Watermark Detection Is Not Attribution: Anthropic's Article 50 Rollout as a Predictive Case Study in the Occurrence Primitive

On 2 August 2026, in response to Article 50 of the EU AI Act, Anthropic began embedding machine-detectable watermarks in text generated by Claude and attaching signed C2PA provenance metadata to generated files. The rollout was applied globally rather than scoped to the European market. It is among the first production deployments of text watermarking across every surface of a frontier model provider's products.

This paper examines that deployment as a predictive case study. It differs from the earlier papers in this series in an important respect: the failure it describes has not yet occurred. Anthropic states the limits of its marks precisely in its own documentation, and the practitioner commentary that has followed states the detectability-versus-attribution distinction correctly. Nothing in the record shows an institution treating watermark detection as attribution.

The argument is therefore not corrective but explanatory. It sets out why the hazard is structural rather than a matter of conceptual confusion — why an assurance regime built on occurrence will infer attribution from detection even where every party understands, in the abstract, that it should not. The mechanism lies in what the mark actually tracks. A watermark records token selection with precision. Token selection is not intellectual authorship, and the two diverge systematically: a translation carries a full mark while its ideas belong to the original author, and a proofreading pass may carry almost no mark while the influence on the text is real.

Because the paper predicts rather than reports, it states the conditions under which the prediction fails, with a date. No measurements, rates or prevalence claims are made.