Anthropic published a blog post on Friday describing how it plans to watermark text produced by its chatbot Claude and what the practical effects will be. The post addressed basic questions such as how the watermarking operates, whether editing can hide it, and how the approach applies to code.
The company said it is implementing watermarking to comply with the EU AI Act’s Transparency Code, which requires AI providers to deploy systems that make AI-generated content identifiable. The announcement follows earlier statements from Anthropic that it would employ watermarking for this reason.
Watermark concept and detectability
Anthropic outlined the watermarking concept by explaining that, when making “low-stakes choices” — for example choosing between the words “overcast” and “grey” — Claude can introduce a pattern into its outputs that is "undetectable to the reader, but is detectable to anyone who has a key that encodes it." The company stated that "watermarking does not impact the quality of Claude’s output" and that "to a reader, a watermarked response is indistinguishable from an unwatermarked one."
More specifically, Anthropic said it will adopt the SynthID-Text approach described by the Google DeepMind team in 2024, and plans to publish a watermark detection API. Anthropic also emphasized that watermarking is different from AI-detection methods used by firms such as Pangram, which look for stylistic "tells" in writing (for example particular constructions) to infer AI use: "Picking up on these patterns is fundamentally different from checking for a watermark."
Editing, rewrites and when the watermark is lost
Could someone remove a watermark by rewriting or editing? Anthropic acknowledged this is possible but said that "light editing probably won’t remove the watermark completely," whereas "a complete rewrite where every word is replaced will." The company added that in the latter case "it’s arguable whether the text can any longer be described as AI-generated."
Whether a watermark will be detectable in text that Claude merely proofread or lightly edited depends on "the length of the text and how heavily Claude has edited it." If only light edits were made, Anthropic said "nearly all the words" will be from the human author and there will be "very little (if anything) for the watermark to attach to."
Effect on code
Anthropic said generated code should carry less of a watermark than other text, because the model must produce functioning code and so often lacks the freedom to choose among multiple equally valid alternatives. "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," Anthropic noted. "But by definition, it will have a negligible effect on the actual code produced."
User reactions and industry context
Users have debated the move since Anthropic first revealed it would use watermarking. On Reddit some described the change as a conspiracy against Claude users, while other commenters said the only reason to oppose watermarking would be to deceive people. Business Insider reported that dozens of users on X claimed to have canceled their Claude subscriptions for this reason.
Anthropic also said Claude will not be the only chatbot generating watermarked text: "other major model developers have signed the same Code of Practice and will be implementing their own watermarks."
