Claude Watermark
Anthropic has announced text watermarking for future Claude models.
It says the watermark adds no hidden characters.
A hidden-character match therefore does not prove a Claude watermark.
Check the characters, or clean supported invisible characters.
What Anthropic has publicly documented
Anthropic describes its text watermark as a statistical pattern created through word-selection choices. Its official explanation says nothing is added to the text and there are no hidden characters. It also says watermark detection estimates the likelihood that Claude was involved; it does not establish that every word was authored by Claude.
Read Anthropic’s explanation of Claude text watermarking.
How the statistical check works
Anthropic says the watermark changes the source of randomness used when Claude chooses among suitable next words. A detector with Anthropic’s key can examine the resulting word sequence for consistency with that pattern and estimate the likelihood that Claude contributed to the text. This is a statistical test of word choices, not a scan for a marker embedded between characters.
Source: Anthropic’s description of the method and its key.
Why the amount and kind of text matter
Anthropic says short samples contain fewer word choices, so watermark detection works less well on them. It also describes factual passages, proofreading, and much code as having fewer equally acceptable choices in which the pattern can appear. Longer passages and text written more extensively by Claude provide more decisions for the detector to evaluate.
Source: Anthropic’s stated detection limitations.
What a hidden-character result proves
A positive result from this site proves only that the supplied text contains a character or markup type covered by the checker. It cannot establish who inserted it, why it is present, or whether Anthropic’s statistical watermark is present.
The Unicode Standard documents legitimate text-layout purposes for characters such as the zero-width joiner and zero-width non-joiner. Their presence is therefore a character-level fact, not proof of authorship or provenance.
Read the Unicode Standard’s description of format characters.
Text watermarks and file credentials are different
Anthropic distinguishes its statistical text watermark from the content credentials attached to supported generated files. For supported image and vector formats, it describes a cryptographically signed C2PA note in file metadata. That file credential is metadata; the text watermark is a pattern in word selection.
Source: Anthropic’s explanation of files and C2PA credentials.
Use the right check
Use the hidden-character checker when you want a read-only count of supported invisible characters, special spaces, and metadata tags. Use the Claude cleaner when you want to remove supported invisible characters and normalize supported special spaces without rewriting or translating the visible text.
Neither action detects or removes the statistical word-choice pattern Anthropic describes. The checker does not have Anthropic’s key, and the cleaner deliberately preserves visible wording instead of replacing the sequence of words.
Questions people ask about Claude watermarks
Can anyone use Anthropic’s official watermark detector?
Not at present. Anthropic says its detection API is in private preview for eligible organizations and certain enterprises, with broader access planned over time. CleanText is not that API and does not claim access to its key.
Can a Claude watermark identify a person, organization, or chat?
Anthropic says its text watermark contains no identifying information and cannot be traced to a particular user, organization, or conversation.
Does a hidden-character count prove that Claude wrote the text?
No. It reports only the covered characters or markup found in the submitted text. Anthropic’s documented watermark is based on word-selection statistics and requires a different test.
Can editing remove Anthropic’s statistical watermark?
Anthropic says light editing probably will not remove it completely, while replacing every word in a complete rewrite will. Its explanation also notes that after such a rewrite, the description “AI-generated” becomes debatable.
Can CleanText remove Anthropic’s statistical watermark?
No. CleanText removes supported invisible characters and normalizes supported special spaces. It does not rewrite the visible words that carry the statistical pattern described by Anthropic.
Source for the answers above: Anthropic’s official text-watermark FAQ.