Anthropic published a blog post Friday answering basic questions about how it will watermark text generated by Claude. The move, announced earlier in the week, is meant to comply with the EU AI Act‘s Transparency Code, which requires AI companies to identify AI-generated content. The announcement stirred debate: on Reddit, one poster called it a “conspiracy against innocent Claude users,” while another said, “The only reason you wouldn’t want this is to lie to people.” Business Insider reported that “dozens” of X users claimed to cancel their Claude subscriptions.

The post starts with a general overview of watermarking. When Claude makes low-stakes choices — such as picking between “overcast” and “grey” to describe weather — it can create a pattern in its responses that is undetectable to the reader but detectable to anyone holding an encoding key. Anthropic says watermarking does not impact output quality: “To a reader, a watermarked response is indistinguishable from an unwatermarked one.” The company will use the SynthID-Text approach outlined by Google DeepMind in 2024 and plans to release a watermark detection API.

Anthropic distinguishes watermarking from AI detection tools like Pangram that look for stylistic “tells” (e.g., the phrase “this isn’t [X], it’s [Y]”). “Picking up on these patterns is fundamentally different from checking for a watermark.”

Regarding evasion, Anthropic says light editing probably won’t remove the watermark completely, while a complete rewrite replacing every word will. In that case, the company argues, it’s debatable whether the text can still be called AI-generated. For text Claude only proofreads or edits, detectability depends on the text’s length and how heavily Claude edited it; after light editing, nearly all words are the human author’s, leaving little for the watermark to attach to.

Code will carry less watermarking because Claude must produce working code and can’t freely choose among equally valid options. Still, where arbitrary choices exist — like comments or particular terms — the watermark can be used. Anthropic says this will have a negligible effect on the actual code produced.

Finally, Anthropic notes that Claude won’t be the only watermarked chatbot: “Other major model developers have signed the same Code of Practice and will be implementing their own watermarks.”