watermarked text to indicate AI involvement, a measure implemented to comply with the EU AI Act's August 2 mandate for AI providers. This initiative is being adopted by Anthropic and several other major AI providers.

The watermarking method subtly influences low-stakes word choices made by the large language model, creating an undetectable pattern that can be identified with a key. This process does not alter the meaning or readability of the generated text, ensuring it's indistinguishable to readers.

Watermarking does not affect the content, creativity, or readability of Claude's outputs. It does not add hidden characters, require extra tokens, or increase generation costs, ensuring no practical impact on quality or expense.

Claude uses a version of Google DeepMind's SynthID-Text approach. While effective for assessing the likelihood of Claude's involvement, it cannot confirm human authorship, identify other AIs, and is less effective on short, highly factual, or lightly edited passages due to fewer word choices.

The watermark applies only to words Claude chooses. For lightly edited human text or code requiring exact outputs (e.g., factual statements, functional code), there's little to no impact, as the model makes fewer or no low-stakes choices for watermarking.