Anthropic's text watermarks signal new front in AI detection
TL;DR
Anthropic plans to add machine readable watermarks to text and files generated by new Claude models. According to Axios the measure is meant to satisfy European transparency rules and applies to models launched in the EU after August 2, with marking available wherever Claude is offered worldwide. Text carries patterns Anthropic describes as imperceptible, while generated media files get digital signatures.
Nauti's Take
Machine readable watermarks bring clarity where detectors previously guessed, and they give editors a verifiable provenance signal. The catch hits everyday practice: proofreading, formatting or translating a human draft can be enough to trigger the mark, and Anthropic itself calls the detection limited.
Teams using Claude for press releases should test their own documents before rollout to see when a mark appears and which metadata the files carry.
Sources