Anthropic Develops New Method to Watermark AI-Generated Text
Claude's outputs will be marked without altering their meaning.
1Anthropic's New Watermarking Method
Anthropic has developed a new watermarking method to mark text generated by its Claude AI model. This technique involves altering inconsequential word choices to create a detectable signature. Traditional watermarks are used on currency or official documents, while digital watermarks add identifiers to electronic data. Anthropic's approach builds on techniques proposed in Google DeepMind's SynthID-Text paper.
2How Large Language Models Work
Large language models operate by predicting the next word in a sequence. For example, when completing a sentence like 'The weather today was cold and…', a model like Claude might choose words like 'cold' or 'gray'. However, it is unlikely to select a word like 'sugary'. Despite this theory, Claude Opus 4.8 can sometimes produce overly elaborate responses.
3How the Watermark is Generated
The watermark is created by deviating from the predicted word to another choice. A different source of randomness is used, which can be detected with a digital key. According to Google DeepMind researchers, this technique introduces a statistical signature into the generated text. This signature helps determine whether the text was produced by a watermarked language model.
4Impact of Watermarking on Text
Anthropic claims that watermarking does not alter the meaning of the text. Internal tests showed no impact on Claude's content, creativity, or readability. In a controlled study, human evaluators found no quality difference between watermarked and non-watermarked responses.
5Limitations of Watermarking
Watermarking is not applied to critical text. Anthropic explains that factual passages have fewer word choices that won't reduce accuracy, making watermarking sparser. Similarly, the watermarking algorithm cannot arbitrarily change method names in code.
6Literature and Watermarking
In literature, the idea that some words are interchangeable could spark debate. While imagining Claude producing sentences like 'It was the best of times, it was the worst of times…' is intriguing, those passing off AI-generated text as serious writing may face social backlash.
7Advantages of Watermarking
Anthropic's watermarking method is minimally intrusive. It does not include personal information and only indicates Claude's involvement in the text's creation. Additionally, the watermark is not fully effective. Light editing won't remove it, but a complete rewrite will.
8Legal Compliance and Cost
Anthropic developed this watermarking method for legal compliance. The company emphasizes that it has a negligible impact on model speed and does not increase costs since it produces no extra tokens. This approach reflects Anthropic's focus on performative compliance.