Anthropic confirmed on August 11, 2026 that new Claude models hide a machine readable signal inside the words they write. Here is how it works in plain terms, and why your code is mostly unaffected.
It is not what most people guess
- Not hidden characters. There are no invisible spaces to strip out.
- Not metadata. Nothing is attached to the response.
- Not a writing tic. It has nothing to do with favourite phrases or punctuation.
The mark lives in which words Claude picks, and nowhere else.
The basic idea
At every step Claude has many valid options. "The cat sat on the ___" could be mat, rug, floor, or carpet, and all of them are fine.
A secret key plus the text so far gets hashed, splitting the vocabulary into two halves called green and red. Green words get a small nudge upward. Repeat a few hundred times and the passage holds far more green words than chance allows. A detector with the same key simply recounts them.
Length is everything
The signal is spread thinly across hundreds of words, so no single word carries it. One sentence tells you almost nothing.
*Words generated. Illustrative shape, not measured values.
Code and math barely carry it
When only one token is correct, there is no freedom left to hide anything in. The nudge is small by design, so it never overrides a confident choice. That is exactly why generated code still runs.
The few flexible spots in code are variable names, comment wording, string contents, and the order of independent lines. A formatter or a rename refactor wipes most of those in seconds.
*Relative, illustrative. Driven by how much word choice each format allows.
What a result actually means
- Mark found: the text may have been processed by Claude. That is not proof Claude wrote it, since people use it to translate, proofread, and summarise their own work.
- No mark found: proves nothing at all. It could be an older model, too short a passage, or text that was edited on the way.
Still unknown
- Anthropic has not published which watermarking method it uses.
- The public detector has not shipped yet.
- There is no published data on whether a mark survives code formatting or a pull request.
Treat it as a weak hint about origin, not an audit trail. If you need real provenance on generated code, log the model, the prompt, and a hash of the raw output yourself.
