Anthropic has detailed an invisible watermarking system for its Claude AI that would flag AI-generated text, though it has not yet described exactly how it works. The system would weave imperceptible signals throughout generated text, allowing detection via an as-yet-undescribed mechanism. The watermark is part of the text, travels when copied and pasted, and may persist through some editing, according to Anthropic. It will be applied at the model level, present across all Claude products, and will not affect meaning, quality, or readability. The detection system could be used on anything from job applications to personal emails, but would be most useful in schools, where teachers increasingly judge whether work is student-written or AI-generated. Anthropic’s commitment comes as part of compliance with new European laws demanding transparency measures, though watermarking would apply worldwide. New models will include it from day one, and existing models will be updated in the future. OpenAI has developed its own text watermarking system but has not rolled it out to ChatGPT publicly. Google has developed SynthID Text for Gemini output. All three companies, plus Apple, apply standard C2PA watermarks to generated images, with Google adding its own watermark designed to survive cropping and re-formatting. Critics have described detection systems as antidotes owned by the same companies profiting from the poison, or as a protection racket.