Anthropic deploys SynthID-Text invisible watermark globally for Claude and C2PA signature for files
Anthropic embedded SynthID-Text invisible watermark in Claude, Claude Code, and API outputs on August 15. Files get C2PA signature metadata. The change is global to meet EU AI Act requirements.
On August 15, Anthropic embedded the SynthID-Text invisible watermark in text generated by Claude, Claude Code, Claude Cowork, Claude Tag, and the Claude API, complying with the EU AI Act transparency requirements for general-purpose AI providers. The approach draws on the SynthID-Text method Google DeepMind published in Nature in 2024, adjusting the probability distribution over low-stakes token choices during generation to create a statistically traceable pattern.
Detection requires a key, and third-party developers will be able to query an API to verify whether a piece of text was processed by Claude. Anthropic's internal testing showed no measurable impact on content quality, creativity, or readability. The signal weakens on short or fact-dense text because the model has fewer low-stakes word choices available. Code has an even smaller signal because syntax constraints narrow model choices, though inline comments can still carry a mark. Light editing will likely not erase the watermark, but replacing every word will.
File outputs such as SVG, PNG, and JPG are signed with C2PA open-standard provenance metadata, allowing downstream tools to confirm whether a file has been modified since Claude produced it. The watermark applies to all Claude models launched on or after August 2, 2026, with earlier models in a transition period. Cloud-partner access through AWS, Google Cloud, and Microsoft Foundry also embeds the watermark.
Some users have canceled subscriptions over concerns that watermarked code deliverables could reveal AI use. Anthropic said it has not observed a noticeable increase in cancellations.