01 Paste text
Learn
Runs on this device
Hidden watermarks Text + image + video 100% local
Detect and remove
hidden AI watermarks.
Scan text, images, and video for supported invisible markers, AI metadata, and embedded Content Credentials—without rewriting words or editing pixels.
Hidden AI text watermark scanner
Paste. Inspect. Clean.
02 See markers
0supported signals found
Paste text to reveal invisible characters.
03 Remove markers
→
Your cleaned text appears here after review.
Supported today
Find the hidden watermark layer.
Every supported finding is concrete, inspectable, and tied to a specific character or signed part of your media file.
Find exact C2PA 2.4 invisible wrappers and structured manifest blocks plus supported zero-width characters, unusual spaces, and controls.
Remove known clutter, normalize unusual spaces, and preserve language-, emoji-, and math-sensitive controls for review.
Read image and video C2PA records, verify integrity and signatures, and show signed AI-generation or AI-editing declarations.
Know what you are scanning
Not every AI watermark lives in the same layer.
Some watermarks are literal data inside a file. Others are woven into word choices or pixels and need the original provider's detector. This tool only claims what it can verify.
| Watermark type | How it works | Detect here | Remove here |
|---|---|---|---|
| C2PA text credentialCopied Unicode text | A complete C2PA manifest store encoded as an invisible, standards-defined block of Unicode variation selectors beginning with C2PATXT. | ✓Structure decoded | ✓Complete wrapper |
| C2PA structured textMarkdown, YAML & code | A manifest URI or embedded Base64 store between fixed C2PA ASCII-armored delimiters in comments or front matter. | ✓Reference decoded | ✓Safe whole blocks |
| Hidden UnicodeCopied text | Literal zero-width characters, unusual spaces, and controls stored in the text. These can come from tools or formatting, but are not a standard LLM watermark. | ✓Supported markers | ✓Safe markers |
| Image metadataPNG, JPEG & WebP | Exif, XMP, comments, editor data, and other information stored around the image rather than in its pixels. | ✓Supported fields | ✓Selected fields |
| Image Content CredentialsPNG, JPEG & WebP C2PA | A signed, tamper-evident record that can describe how an image was created or edited, including generative-AI use. | ✓Signature validated | ✓Embedded copy only |
| Video Content CredentialsMP4, MOV & M4V C2PA | A signed C2PA record embedded in the video container. Its actions can explicitly declare AI generation, AI editing, or algorithmic enhancement. | ✓Signature + AI claim | ×Read-only today |
| Claude text watermarkAnnounced August 2026 | Anthropic says supported Claude models weave an imperceptible mark into generated text that survives copy and paste. Its technical format and detector have not been published yet. | ×Detector forthcoming | ×Method undisclosed |
| Statistical text watermarkGemini SynthID Text | A probability pattern woven into the model's token and word choices during generation. There is no hidden character to delete. | ×Provider detector needed | ×Not removable here |
| Pixel-level watermarkSynthID Image | An imperceptible signal embedded directly into generated image pixels and designed to survive common edits. | ×Provider detector needed | ×Pixels stay untouched |
| Frame-level watermarkSynthID Video | An imperceptible signal embedded across generated video frames. It is not the same as a C2PA credential in the file container. | ×Provider detector needed | ×Frames stay untouched |
| Visible watermarkLogo or label | Text, a logo, or a disclosure rendered into the image itself. It is visible content, not hidden file data. | ×No pixel analysis | ×No pixel editing |
The practical rule: supported hidden characters, C2PA text credentials, image metadata, and embedded image or video C2PA credentials can be shown here. Video credentials are read-only. Watermarks woven into word choices, pixels, or video frames still require the provider's detector. Text checker coverage · Content Credentials · SynthID Text
APrivate by design
Files and text stay in this browser tab.
BDeterministic
No prompts, model calls, or probabilistic claims.
CExact findings
Every text result maps to a specific Unicode code point.
Frequently asked questions
What can it actually do?
This detector finds supported machine-readable markers in copied text, images, and video. It does not remove visible logos or detect every model-level watermark.
What does a result of 0 mean?
It means the pasted text contains no supported literal hidden character or complete C2PA text wrapper. Claude and SynthID remain not testable, so zero is not proof that every watermark is absent or that a person wrote the text.
Does it validate Content Credentials?
For supported PNG, JPEG, WebP, MP4, MOV, and M4V files, the official C2PA browser library checks the manifest, asset integrity, signature, and configured trust list. For copied text, this release decodes the C2PA 2.4 wrapper structure but does not yet validate its signature or text binding.
Which image formats can it inspect?
PNG, JPEG, and WebP. The parser validates each container before reporting supported Exif, XMP, comments, editor data, timestamps, or embedded C2PA structures.
Can it check an AI watermark in video?
It can validate embedded C2PA Content Credentials in MP4, MOV, and M4V files and show a signed AI-generation or AI-editing declaration when one exists. It cannot detect SynthID embedded across video frames; that requires Google's compatible detector.
Can it remove AI text watermarks?
It can remove complete C2PA 2.4 text credentials and supported literal hidden characters. It cannot remove statistical systems such as SynthID Text, which are patterns in word choice rather than attached characters.
Can it detect C2PA credentials in copied or structured text?
Yes. It decodes the invisible C2PATXT wrapper used for unstructured text and the ASCII-armored manifest blocks defined for Markdown, front matter, source-code comments, and similar structured text. Each format has a conservative cleanup policy.
Will it detect Claude's new text watermark?
Not yet. Anthropic describes a model-level mark but has not published its promised technical detector details. A result of zero only means no supported literal signal was found.
What did the 10,000-response test show?
Only 15 responses contained supported hidden characters, and most findings were normal emoji controls. Google's 3,000 watermarked samples and 3,000 clean controls both showed zero because SynthID Text was not stored as hidden Unicode.
Will cleaning damage text or image pixels?
Text cleanup changes only supported safe markers and complete C2PA wrappers; language-, emoji-, and math-sensitive controls remain for review. Image cleanup removes selected container metadata while preserving render-relevant image data.
Does my content leave this device?
No. Text and supported image or video files are inspected inside the browser without an upload or an LLM call.
Evidence, not guesswork
Understand the result.
Clear definitions, primary sources, and our own testing—so a watermark check never turns into an unsupported authorship claim.
GuideWhat is an AI watermark?
Token patterns, pixel signals, provenance, and literal file data.
→Text checkerWhat can a paste scan detect?C2PA text, hidden Unicode, Claude, SynthID, and honest zero results.
→Image checkerWhat can an image scan detect?C2PA, metadata, pixel watermarks, visible marks, and honest negative results.
→Removal guideWhat can be removed without rewriting?Clean literal payloads while preserving words and sensitive controls.
→TerminologyWhat is an AI signature?Watermarks, fingerprints, cryptographic signatures, and style clues.
→C2PAWhat does validation prove?Understand present, valid, trusted, invalid, and absent results.
→New in C2PA 2.4Can copied text carry C2PA?How hidden variation selectors encode a removable text credential.
→Markdown & codeHow structured text carries C2PAASCII-armored manifest references in comments and front matter.
→ClaudeCan Claude's mark be checked?What Anthropic has published—and what remains unavailable.
→GeminiWhy a Unicode scan cannot see SynthIDKeys, token probabilities, and provider detector requirements.
→Original test10,000 AI responses scannedOur datasets, method, findings, and limitations.
→