Illustration: a pirate's spyglass scan-beam catches a worried ghost half-hidden inside a document

The crew's spyglass is on your hidden marks.

Claude Watermark Remover#

MarkMyAss removes and verifies supported signals — not everything people call "the Claude watermark"#

If you landed here searching for "Claude watermark remover," you probably mean one of three genuinely different things. MarkMyAss is built around keeping them separate, because a tool that quietly conflates them is a tool that can lie to you without saying anything false.

Mechanism	MarkMyAss's status
Hidden Unicode characters in text	Supported — detect and remove
File metadata (EXIF/XMP/IPTC/PDF DocInfo)	Supported where applicable — detect and remove
C2PA / Content Credentials container	Partial — structural detection/removal, not cryptographic validation
Claude statistical model watermark	Not currently publicly verifiable — see why below

MarkMyAss will not tell you it removed something it can't actually confirm. The last row is the important one, and most "watermark remover" tools online skip past it.

Clean a file or text now →

What MarkMyAss actually does here#

1. Hidden Unicode — fully supported#

Text copied from a chat interface can carry invisible Unicode characters: zero-width spaces, Unicode "Tags" block characters, stray bidi marks. These aren't a Claude-specific signature — any text-producing pipeline can introduce them — but they're real, detectable, and MarkMyAss removes them deterministically while preserving characters that are actually load-bearing (emoji sequences, right-to-left scripts, French typography).

ghostmark inspect-text "your pasted text here" --json
ghostmark clean-text "your pasted text here"

Full methodology: /lab/hidden-unicode.

2. File metadata — supported where applicable#

If you exported a document, image, or PDF through some pipeline downstream of Claude, that pipeline may have embedded ordinary file metadata (author, producer, XMP fields). MarkMyAss strips this byte/segment-level for JPEG, PNG, WebP, and PDF, and independently verifies removal with ExifTool.

ghostmark clean document.pdf
ghostmark verify document.ghostmark.pdf --original document.pdf

Full methodology: /ai-metadata-cleaner and /lab/pdf-metadata.

3. C2PA / Content Credentials — partial#

As of August 2026, Anthropic's own documentation states that Claude attaches signed C2PA provenance metadata to supported image outputs (SVG, PNG, JPG) — see Anthropic Help Center. MarkMyAss detects and strips the JUMBF container structure a C2PA manifest lives in, and cross-checks with the official c2patool where installed. This is structural container removal, not cryptographic manifest validation — MarkMyAss doesn't forge or verify signatures.

Full methodology: /c2pa-remover and /lab/c2pa.

4. The statistical text watermark — honestly, not yet#

This is what most people actually mean by "Claude watermark": an imperceptible bias in how the model samples tokens, meant to be detectable later without changing what the text says. Anthropic confirmed on August 11–12, 2026 that this mechanism is real and rolling out to models launched after August 2, 2026, but its own documentation states plainly that no public detector exists yet — "we'll share details on detection mechanisms in forthcoming technical documentation."

MarkMyAss cannot detect, remove, or verify removal of a mechanism with no published, independently reproducible detection method. Any tool that claims otherwise is either measuring something else (hidden Unicode, formatting) and mislabeling it, or guessing. ghostmark inspect-text reports this row explicitly as UNKNOWN / NOT CURRENTLY VERIFIABLE rather than a fabricated pass.

See the MarkMyAss Claude Watermark Lab methodology → for the full breakdown, sources, and what would change this page.

FAQ#

Does MarkMyAss remove Claude's invisible statistical watermark? No. No public detector exists for MarkMyAss (or anyone outside Anthropic) to test against, so there's nothing to verify removal of. MarkMyAss reports this status as unknown rather than claiming success.

Is hidden Unicode the same thing as the Claude watermark? No, and MarkMyAss never calls it that. Hidden Unicode is a text-encoding artifact any source can introduce; it's unrelated to model-level statistical watermarking. See /lab/claude-watermark for the full distinction.

Will removing metadata make an image "undetectable" as AI-generated? No claim like that exists on this site. MarkMyAss removes the specific, named signals in the table above and reports exactly what it did and didn't verify — not a synthetic confidence score.

What if Anthropic publishes a detector later? MarkMyAss already has a pluggable interface for this (src/ghostmark/detectors/statistical.py) and would implement it the moment a reproducible methodology exists. The Lab page tracks this.