Starting August 2, 2026, Anthropic has integrated Google DeepMind’s SynthID-Text technology into all new Claude models. This system embeds invisible statistical fingerprints into AI-generated text, allowing detection via a cryptographic key. The implementation applies globally to all Claude deployments, rather than being restricted to European users.
The technology functions at the token-selection layer, biasing word choices in low-stakes contexts where multiple options are interchangeable. Anthropic states that this process maintains output quality and readability. Evidence from Google’s deployment of SynthID-Text on Gemini, which involved approximately 20 million responses, suggests no measurable degradation in user experience.
The watermark remains effective through minor modifications, such as light editing or paraphrasing, though it does not survive a full rewrite. Anthropic plans to retrofit existing production models with this technology by December 2, 2026. Additionally, the company will implement C2PA metadata for supported image files.
This initiative aligns with the EU AI Act’s Article 50(2) Code of Practice, joining similar commitments from Google, Meta, Microsoft, and OpenAI. Anthropic also intends to release a detection API to enable third-party verification of Claude-generated content.
Recent research indicates potential side effects of the watermarking process. Findings suggest that modifications to token selection may influence tool invocation and alter the probability of a model adhering to its safety guardrails.
