Anthropic’s Claude Watermarks Spark Rapid Development of Removal Tools

Anthropic announced that its Claude large-language models will embed an imperceptible, machine-readable watermark in every piece of text they generate to satisfy the EU AI Act, which requires labeling synthetic content and imposes fines of up to 3 percent of annual turnover for non-compliance.

Wired Magazine

The AI Act, adopted earlier this month, obliges providers to create “marking” mechanisms for generated content. Anthropic’s watermark subtly influences word choice during generation, leaving a pattern detectable only with the appropriate key. The company says the watermark is solely for testing whether Claude produced or processed the content and does not affect ownership or user rights.

ZDNet

• Within four hours of the announcement, developer Guillaume Meyer released an open-source “watermark remover” on GitHub, which has been bookmarked over 20,000 times on X and attracted more than 100 contributors.

Wired Magazine

• A related GitHub project, Watermarks Remover, has earned more than 15,000 stars and thousands of forks.

Gizmodo

• Google search interest for “AI watermark remover” rose 60 percent week over week, according to Business Insider.

Gizmodo

Anthropic explained that the watermark works by replacing the model’s usual random word-selection process with one that incorporates a secret key and a few preceding words, creating a detectable pattern without changing overall randomness.

ZDNet

• Guillaume Meyer, creator of the removal tool, argues that the watermark “is a very significant distinction” from content attribution and opposes the technique.

Gizmodo

• John Gruber, co-creator of Markdown, says the watermark’s influence on word choice is “unacceptable.”

ZDNet

• Sara Simeone, founder of NoCodeLab.ai, notes that AI labs have long scraped creators’ content without consent and now require attribution that could further monetize that scraped material.

ZDNet

• Anthropic’s blog states the watermark is “machine-readable” but does not yet provide a public detector, leaving users unable to verify marked text.

Wired Magazine

• While the EU AI Act mandates marking, it does not prohibit removal, creating a regulatory gap that developers are already exploiting.

Gizmodo

• “I am all for content attribution,” — Guillaume Meyer

Gizmodo

• “Instead of using an arbitrary random number generator to pick the next word, watermarking uses the key and a few words that come before to settle what word the model should pick,” — Anthropic

ZDNet

Anthropic plans to apply the watermark to text from several Claude models but has not detailed timelines for public detection tools or how the watermark will behave on smaller edits. The “cat-and-mouse” dynamic suggests further removal utilities and possible policy refinements will emerge as stakeholders adapt to the new labeling requirement.

ZDNet