Anthropic is implementing watermarking across all content generated by Claude, with different approaches depending on the medium. For text, the watermark works by subtly altering the probability of word choices in ways imperceptible to human readers—yet the mark persists even after the content is edited or copied. Documents and images use open-standard C2PA metadata for watermarking instead.
The move directly addresses the European Union's AI Act requirements, though Anthropic plans to roll out watermarking globally rather than limiting it to EU users. The strategy reflects growing pressure on AI companies to make generated content traceable and verifiable, particularly as regulators worldwide grapple with deepfakes, misinformation, and the need to distinguish human-created from AI-generated material.