Anthropic announced that future versions of its Claude models will generate text that includes a watermark. This watermark is intended to help determine the likelihood that Claude was involved in writing the text. The company, along with several other major AI providers, is implementing this change to comply with the EU AI Act. The announcement comes as part of a broader effort to address regulatory requirements for AI-generated content. The watermarking method does not affect the quality or content of the output, and it remains undetectable to readers. The company emphasized that the watermarking process does not alter the randomness of word selection in Claude’s responses. Watermarking uses a key to encode a pattern in the text, which can be detected by those with access to the key. This method allows for verifying whether a text was likely generated by Claude without impacting the final output. Source: anthropic

Watermarking works by using low-stakes word choices to leave a pattern in Claude’s responses. These choices are made at random, but the source of the randomness is different. Instead of using an arbitrary random number generator, the watermarking method uses a key and a few preceding words to determine the next word. This creates a detectable pattern that can be verified with the key. The watermarking process does not add any hidden characters or alter the text’s content. It also does not require extra tokens or increase costs. The watermarking method does not provide identifying information and cannot be traced to a specific person, organization, or chat. The watermark is not specific to Claude and is part of a broader industry effort to comply with the EU AI Act. Source: anthropic

The watermarking method is based on the SynthID-Text approach published by Google DeepMind in a Nature paper in 2024. It belongs to a family of approaches that include a proposal by Scott Aaronson in 2022. The watermarking method only changes the source of randomness used to pick among words. It does not confirm whether the text was human-written or written by a different AI. The effectiveness of watermarking is limited, especially on small samples where there are fewer word choices. The watermark is less effective on factual passages where there are fewer choices to make. It is also less effective on code where an exact output is required. Source: anthropic