Knowledge base

AI watermarks,
without the guesswork.

Direct answers about signals hidden in text, images, and video, the detectors they require, and the conclusions you can—and cannot—draw from a scan.

01 · Start hereWhat is an AI watermark?

The four layers people call watermarks—and why each needs a different detector.

02 · Run a text checkWhat can an AI text watermark checker detect?

A direct map of C2PA text, hidden Unicode, Claude, SynthID, and what a zero result proves.

03 · Run an image checkWhat can an AI image watermark checker detect?

A direct map of C2PA, metadata, SynthID pixels, visible marks, and what a negative result proves.

04 · Clean safelyHow to remove an AI text watermark

Which literal signals can be removed without rewriting and which statistical marks cannot.

05 · ClaudeCan Claude's watermark be detected?

Anthropic's announced model-level mark, C2PA file provenance, and the detector details still forthcoming.

06 · GeminiHow SynthID Text detection works

Why configured keys and token analysis—not a hidden-character scan—are required.

07 · TerminologyWhat is an AI signature?

Separate statistical fingerprints, cryptographic signatures, provenance, and literal file data.

08 · Image provenanceHow C2PA validation works

What a valid Content Credential proves, what trust means, and what absence does not prove.

09 · Copied textHow C2PA text watermarks work

The exact hidden variation-selector wrapper defined in C2PA 2.4, including detection, removal, and validation limits.

10 · Markdown and codeHow C2PA works in structured text

Manifest references in comments or front matter, local Base64 decoding, and conservative whole-block removal.

11 · Audit the checkerDetection methodology

Exact coverage, cleanup safeguards, scanner versions, evidence reports, and known limits.

12 · WebP imagesCheck WebP metadata and C2PA

How WebP stores Exif, XMP, and Content Credentials—and how safe removal is verified.

13 · Original test10,000 AI responses scanned

Our method, results, and the evidence that hidden characters do not reveal SynthID Text.