Anthropic adds Claude text watermarks citing anti-distillation defense

Anthropic is reportedly implementing invisible watermarking in Claude-generated text, a move observers attribute to anti-distillation defenses rather than EU AI Act compliance. Critics note that current European regulations do not mandate output watermarking for general-purpose models, suggesting the technical measure targets unauthorized model cloning instead. Distillation attacks involve training smaller competitor models on proprietary outputs to replicate capabilities at lower cost. By embedding detectable patterns in generated text, Anthropic aims to identify and block synthetic data harvesting by rival labs. This development marks a strategic pivot toward technical intellectual property enforcement amidst intensifying competition in the foundation model market. Industry analysts suggest this preemptive defense could become standard practice as open-weight models increasingly leverage proprietary API outputs for training. The initiative highlights growing tension between open research norms and corporate IP protection strategies in generative AI.