Anthropic details how Claude watermarks text to comply with EU AI Act

On Friday, Anthropic published a blog post to clarify how it will watermark text produced by Claude, the company's AI chatbot. The watermarking initiative was announced earlier in the week and is required by the EU AI Act's Transparency Code, which obligates AI companies to implement systems that make AI-generated content identifiable.

The technical approach uses SynthID-Text, a method outlined by the Google DeepMind team in 2024. The watermark works by encoding a pattern into Claude's responses when the model makes "low-stakes choices"—such as selecting between synonymous words like "overcast" or "grey" to describe weather. The encoded pattern is imperceptible to human readers but detectable by anyone holding a decryption key. Anthropic 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." The company plans to release a watermark detection API. Anthropic also clarified that watermarking is fundamentally different from AI detection methods offered by companies like Pangram, which search for telltale stylistic patterns (such as the phrase "it isn't [X], it's [Y]") in the writing itself.

Anthropicaddressed several practical concerns. Regarding editing, the company acknowledged that light editing will probably not remove the watermark entirely, but a complete rewrite in which every word is replaced will eliminate it—though Anthropic noted that such extensive rewriting raises the question of whether the text is still meaningfully AI-generated. For text that has been proofread or lightly edited by Claude, the detectability of the watermark depends on "the length of the text and how heavily Claude has edited it." If only lightly edited, "nearly all the words" will have come from the human author, leaving "very little (if anything) for the watermark to attach to." Code will experience minimal watermarking because the model requires freedom only in areas of arbitrary choice—such as comments—and the watermark "will have a negligible effect on the actual code produced."

The announcement comes as Claude users have engaged in heated debate over the watermarking mandate. On Reddit, some users characterized the move as a conspiracy against innocent Claude subscribers, while others argued that refusing watermarking would be dishonest. Business Insider reported that "dozens" of users on X have claimed to cancel their Claude subscriptions in protest. Anthropic also noted that the watermarking requirement is not unique to Claude: "Other major model developers have signed the same Code of Practice and will be implementing their own watermarks."