Anthropic Adds Invisible Watermarks to Claude-Generated Content

Anthropic is introducing invisible, machine-readable watermarks into content generated by its Claude AI models. Unlike a traditional visible watermark, the mark cannot be seen by readers and is designed to remain associated with the text when it is copied and pasted. Anthropic says the system is intended to improve transparency and make it possible to identify AI-generated material without changing the appearance of the writing.

The change is connected to new transparency requirements under the European Union’s AI Act, particularly rules requiring AI-generated content to be identifiable by machines. Anthropic says the watermarking approach is being applied globally rather than being restricted only to European users. New Claude models released from August 2, 2026 are covered, with existing models being updated as well.

Anthropic is also using a different system for files. Supported images and other files can contain C2PA provenance metadata, which provides information about the content’s origin and history. The company is developing detection tools that could allow Anthropic and eventually third parties to determine whether content carries the Claude watermark or provenance information.

The move could have major implications for education, publishing, business and online media. AI-generated writing is increasingly difficult to distinguish from human writing based solely on appearance, so machine-readable provenance could give organizations another way to determine whether AI was involved in producing content. However, Anthropic acknowledges that the system has limitations, particularly when content is heavily rewritten, translated or otherwise transformed.

Anthropic’s decision also raises questions about how the wider AI industry will respond. If invisible watermarking becomes a common standard, AI-generated text could eventually carry a persistent digital signal across different platforms and services. At the same time, the effectiveness of the approach will depend on reliable detection and its ability to distinguish fully AI-generated material from text that was only lightly edited or assisted by Claude.