Claude’s Invisible Text Watermarks Doesn’t Work the Way You Think
Anthropic has published the answer to a frequently asked question from the past few days: How its new invisible Claude watermarks work.
First announced last Wednesday, these watermarks will appear in all text generated by supported Claude models launched on or after Aug. 2, 2026—Claude Code, Claude Cowork, Claude Tag, and its API—as well as in supported Claude options through AWS, Google Cloud, and Microsoft Foundry. Older models will gradually be retrofitted with the watermarking technology.
As revealed in a new Anthropic blog post, these watermarks don’t add any extra characters (hidden or visible) or metadata to a piece of text. Instead, it uses patterns built into the sequence of word choices that Claude makes. In simpler terms, it relies on patterns in how Claude assembles its verbiage.
The model chooses words based on a secret key and the words that came before. This creates a pattern that a detector can later check to estimate how likely Claude was to have helped write the text. Anthropic says the watermark does not change the meaning, quality, or readability of a Claude-generated passage, and human readers cannot see it.
The system purportedly works best with natural language because there are many different ways to convey the same thing. For code, it is harder to inject the watermark since it’s far less flexible; as a result, the watermark can only be added in comments or variable names.
Meanwhile, translations receive full watermarking because Claude chooses every word. However, if Claude only fixes grammar or makes small edits, most words remain human-written, so detection may not work well.
Anthropic says it will give users a detection API to check whether text likely involved Claude. Detection works better on longer passages because more word choices give the detector more data.
The company also says a positive result means Claude likely had a hand in the text, but does not mean that it wrote everything. Light editing probably will not erase the watermark, but heavy rewriting can remove it completely.
