What is an AI watermark?
An AI watermark is a machine-readable signal associated with generated content. It may be woven into word choices, embedded in pixels or audio, or stored as signed provenance metadata. These are different systems—not one universal mark.
You cannot check every AI watermark with one scan. A detector must look in the same layer where the mark was placed and, for keyed systems, usually needs the provider's detector or secret key.
The four layers called “AI watermarks”
A model biases token choices according to a keyed pattern. The watermark lives in the sequence—not in an invisible character. Google's SynthID Text is in this category.
A signal is embedded in image pixels or audio samples and designed to survive common edits. Detection analyzes the media itself.
C2PA Content Credentials attach a tamper-evident history to media or can be encoded directly into copied text. This is provenance, not a statistical guess.
Content may contain Unicode controls, C2PA text wrappers, Exif, XMP, comments, or application metadata. These are inspectable and often removable, but their presence alone does not prove AI authorship.
What this AI watermark checker detects
Check AI Watermarks scans pasted text for a defined set of invisible or unusual Unicode characters, exact C2PA 2.4 invisible wrappers, and structured manifest blocks. For PNG, JPEG, and WebP images, it reads supported metadata and uses the official C2PA browser library to inspect and validate embedded Content Credentials. It does not rewrite prose or alter image pixels.
| Signal | Can check here? | Why |
|---|---|---|
| Supported hidden characters | Yes | They are literal code points in the pasted text. |
| C2PA 2.4 text credential | Yes, structurally | The exact wrapper is decoded and removable; signature and text binding are not yet validated. |
| C2PA structured-text reference | Yes, structurally | The reference is decoded locally; supported whole-line or front-matter blocks are removable. |
| PNG/JPEG/WebP metadata | Yes | It is stored in supported file segments or RIFF chunks. |
| Embedded C2PA credentials | Yes | The manifest and cryptographic validation data are readable. |
| SynthID Text | No | It is a statistical token pattern requiring the appropriate detector. |
| Pixel-level SynthID | No | This checker deliberately does not analyze or edit pixels. |
See the exact pasted-text coverage or see the exact image-checker coverage.
What does “no watermark found” mean?
It means no supported signal was found in the layers that were scanned. It does not prove the content is human-made. It also does not rule out a provider-only watermark, an unsupported format, an externally stored credential, or a signal damaged by editing.