Anthropic is implementing text watermarking in Claude, citing compliance with the EU AI Act. The company claims the system will not add text or hidden characters, require extra tokens, increase costs, or allow watermarked content to be traced to a specific person, organization, or chat.
Anthropic claims the watermark will be invisible
The company’s FAQ describes the watermark as a machine-detectable pattern embedded in the model’s word choices. Anthropic asserts that the difference between watermarked and unwatermarked text will not be distinguishable to readers and will have “no practical impact” on Claude’s output quality or content.
The mechanism appears to rely on how language models select text. Claude generates responses one word at a time from a list of plausible candidates. Given the phrase “The weather today was cold and…,” for example, “sugary” would be unlikely, while several other words could plausibly follow. Anthropic’s explanation states that, when multiple choices would preserve the meaning of a sentence, a random number can determine which word is selected.
A watermarking system can influence those selections in a statistically detectable way without adding visible markers. A demonstration labeled “The story (watermarked)” reports that 298 of 476 selected word choices were heads in a coin-flip analogy, or 62.6%, compared with the 50% expected by chance. It assigns the result a z-score of 5.50 and claims that a writer without the detection key would have a one-in-48-million chance of scoring as highly.
The demonstration highlights ordinary words and phrases such as “was,” “a little girl,” “loved,” “the park,” “One day,” and “gave her.” That design is intended to make the watermark difficult for readers to identify while allowing a detector with the relevant key to assess whether a passage likely came from Claude.
A browser-based project also claims to implement the SynthID algorithm used for Claude’s text watermarking, using a small language model that runs in the browser.
Critics dispute the quality claims
The announcement drew criticism from users who argue that selecting different words necessarily changes the output, even if the overall meaning remains similar.
One response directly challenged Anthropic’s statement that the watermark would not affect quality or content, arguing that the system “change[s] the wording.” Another user pointed to the FAQ’s more qualified language, including references to sentences being “largely the same,” “no statistically significant differences,” a “negligible effect on the actual code produced,” and a “negligible impact on the speed of models.”
Those objections focus on areas where small wording changes can matter, including legal documents, technical writing, and code. A separate reply asked whether the watermarking applies to code. The EU rules summary provided with the announcement lists source code among limited exceptions to the text-marking requirement, while also describing Anthropic’s claims about a negligible effect on code generation.
Several users also objected to the policy on ownership and consent grounds. Critics argued that people may submit human-written material for grammar correction or reformatting and then receive output that could be identified as AI-generated. Others accused Anthropic of applying a watermark to work based on material gathered from the internet.
The EU rules are technical rather than visible labels
The EU text rules described in the supporting material require providers of generative AI systems, including GPAI systems, to mark synthetic text outputs in a machine-readable format so they can be detected as AI-generated or manipulated. The requirement is described as technical rather than as a visible label intended for ordinary readers.
The summary lists exceptions for short outputs, source code, pure machine-to-machine use, and standard editing. It also states that published AI-generated or manipulated text concerning matters of public interest—such as politics, public health, administration, or the environment—must be clearly labeled by deployers. That labeling is not required when genuine human review or editorial control takes place and a person or organization accepts editorial responsibility.
Anthropic claims that other major model developers signed the same Code of Practice and will also implement watermarking. The company did not identify those developers in its announcement, prompting at least one response to ask who had signed the agreement.
The rollout also triggered complaints from users who questioned why a San Francisco-based company would apply rules associated with European law to users elsewhere. One response characterized the decision as complying with foreign regulation beyond what the law requires, while another called for canceling Claude subscriptions.
Anthropic’s FAQ presents watermarking as an imperceptible technical safeguard. The reaction suggests that users are evaluating it as a change to the model’s behavior, regardless of whether the resulting text remains readable and semantically similar.
Source: Anthropic on X
