Field notes / AI provenance
AI watermark · provenance

AI watermark detector: C2PA, SynthID and what a result really means

There is no single universal “AI watermark.” Some systems attach signed provenance metadata, some embed imperceptible signals in the pixels, and some detectors only estimate synthetic-looking patterns.

Published September 15, 2026 · CreatorPrivacyKit Research

Strong positive

A trusted provenance signal can associate a file with a participating generator or export workflow.

Layered checks

C2PA, SynthID and pixel classifiers are different evidence sources and should be reported separately.

Weak negative

“No supported signal found” does not prove that an image is human-made.

Three things people call an “AI watermark”

LayerExampleWhat it is
Signed provenanceC2PA / Content CredentialsA cryptographically signed record attached to media and its history
Imperceptible watermarkSynthIDA hidden signal embedded into generated content and detected by supported systems
Statistical detectorVisual AI classifierA model that estimates whether pixel patterns resemble synthetic content

These layers answer different questions. C2PA can tell you about a signed creation or editing chain when credentials are present. SynthID can indicate that supported content carries a participating watermark. A visual detector can provide probabilistic evidence even when no metadata exists, but it can also be wrong.

OpenAI verification now checks more than metadata

OpenAI's current verification flow checks supported files for provenance signals associated with OpenAI tools, including trusted C2PA manifests and supported SynthID watermarks. A positive result can indicate that the content was generated by or exported from OpenAI tools. It does not establish factual accuracy, ownership, identity or context.

This matters for detector design: a useful report should say which signal was found instead of collapsing everything into one “AI / not AI” badge.

Google SynthID is vendor-scoped evidence

Google DeepMind's SynthID embeds imperceptible watermarks into supported AI-generated media. Its detector can identify supported Google AI content. A missing SynthID signal is useful information about that ecosystem, but it does not rule out OpenAI, Midjourney, open-source diffusion models or other generators.

Why screenshots change the problem

A screenshot can destroy file-level provenance because the screenshot is a new file containing recaptured pixels. That often removes the original C2PA container data. Pixel-domain watermarking may survive some transformations better, depending on the system. This is one reason a detector should inspect multiple layers rather than rely on metadata alone.

How to read a detector result

  1. Positive C2PA: inspect the signer and claims, not just the existence of a manifest.
  2. Positive watermark: note which vendor/system the watermark is associated with.
  3. Positive visual classifier: treat it as model evidence, not cryptographic proof.
  4. No supported signal: report the negative result narrowly.
  5. Conflicting signals: preserve the original file and investigate transformations instead of forcing a binary answer.

What a good AI image checker should show

At minimum: file metadata, C2PA status, supported watermark checks, image dimensions and a clearly labeled visual-classifier score if one is used. It should also expose limitations: which vendors are supported, whether URL checks require server fetching, and whether an edited derivative can still be linked to an original provenance chain.

Do not overclaim: “No supported AI watermark detected” is a valid result. “This image is definitely human-made” is a much stronger conclusion and usually is not justified by that result alone.

Why this search category is growing

As major AI vendors expose provenance verification, “AI image detector” is splitting into more precise intents: “C2PA checker,” “Content Credentials checker,” “SynthID detector,” “ChatGPT image detector” and “AI watermark detector.” The best tools will explain these signals rather than pretend they are interchangeable.

Check provenance by layer

CreatorPrivacyKit is building a privacy-first image provenance workflow that combines file metadata and supported provenance checks without turning uncertainty into a fake binary verdict.

Explore provenance tools →