Anthropic Deploys Token-Bias Watermarking in Claude Models to Meet EU AI Act Requirements

Anthropic's initial support page titled 'How Claude Marks AI-Generated Content' asserted imperceptible watermarks that preserve readability and semantics. The revised document 'How Claude’s Text Watermark Works' describes steganographic adjustment of next-token probabilities at each step, favoring words from predetermined green lists to enable probabilistic detection downstream. This implements requirements under the EU AI Act for labeling synthetic content.

James Padolsey's interactive analysis demonstrates the mechanism: repeated biased coin-flip equivalents across token decisions accumulate statistical signals detectable above random baselines. Anthropic provides no public benchmarks on detection false-positive rates or degradation in perplexity scores. Primary research papers on similar techniques, including those from 2023-2025, report 5-15% increases in cross-entropy loss when watermark strength exceeds 0.7 bias ratio.

The approach directly addresses misuse vectors such as undetected AI-generated disinformation campaigns, yet introduces systematic alteration of output distributions. This creates tension with developer expectations for deterministic tool behavior and raises verification challenges for non-English tokenizers where green-list construction is less stable.

Operational deployment across all Claude endpoints worldwide begins within 90 days. Enterprises using Claude for regulated content generation must now log watermark detection thresholds separately from model outputs to maintain audit trails.