Anthropic’s new provenance system for Claude combines two technologies that are easy to conflate: an imperceptible watermark carried by generated text and digitally signed metadata carried by supported files. They address the same broad question—where did this content come from?—but they operate in different ways and fail in different ways.
The rollout is tied to the European Union’s Article 50 AI transparency requirements, which began to apply on August 2, 2026. New Claude models released on or after that date support marking from launch. Models that were already available are still being updated, and older systems receive a limited transition period through December 2 for the machine-readable marking obligation.
For text, Anthropic says the watermark is woven into output at the model level. It is designed to be imperceptible and to avoid changing meaning, quality or readability. Because the signal is part of the text rather than a document header, it can travel through copy and paste and may remain detectable after some editing.
Anthropic has not disclosed the specific encoding method. Modern language-model watermarking can in principle use statistical properties of token selection, but there is no public basis for assigning that mechanism—or any particular alternative—to Claude’s new system. The company has said more technical documentation and third-party detection support are forthcoming.
That uncertainty matters for evaluation. A useful watermark needs both low false positives and enough robustness to survive ordinary transformations. Anthropic acknowledges that substantial editing, paraphrasing, translation or mixing with other text can make detection harder. As a result, the absence of a detectable mark cannot establish human authorship.
A positive detection is not a full authorship test either. The mark may have been introduced because Claude translated or edited a human-written passage. A detector can potentially identify model participation, but it cannot infer the intellectual contribution of the model from the signal alone.
File provenance is more explicit. Anthropic says supported generated files will include digitally signed provenance metadata, using C2PA for supported media. The C2PA architecture stores cryptographically verifiable claims about an asset’s provenance in signed manifests. Those claims can document origin and transformations in a way that downstream software can validate.
Cryptographic verification does not make metadata permanent. A file can be separated from its provenance information during processing, and some platforms may strip metadata. C2PA therefore includes mechanisms intended to improve durability, but the practical survival of provenance still depends on the workflow.
Nor does C2PA establish factual truth. It can verify that a signer made a provenance claim and that the signed claim has not been altered without detection. It cannot determine whether an AI-generated image depicts a real event or whether text in a generated document is accurate.
Anthropic plans to deploy supported marking globally across Claude surfaces, including the API and supported cloud channels. This will create a large real-world experiment in model-level text provenance. Once third-party detection tools are available, researchers will be able to measure how the watermark behaves under common edits, translations and transformations. Those results—not the existence of a watermark alone—will determine how scientifically useful the signal becomes.
