Anthropic has begun rolling out machine-readable watermarking across its Claude models. The technology is designed to help distinguish AI-generated content from human-created work and applies across Claude products and services, including API services, Claude Code, and enterprise deployments.
The system uses two main methods to mark content. It embeds an imperceptible watermark directly into text generated by Claude and adds signed provenance metadata to supported files, including PNG, JPG, and SVG formats, using the C2PA open standard.
These markers are designed to be detected by machines rather than seen by users. Anthropic says the signals can help indicate AI involvement but are not absolute proof, as watermarks can be weakened or removed through substantial editing, while metadata can be stripped from files.
The rollout is linked to the European Union’s AI transparency requirements, which took effect on August 2, 2026. New Claude models launched in the EU from that date are required to carry the relevant markers, while older models are being given a transition period. Anthropic plans to bring the watermarking to existing models by December 2, 2026.
The move comes as AI companies continue to explore ways of identifying machine-generated content. OpenAI previously shut down its AI text classifier after the company found that it correctly identified only 26% of AI-written text in one evaluation, while incorrectly flagging human-written text 9% of the time.
OpenAI also developed an unreleased text watermarking system that was reportedly 99.9% effective when enough new text was generated. The company chose not to release the system publicly amid concerns about its impact on users and the possibility of circumvention.
Google DeepMind has also been developing similar technology through SynthID. In May 2024, the company announced SynthID for text generated through the Gemini app and web experience, adding to its existing watermarking work for AI-generated media.
