Anthropic's 'Watermark' Text Adulteration in Claude Is a Perversion of Writing

When I wrote about Anthropic's announcement that all Claude models would begin watermarking their generated text, we were left to speculate on the specifics of the process. Anthropic offered no clear description of how the watermarking would work, despite the provocative title "How Claude Marks AI-Generated Content." Initial speculation suggested the use of invisible Unicode characters within the text, but this turned out to be incorrect.

Instead, Anthropic plans to use a form of steganography, where the choice of words during inference time will leave fingerprints detectable later. This technique, based on word lists labeled "green" and "red," subtly alters the likelihood of specific words being selected at each token generation point. With enough text, it becomes possible to analyze the distribution of these words and determine if they align with the specific watermarking system used by Anthropic's Claude models.