How Claude’s Invisible Watermark Works—and What You Can Verify

What Anthropic announced

Anthropic describes an imperceptible, model-level watermark for text from supported Claude models. The mark is described as part of the generated wording rather than a visible symbol or a file property. That distinction matters: a model-level statistical pattern is not the same thing as a zero-width character that anyone can find with a Unicode scanner.

Coverage can change by model, launch date, product surface, language, and output type. A careful tool should therefore point readers to Anthropic’s current documentation instead of assuming that every piece of text associated with Claude has the same marking behavior.

Model-level text watermark versus hidden characters

Hidden Unicode is directly observable. A scanner can enumerate U+200B zero-width spaces, U+00A0 non-breaking spaces, bidirectional controls, soft hyphens, variation selectors, tag characters, and unusual control codes. Those characters often have ordinary origins: websites, document editors, copied email, language shaping, emoji sequences, or publishing systems.

A model-level watermark is different. It can be encoded through subtle choices made during generation and evaluated statistically by a detector designed for that scheme. Removing a hidden character does not demonstrate that a model-level pattern changed, and finding one does not demonstrate that Claude placed it there. This site keeps those findings in separate result cards for exactly that reason.

File provenance and C2PA are a third category

Files can carry signed provenance metadata. C2PA is a technical standard for recording claims about where media came from and what happened to it. That information belongs to a file or asset and can be checked with software that understands the signed manifest. Plain text copied into a textarea usually does not preserve the surrounding document container or its signed file metadata.

A future file checker could inspect C2PA, EXIF, XMP, IPTC, PDF properties, or document metadata. That would still be a file-provenance feature, not a substitute for an official detector for a model-level text watermark. The launch version intentionally focuses on plain text and does not pretend that file verification has already been implemented.

Which Claude models and outputs are covered?

Consult Anthropic’s live support article for the exact current scope. New models may adopt a policy before older models, and supported text may differ from images or downloadable files. Output made from fixed strings, short code, JSON, quotations, or heavily constrained templates may also provide less room for a statistical marking method than ordinary prose.

Why detection is probabilistic and limited

Statistical detectors need enough eligible text. Short passages contain fewer generation choices, while edited passages may no longer retain enough of the original pattern. Translation, strong paraphrasing, combining text from several authors, or inserting fixed material can all change the evidence a detector sees. A result should be interpreted within the detector’s documented operating conditions rather than treated as an authorship verdict.

These limits are also why sentence length, vocabulary variety, repeated phrases, and paragraph shape are not Claude detectors. They can describe a sample, but human writing and writing from many different systems overlap on those measurements. This site reports them only as descriptive statistics.

What editing, translation, and short text can change

Light copyediting may preserve much of the original text. A strong rewrite can replace a large portion of the wording and structure. Translation changes nearly every token. Short excerpts may never contain enough evidence for a reliable statistical test. None of these observations lets an independent site claim that a mark has definitely been removed—especially when the official third-party verifier is not available.

What this site checks locally

The local inspector reads the Unicode code points in the pasted text, records code-point and UTF-16 positions, shows context, groups similar signals, and explains legitimate uses. It also checks line endings, trailing whitespace, common HTML and Markdown artifacts, word count, sentence length, paragraph count, vocabulary variety, and repeated two-word phrases.

Cleanup is selective. It can normalize line endings, trim trailing spaces, replace selected Unicode spaces, remove soft hyphens, remove a conservative subset of zero-width formatting, remove bidirectional controls when selected, and apply NFC normalization. ZWJ, ZWNJ, variation selectors, and emoji tags are protected by default. Code, URLs, email addresses, citations, and valid JSON are protected where possible. Every cleanup shows a diff before apply.

What rewriting can and cannot prove

Optional rewriting changes expression at a chosen strength and tone. The server prompt instructs the provider to preserve names, facts, numbers, dates, URLs, citations, code, JSON keys, commands, and file paths. Basic checks warn if protected facts disappear. The text is uploaded only after the user chooses Rewrite Text.

A rewrite is not a verification step. It does not prove that a Claude watermark existed, that one was removed, or that another detector will classify the output in a particular way. It is an editing feature with an explicit review requirement.

Claude Code, code, JSON, and fixed strings

Source code and structured output contain many fixed tokens and strict syntax choices. A generator has less freedom to choose alternative wording, and users frequently mix generated fragments with existing code. For those reasons, origin claims based only on a pasted snippet deserve extra caution. The local tool can still reveal hidden characters and formatting that might break a command, identifier, comparison, or parser.

Current status

As of 2026-08-11, the site does not have an official public third-party method it can use to verify Anthropic’s model-level text watermark. The product therefore displays “Not publicly verifiable yet” and provides no Claude probability score. If Anthropic publishes a supported mechanism, the status, model coverage, test corpus, and limitations will be updated before any official-verification result is added.

Official and technical sources

• Anthropic: How Claude marks AI-generated content
• C2PA technical specification
• Cloudflare Turnstile server-side validation