Anthropic Unveils How Claude's New Watermarks Operate

Anthropic recently released a blog post addressing fundamental inquiries regarding its implementation of watermarking for text produced by its AI chatbot, Claude. The company sought to clarify aspects such as the operational mechanism of watermarking, its susceptibility to removal through editing, and its implications for generated code.

This initiative has sparked considerable discussion among Claude's user base, following Anthropic's earlier announcement this week that it would be incorporating watermarks. This measure is intended to satisfy the EU AI Act's Transparency Code, which mandates that AI developers implement systems capable of identifying content generated by artificial intelligence.

Reactions on social platforms have been varied; for instance, a Reddit user described the policy as a "conspiracy against innocent Claude users," while another asserted, "The only reason you wouldn’t want this is to lie to people." Furthermore, Business Insider reported that "dozens" of users on X (formerly Twitter) have reportedly canceled their Claude subscriptions in protest of the change.

Anthropic's recent publication commences with a broad explanation of the watermarking principle. It details that during instances of "low-stakes choices"—such as selecting "overcast" versus "grey" to describe weather conditions—Claude is able to embed a subtle pattern within its output. This pattern remains "undetectable to the reader, but is detectable to anyone who has a key that encodes it."

The company affirmed that "Watermarking does not impact the quality of Claude’s output." It further clarified, "To a reader, a watermarked response is indistinguishable from an unwatermarked one."

Providing further specifics, Anthropic stated its intention to adopt the SynthID-Text methodology, which was introduced by the Google DeepMind team in 2024. The company also plans to launch a dedicated watermark detection API. It emphasized that this watermarking technique differs fundamentally from other AI detection methods, such as those employed by companies like Pangram, which analyze stylistic "tells" in writing (e.g., the phrase structure “his isn’t [X], it’s [Y]”) to infer AI generation. Anthropic underscored this distinction, stating, "Picking up on these patterns is fundamentally different from checking for a watermark."

Addressing the possibility of obscuring the watermark through rewriting, Anthropic acknowledged it is feasible. However, the company noted that "light editing probably won’t remove the watermark completely," whereas "a complete rewrite where every word is replaced will" effectively eliminate it.

The company added, "In the latter case, of course, it’s arguable whether the text can any longer be described as AI-generated."

Regarding the detectability of watermarks in text merely proofread or lightly edited by Claude, Anthropic explained that this would be contingent upon "the length of the text and how heavily Claude has edited it." In instances of minimal editing, "nearly all the words" would originate from the human author, leaving "very little (if anything) for the watermark to attach to."

Conversely, code is anticipated to exhibit a less prominent watermark compared to other forms of text. This is because the model's primary objective is to generate functional code, thereby limiting its flexibility to select among various equally valid linguistic alternatives.

Anthropic clarified, "Having said that, in areas where there is an arbitrary choice between particular words or terms within the code, the watermark can be used, such as comments within code." The company concluded, "But by definition, it will have a negligible effect on the actual code produced."

Finally, Anthropic indicated that Claude would not be unique in generating watermarked text, as "other major model developers have signed the same Code of Practice and will be implementing their own watermarks."