SAN FRANCISCO — Anthropic has implemented a system to weave imperceptible watermarks into text generated by its Claude models, applying the change to all models launched on or after Aug. 2, 2026. The company also attaches signed C2PA provenance metadata to supported file types, including.svg,.png and.jpg.
The policy follows Anthropic's commitment to the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content. The company signed the code as a provider of both generative AI models and generative AI systems, a distinction that carries different obligations under the Act.
Aug. 2, 2026, marks the date from which the EU AI Act's code applies to newly launched models. Anthropic's decision to apply watermarking globally, rather than limiting it to the European Union, stands out. Nothing in the EU code requires text generated for users in the United States or India to carry such marks.
That global application has two plausible readings. It could reflect a principled consistency in Anthropic's approach to AI transparency. Or it may be the most cost-effective engineering path — avoiding the complexity of maintaining separate inference pipelines for different geographic regions.
The watermark is embedded in the text itself, not merely attached as metadata. Anthropic said the watermark travels with text when copied and pasted, and may persist through some editing, making it resistant to removal by simple text editors or retyping.
Text watermarking typically works by subtly influencing token selection during the model's sampling process. At each step, the model's choice among statistically similar next tokens is biased according to a hidden key, creating a detectable statistical signature across a longer passage.
For generated files, C2PA metadata provides a digital signature of provenance, allowing verification of a file's origin and whether an AI system produced it. The standard is gaining traction across the digital content industry.
Online reaction has been largely negative among paying users. Commentary frequently conflated text watermarking with file metadata, missing the distinctions in Anthropic's approach.
Anthropic has not published its specific technical scheme for text watermarking. The company said forthcoming technical documentation will detail its method, providing clarity on the robustness and detectability of the embedded marks.
For enterprise customers, imperceptible watermarks and provenance metadata offer a verifiable layer of content authenticity — increasingly relevant as businesses integrate generative AI into content creation, marketing and legal workflows. The ability to trace content origin can reduce risks tied to misinformation or intellectual property disputes.
The global rollout sets a precedent for how frontier AI companies handle regulatory compliance across jurisdictions. Whether a single global standard proves cheaper than segmented regional approaches will affect operational overhead and product development cycles for every major model provider watching Anthropic's execution.
