Image Courtesy: Getty Images
Anthropic has begun adding invisible, machine-readable watermarks to text generated by new Claude models, giving organizations a way to identify AI-generated writing even when it looks completely ordinary to human readers.
The watermarking applies to Claude models launched on or after August 2, 2026, and Anthropic says the system will be used worldwide rather than being limited to European users. The company introduced the measure as part of its transparency commitments under the European Union’s AI Act.
The watermark is embedded directly into the generated text and is designed to remain invisible without specialized detection tools. Anthropic says it travels with text when users copy and paste it and can survive some editing, although extensive rewriting, translation, or mixing the output with other material can weaken or remove the signal.
The marking is applied at the model level, meaning it can follow Claude-generated content across several products and services. This includes Claude, the Claude API, Claude Code, Claude Cowork, and Claude Tag. Customers accessing supported models through cloud platforms including Amazon Web Services, Google Cloud, and Microsoft Foundry will also receive marked output.
Anthropic plans to release technical information and detection tools so users and third parties can identify supported Claude watermarks. Older Claude models will not immediately receive the same treatment, although Anthropic says it is working to add marking capabilities during the EU AI Act’s transition period.
For files, Anthropic is taking a different approach. Supported generated files, including certain image formats, can contain digitally signed provenance information based on the Coalition for Content Provenance and Authenticity, or C2PA, standard. The metadata can indicate that Claude processed a file and whether its provenance information has been altered.
Anthropic cautions that neither system should be treated as definitive proof that a person or AI created a piece of content. Claude can edit, summarize, or translate human-written material, while heavy editing can make watermarks harder to detect. Files can also lose provenance metadata through ordinary processing.
The change could have major implications for publishers, educators, businesses, and online platforms dealing with AI-generated material. As AI writing becomes harder to distinguish from human work, machine-readable provenance gives organizations another tool for assessing where content came from.
