Anthropic explains how Claude's new watermarks work

Anthropic published a blog post Friday answering basic questions about how it will watermark text generated by its chatbot Claude — how the watermarking works, whether it can be hidden through editing, and how it affects code. The post follows days of user debate after Anthropic revealed it would add the watermarking to comply with the EU AI Act’s Transparency Code, which requires AI companies to use systems that make it possible to identify AI-generated content. Some Claude users pushed back sharply: one Reddit poster called the move a conspiracy against innocent users, while Business Insider reported that “dozens” of users on X claimed to cancel their Claude subscriptions over it.

According to Anthropic, watermarking works by having Claude favor certain words over equally valid alternatives in “low-stakes choices” — such as “overcast” versus “grey” — creating a pattern invisible to readers but detectable with a key that decodes it. “Watermarking does not impact the quality of Claude’s output,” the company said. “To a reader, a watermarked response is indistinguishable from an unwatermarked one.” Anthropic is using the SynthID-Text approach that Google DeepMind outlined in 2024, and plans to release a watermark detection API; it said the approach is fundamentally different from AI-detection tools like Pangram, which instead look for stylistic “tells” in writing.

Rewriting text can remove the watermark, Anthropic acknowledged, but light editing probably won’t erase it completely, while a complete rewrite where every word is replaced will. Whether a watermark survives in text Claude only proofread or lightly edited depends on the length of the text and how heavily it was changed — if editing is light, there may be little for the watermark to attach to, since most of the words remain the human author’s own. Code carries less of a watermark than prose, Anthropic said, since the model usually has no freedom to choose between equally valid working solutions, though it can still watermark arbitrary choices like comments. Anthropic added that Claude won’t be the only chatbot producing watermarked text, since other major model developers have signed the same EU Code of Practice and will roll out their own watermarks.

Why it matters

Anthropic’s explainer is as much a trust exercise as a technical one — an attempt to show skeptical users that EU-mandated transparency doesn’t have to come at the cost of Claude’s usefulness, even as rivals prepare to roll out their own watermarks under the same rules.