Spot the ghost before it stows away.

Claude Watermark Detector#

What "checking for a Claude watermark" can and can't actually tell you#

This page is about detection, not cleaning. If you have a file or piece of text and want to know what MarkMyAss can genuinely check for it — and what a "not found" result does and doesn't prove — read this before you trust any tool's answer, including MarkMyAss's.

Check a file or text →

What MarkMyAss tests#

Running ghostmark inspect or ghostmark inspect-text checks for four independent, named signals — never a single blended "AI score":

1. Hidden Unicode characters — deterministic, rule-based scan.
2. File metadata (EXIF/XMP/IPTC/PDF DocInfo) — deterministic, and cross-checked against ExifTool if installed.
3. C2PA / Content Credentials container — structural JUMBF-marker scan, cross-checked against c2patool if installed.
4. Statistical model-level text watermark — reported explicitly as UNKNOWN, because no public detector exists (see below).

Each signal gets its own row: FOUND, NOT FOUND, or UNKNOWN. There is no combined percentage, because a combined percentage would have to average a real measurement with a guess.

What "NOT FOUND" means (and doesn't mean)#

NOT FOUND for hidden Unicode or file metadata means MarkMyAss's detector actually ran its checks against the content and found none of the specific patterns it looks for. It is a real, reproducible result — run it again and you'll get the same answer, because the detection rules are public and deterministic (not a black-box model).

NOT FOUND does not mean:

• "This text is definitely human-written." MarkMyAss checks for specific technical signals, not authorship in general.
• "This text has no statistical watermark." MarkMyAss has no way to check that signal at all right now — see below — so it never reports NOT FOUND for it. It reports UNKNOWN.
• "No AI provider anywhere embeds anything undetected by MarkMyAss." MarkMyAss only implements the specific mechanisms listed above.

What MarkMyAss does NOT prove#

The single most important limitation: MarkMyAss cannot tell you whether text was generated by Claude specifically, because Anthropic has not published a public, independently runnable detector for its statistical text watermark. Anthropic's own August 2026 announcement confirms the mechanism exists but states detection tooling is "forthcoming" — see Anthropic Help Center and the full writeup at /lab/claude-watermark.

Any tool online claiming to detect "the Claude watermark" today is either checking a different, unrelated signal (hidden Unicode, formatting quirks, writing-style heuristics) and mislabeling it, or producing an unverifiable guess. MarkMyAss's own inspect output makes this distinction explicit rather than hiding it in fine print.

Supported vs. unsupported signal types#

FAQ#

Can I detect if a specific piece of text came from Claude? Not with certainty, using any publicly available tool as of this writing — MarkMyAss included. What you can check: whether it contains hidden Unicode, or (for files) embedded metadata that might indicate its origin.

Why doesn't MarkMyAss just estimate a probability? Because there's nothing to base an honest probability on. A provider's private statistical watermark is only checkable by the provider holding the detection key. Anything calling itself a "Claude detection score" today is not measuring what it claims to measure.

Does a clean detection result mean I can safely claim something is human-written? No. MarkMyAss reports what its named detectors found or didn't find — that's a narrower, more honest claim than "this is/isn't AI-generated."

Check a file or text → Run the inspector

Sources#

• Anthropic Help Center — "How Claude marks AI-generated content"
• ExifTool
• C2PA specification