Anthropic Starts Watermarking Claude Text as the European Union's Transparency Rules Take Effect

Anthropic has begun embedding an invisible watermark in text produced by Claude models released on or after August 2, and attaching provenance metadata to generated files using the Coalition for Content Provenance and Authenticity (C2PA) standard. The marking applies across the application programming interface (API), Claude Code, Claude Cowork and deployments on Amazon Web Services, Google Cloud and Microsoft Foundry, with a grace period for existing systems running to December 2026.

The method is a version of SynthID-Text, a technique Google DeepMind published in the journal Nature in 2024. Rather than biasing the model's output probabilities toward a fixed vocabulary list, it manipulates the source of randomness used when the model samples each token, so a secret key combined with the preceding text jointly determines what comes next. Anthropic reports no statistically significant quality difference in human ratings between watermarked and unwatermarked output.

The limits are well documented and were pointed out immediately. Detection degrades on short passages, where there are too few sampling decisions to carry a signal, and the mark weakens under paraphrasing, translation, summarization by a second model, splitting into fragments, or conversion into structured data. It also answers a narrower question than most people assume: how likely it is that a passage passed through Claude, not who wrote it, and not whether a different model produced it. The mark can appear even when Claude only corrected spelling, which is why developer reaction has been mixed.

The trigger is regulatory. Article 50 of the European Union's Artificial Intelligence Act became enforceable on August 2 and requires generative providers to mark outputs in machine-readable form.