Anthropic’s move to add invisible watermarks to Claude-generated content is already facing a familiar problem: people are trying to strip or defeat the label as soon as it appears.
WIRED reports that Anthropic introduced the watermarks last week to comply with new European Union rules on AI-generated material. Within hours, coders were discussing overrides online, raising doubts about whether invisible signals can reliably identify machine-written content once users can edit, reformat, or route text through other tools.
Watermarking is attractive because it promises a quiet technical answer to a social problem: knowing when text came from AI. But the report highlights the weakness of approaches that depend on cooperation from users and platforms after the content leaves the original model.
The episode does not mean all provenance systems are useless. It does show that watermarking alone is unlikely to settle questions about disclosure, enforcement, or authenticity, especially for developers who can test the boundary directly.