Anthropic Embeds Hidden Watermarks in Claude Models

Anthropic has announced that it will integrate Google's SynthID-Text watermarking technology into its Claude AI models. Starting August 2, 2026, every new model will have a hidden statistical fingerprint baked into its outputs.

The watermarks are designed to be detectable by anyone holding the right cryptographic key and can be used to identify whether text was generated by Anthropic's AI. This move aligns with the EU AI Act's Article 50(2) Code of Practice, which requires AI platforms to make their outputs identifiable.

The SynthID-Text watermark operates at the token-selection layer of a language model and subtly biases the randomness used in word choice. It preserves output quality, readability, and user experience, according to Anthropic, with Google's own deployment on Gemini validating this claim across 20 million responses.

Existing Claude models will be retrofitted with watermarks by December 2, 2026. Anthropic will also attach C2PA metadata to supported image files. The company's forthcoming detection API will allow third parties to check whether text was generated by Claude.