MarkMyAss

Lab: "Claude Watermark"#

Current status: Unknown · Detect: Unknown · Remove: Unknown · Last tested: 2026-08-13

← Back to the Lab · ← Back to the GhostMark cleaner

People searching for "Claude watermark" usually mean one of three completely different mechanisms. Conflating them is the single most common source of confusion in this space, so this page exists specifically to keep them separate.

The three things people mean#

1. File / metadata provenance#

If you ask Claude to help produce a file (a document, an exported image, etc.) through some pipeline, that pipeline may embed ordinary file metadata: an author field, a "Producer" string, XMP data, or (for images) a C2PA Content Credentials manifest. As of the same August 2026 announcement above, Anthropic states that supported image outputs (SVG, PNG, JPG) now get signed C2PA provenance metadata attached directly by Claude, not just by a downstream export pipeline -- see the same Anthropic Help Center article. Either way, this is the same category of signal any C2PA-aware document/image tool can add, and it's exactly what GhostMark's PDF-metadata, EXIF/XMP, and C2PA detectors target -- see /lab/pdf-metadata and /lab/c2pa for what GhostMark can and cannot do with a signed manifest specifically (short version: GhostMark detects/strips the JUMBF container structurally; it does not validate or forge cryptographic signatures).

GhostMark's status: Supported. Detect: Yes. Remove: Yes. Independently verified: Yes, via ExifTool (and c2patool for the C2PA container).

2. Hidden Unicode characters in text#

Some people call invisible Unicode characters (zero-width spaces, Unicode "Tags" block steganography, etc.) found in AI-generated text a "watermark." This is not correct, and GhostMark does not call it one. Hidden Unicode is a text-encoding artifact that can appear in text from any source -- it is not a statistical signature tied to a specific model, and its presence doesn't identify which model (if any) produced the surrounding text. See /lab/hidden-unicode for the full, separate writeup of this mechanism.

GhostMark's status: Supported (as its own, distinct signal). Detect: Yes. Remove: Yes (with load-bearing characters like ZWJ/NBSP preserved by default). Independently verified: Yes, deterministically.

3. Statistical / model-level text watermarking#

This is what people usually mean when they ask "does Claude watermark its text so it can be detected later." The idea: a provider could bias how its model samples tokens in a detectable-but-imperceptible way, such that only the provider (holding a private key/seed) could later run a statistical test against a piece of text and estimate the odds it came from their model.

On August 11-12, 2026, Anthropic publicly confirmed this mechanism is real, not hypothetical. Its support article states Claude "embeds an imperceptible watermark directly into the text itself" that "will travel with the text when it's copied and pasted elsewhere," rolling out to models launched on or after August 2, 2026 (with older models being retrofitted), across claude.ai, the API, Claude Code, Claude Cowork, and Claude Tag, worldwide. Source: Anthropic Help Center -- "How Claude marks AI-generated content".

That same article is explicit that no detector exists publicly yet: "We're also working to enable users and other third parties to detect Claude's embedded watermarks and provenance metadata... We'll share details on detection mechanisms in forthcoming technical documentation." Until that documentation ships, there is nothing independently runnable for GhostMark (or anyone outside Anthropic) to implement or test against. GhostMark will not fabricate a detection result for a mechanism nobody outside the provider can currently verify.

GhostMark's status: Unknown. Detect: Unknown. Remove: Unknown. Independent verification: No public verifier exists yet.

Why GhostMark reports "Unknown" instead of a score#

Reporting a confidence percentage here -- "87% likely watermarked," "0% detectable" -- would require either (a) Anthropic's own private detection key, which GhostMark does not have and never will, or (b) an independently published, peer-reviewed detection methodology GhostMark could implement and that others could verify GhostMark implemented correctly. Neither exists publicly today. Any tool claiming otherwise is either guessing or measuring something else (like hidden Unicode or formatting) and mislabeling it as "the watermark."

ghostmark inspect and ghostmark inspect-text report this explicitly as a distinct row: Claude statistical watermark: UNKNOWN / NOT CURRENTLY VERIFIABLE. Implementation and the plug-in interface a future real detector would use: src/ghostmark/detectors/statistical.py.

What would change this page#

Anthropic has already committed to publishing "forthcoming technical documentation" on detection. When that ships -- or if an independent researcher publishes a reproducible detection methodology and open code first -- GhostMark would implement it behind the same StatisticalWatermarkDetector interface already defined for this purpose, and this page's status would change from Unknown to whatever the evidence actually supports. Until then, this page states the current state of public knowledge, not a guess.

Reproducible test commands#

ghostmark inspect-text "any text you like" --json
# -> "statistical_claude": {"status": "unknown", ...}

Sources#

Something outdated or inaccurate?#

If Anthropic (or anyone) publishes something that changes the facts on this page, please tell us: open an issue or submit a pull request against src/ghostmark/web/content/lab/claude-watermark.md.

Last reviewed: 2026-08-13