Anthropic Details How Invisible Watermarks Will Mark Claude's Text

SAN FRANCISCO —

Anthropic laid out, in a blog post published Friday, exactly how it plans to hide digital watermarks inside text written by its Claude chatbot, a move the company says is needed to comply with the European Union's AI Act. The law requires that AI-generated audio, images, video and text carry machine-readable marks so the material can be identified as artificial, according to The Verge.

Anthropic said its system is "a version" of SynthID-Text, an open-source watermarking method built by Google DeepMind, The Verge reported. Google's own Gemini chatbot has used SynthID-Text since 2024. OpenAI has not said whether it will add similar watermarks to ChatGPT, though it will face the same EU requirement, according to The Verge.

How the hidden marks work

Anthropic's explanation centers on what it calls low-stakes word choices, moments when Claude could pick from several roughly equivalent words without changing the meaning of a sentence. The company's example: after "The weather today was cold and," Claude is very unlikely to pick "sugary" but could reasonably choose "overcast" or "grey." Normally that choice is settled by a random number. With watermarking turned on, the choice is instead guided by a secret key combined with the preceding words, leaving a pattern that a reader can't spot but that becomes visible to anyone holding the key, Anthropic said, according to The Verge and TechCrunch.

Anthropic told TechCrunch it plans to release a detection tool through an API and said the approach is not the same as services such as Pangram, which hunt for telltale phrasing patterns in AI writing rather than checking for an embedded key. The company also addressed editing: light edits to watermarked text probably won't erase the pattern, but replacing every word in a rewrite will, Anthropic said, adding that at that point it's debatable whether the text still counts as AI-generated. Computer code, Anthropic said, will carry a much weaker watermark than ordinary prose because working code leaves the model far less room for arbitrary word choices, though comments within code can still carry a mark.

Backlash from some Claude users

The explanation followed several days of complaints after Anthropic first disclosed the watermarking plan, according to TechCrunch. On Reddit, one user described the move as a conspiracy against Claude subscribers, while another argued anyone opposed to it must be trying to hide AI use, TechCrunch reported. Business Insider reported that dozens of users on X said they had canceled their Claude subscriptions in response.

Anthropic noted that other AI developers have signed on to the same EU Code of Practice and will roll out their own watermarking systems, meaning Claude will not be the only chatbot marking its text this way, according to TechCrunch.