Anthropic Details How Claude's Text Watermarking Will Actually Work

Anthropic published a blog post on Friday explaining the technical mechanics behind Claude's upcoming text watermarking system, addressing user concerns that emerged after the company announced the feature earlier in the week.

The watermarking is required to comply with the European Union's AI Act Transparency Code, which mandates that AI companies deploy systems capable of identifying AI-generated content.

How it works

According to Anthropic, the watermark operates by embedding patterns into "low-stakes choices" Claude makes during text generation — for instance, choosing between synonyms like "overcast" and "grey" to describe weather. The company says this creates a pattern that is "undetectable to the reader, but is detectable to anyone who has a key that encodes it."

"Watermarking does not impact the quality of Claude's output," Anthropic stated, adding that a watermarked response is indistinguishable from an unwatermarked one to human readers.

Anthropic confirmed it will use the SynthID-Text method, an approach Google DeepMind published in 2024. The company also plans to release a watermark detection API, though pricing and a launch date were not disclosed.

Anthropic drew a distinction between its watermarking system and AI-detection tools from companies like Pangram, which scan for stylistic "tells" in writing — such as the construction "this isn't X, it's Y" — rather than checking for an embedded signal. "Picking up on these patterns is fundamentally different from checking for a watermark," the company said.

Can the watermark be removed?

Anthropic acknowledged that editing text can strip the watermark, but the effect depends on the extent of the edit. "Light editing probably won't remove the watermark completely," the company said, while "a complete rewrite where every word is replaced will." Anthropic added that in the case of a full rewrite, "it's arguable whether the text can any longer be described as AI-generated" at all.

For text that Claude only proofreads or lightly edits, the watermark's presence depends on "the length of the text and how heavily Claude has edited it." If edits are minimal, Anthropic said, "there's very little (if anything) for the watermark to attach to," since most of the words originated with the human author.

Impact on code

Anthropic said generated code will carry a weaker watermark signal than prose, because the model has less freedom to choose between arbitrary options when producing functional code. The exception is comments, where the model can choose between equally valid wording, allowing some watermark signal to persist there. The company said this will have "a negligible effect on the actual code produced."

Anthropic noted that Claude will not be alone: other major model developers who signed the EU's Code of Practice are also expected to implement their own watermarking systems.

Backlash

The announcement has drawn criticism on social media. Reddit users have debated the move, with one commenter framing it as a conspiracy against users and another arguing the only reason to oppose it is "to lie to people." Business Insider reported that "dozens" of users on X claimed to have canceled their Claude subscriptions over the change.

What this means

This is a compliance-driven feature rollout, not a new model or architecture change — Anthropic is retrofitting an existing detection mechanism (SynthID-Text) onto Claude's output to satisfy EU regulatory requirements, not altering how the model itself is trained or weighted. The practical impact on typical users appears minimal, since watermarking targets long-form, mostly-untouched AI text rather than code or edited work. The bigger open question is adoption: if watermark detection APIs stay siloed per-company rather than becoming interoperable, their usefulness for third parties trying to verify AI content across different model providers will be limited.