How Claude’s text watermark works
Aug 14, 2026 — Anthropic
Future Claude models will generate text that contains a watermark. This is a way of determining the likelihood that Claude was involved in writing the text, and we, along with several other major AI providers, are implementing this change to comply with the EU AI Act.
Summary of key points
- Watermarking uses a method that does not have any practical impact on the quality or content of Claude’s outputs
- The difference between watermarked and un-watermarked text will not be distinguishable to readers
- Nothing is added to the text and there are no hidden characters
- Watermarking doesn’t require extra tokens, and will not be more expensive
- Watermarking carries no identifying information and can’t be traced to a specific person, organization, or chat
- As of August 2, the EU requires AI providers serving its market to mark AI-generated content. Other major model developers signed the same Code of Practice and will implement their own watermarks
What is watermarking?
Large language models like Claude work by generating one word at a time. Each time the model decides on the next word, it chooses among a list of potential candidates. In low-stakes passages where either word choice would preserve meaning (e.g., “overcast” vs “grey”), watermarking uses a different source of randomness — a key plus preceding words — to settle word choices. The pattern is undetectable to readers but detectable with the key.
Technical method
Claude’s text watermark is a version of the SynthID-Text approach published by Google DeepMind in a Nature paper in 2024. It belongs to a family of approaches going back to Scott Aaronson’s 2022 proposal.
Limitations:
- Can only answer “What is the likelihood this was partly written by Claude?”
- Cannot confirm human authorship or identify other AI systems
- Less effective on short samples and factual passages with few interchangeable word choices
- Code has generally less watermarking because exact tokens are required
Impact on outputs
Anthropic reports no impact on content, creativity, or readability in internal testing. Google DeepMind found no statistically significant thumbs-up/down differences in Gemini traffic tests. Human raters saw no quality difference in controlled side-by-side comparisons.
Watermarking has negligible impact on speed; no extra tokens means no price increase.
Compliance context
Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026 alongside ~190 signatories. Watermarking is applied globally at launch because Anthropic does not yet have a durable way to scope by region.
Models launched before August 2, 2026 have a transition period; watermarking for older models will roll out over coming months.
Detection and other media
- A watermark detection API is planned; implementation details are being worked out
- For PNG, JPG, and SVG files, Claude attaches C2PA provenance metadata (cryptographically signed) rather than embedded watermarks
- Light editing may not remove watermarks; complete rewrites will
- Watermarks apply to translations (every word chosen by Claude)
What watermarking does not prove
A watermark can only determine that Claude was likely involved. It cannot distinguish “Claude wrote this” from “Claude heavily edited this.” It does not change ownership or legal responsibility under Anthropic’s terms.