Nebula AI Detector
AI image detection · Methodology

AI Image Detection: Metadata vs C2PA vs Watermarks

The short answer: no single method can settle the question on its own. Metadata is the most precise but evaporates the moment an image is re-encoded. Pixel classifiers are the most general and carry a real false-positive rate. C2PA credentials and watermarks are the most trustworthy — but only cover part of the pipeline. This guide breaks each one down and cross-checks them.

Last updated: 2026-11-11

Most people searching for an "AI image detector" want one of two things: which tool is the most accurate, or why was my photo flagged. Both questions need a more basic answer first: what are these tools actually looking at? They all present the same UI — upload, get a percentage — but the underlying techniques differ enormously in reliability.

1. Detection is actually four different things

MethodWhat it inspectsStrengthMost common failure
Metadata / generator signaturesGenerator name, parameters and AIGC labels written into the fileConclusive on a hitPlatform re-encoding, screenshots, edits
Pixel classifiersTexture, noise and frequency characteristics of the image itselfProbabilistic; measurable false positivesNew generators, recompression, partial edits
C2PA Content CredentialsA signed chain of custody: who made it, who changed itConclusive on a hitGenerator not onboarded / signature stripped
Generation-time watermarksAn invisible mark embedded when the image was producedConclusive on a hitCropping, repainting, re-encoding

The crucial distinction is the type of evidence. Metadata, C2PA and watermarks are conclusive — the mark is either there or it is not. A pixel classifier is an inference about what the image looks like. Treating an inference as proof is the most common misuse of these tools.

2. Metadata and generator signatures: precise, but easy to erase

Most generators write something into the file on export. Common traces:

This evidence has one fatal precondition

Metadata lives only in the original file. Every major social platform re-encodes uploads, and the generator parameters are gone on arrival; screenshots do the same. So metadata checks are near-perfect on a freshly downloaded generated file and close to useless on an image you found on the open web.

3. Pixel classifiers: the broadest coverage, the most false positives

Tools in this family (AI or Not, Hive, Illuminarty, WasItAI, Sightengine and others) ignore metadata and run a trained model over the pixels. The upside is coverage: any format from any source can be run through them. The downside is that the verdict is inherently a probability, and the measured false-positive rate is not small.

Why editing is so damaging

Pixel classifiers learn the statistical signature of generated images. Re-compression, text overlays, cropping and colour grading all disturb exactly those signals — which is why the same tool can score the original and the re-encoded copy differently. That inconsistency is expected behaviour, not a bug.

4. C2PA Content Credentials: provenance, not detection

C2PA (the Coalition for Content Provenance and Authenticity) defines provenance credentials: the producer writes a signed chain into the file recording who generated it and what was changed afterwards. Anyone can check it with the official verifier, Content Credentials Verify. Adobe, Microsoft, the BBC and a growing list of camera makers ship it.

It is routinely misunderstood: C2PA verifies "is there a trusted signature", not "is this AI-made". An entirely real old photograph can carry a camera-maker signature, and an AI image can carry no signature at all if the generator never onboarded. Signatures can also be deliberately stripped — some tools even ship a "remove metadata" button to demonstrate it.

5. Generation-time watermarks: the most reliable, the narrowest coverage

Google's SynthID is the best-known example: the model embeds an invisible watermark in the pixels at the moment the image is produced, and detection extracts and compares it. This is currently the lowest false-positive category — it is not asking "does this look like AI", it is asking "is my specific mark in here".

The limitation is equally clear: only generators that embed a watermark are recognisable this way. Third-party models, tools that never onboarded, and images that have been cropped, repainted or heavily re-encoded lose the signal.

6. The layer Western tools mostly ignore: China's AIGC labelling standard

Since 1 September 2025, China has enforced the mandatory national standard GB 45438-2025 (Cybersecurity technology — Labelling methods for content generated by artificial intelligence). It requires that AI content be labelled through generation and distribution, and it splits labelling into two kinds:

This dimension is essentially blank in the English-language market. AI or Not, Hive, Sightengine and Illuminarty all discuss pixel classification and metadata, and none of them touch national AIGC labelling standards. For content moderation on Chinese platforms, an explicit-label hit is hard evidence that is entirely independent of any pixel judgement — worth checking on its own.

7. How to combine them: a workable order of checks

  1. Check metadata and AIGC labels first — a generator signature or a national-standard label settles it, at the lowest cost.
  2. Then look for C2PA credentials and watermarks — a valid signature chain or an extractable watermark is equally conclusive.
  3. Use pixel classification last — at this point it is supporting evidence for a tendency, not a verdict on its own.
  4. Account for the file's provenance — for a screenshot or a platform download, assume the first two layers are gone. Do not read a missing metadata hit as "not AI".

8. Three common misconceptions

The one-line summary

Chain of evidence first, probability second. Metadata, credentials and watermarks are proof; a pixel score is an estimate. Confusing the two is why both false positives and false negatives go unexplained.

Frequently asked questions

Which AI image detection method is the most accurate?

Judged on a single verdict, generation-time watermarks and C2PA credentials are the most reliable — they check for a verifiable mark rather than a resemblance, so they rarely mislabel a genuine photograph; but their coverage is narrow, limited to onboarded pipelines. Pixel classifiers have the widest coverage and the highest error rate, with third-party audits measuring real photographs misclassified as AI in the 13%–40% range depending on the tool. Metadata sits in between: conclusive when present, but only if you have the original file. Cross-checking the three is the only way to get a defensible answer.

Does C2PA prove an image was AI-generated?

Not directly. C2PA Content Credentials verify the integrity of a chain of custody — who produced or edited the file and what happened to it since. A real photograph can carry a camera-maker signature, and an AI-generated image may carry no signature at all if the generator never onboarded. It proves traceability, not authorship by AI, and is better used to judge whether an image's edit history is complete and credible.

Why was my real photo flagged as AI?

Three usual causes: over-processing — aggressive denoising, sharpening, HDR merging and heavy compression all shift the pixel statistics towards generated-looking; extreme capture conditions — very shallow depth of field, blown highlights, plain backgrounds and heavy skin retouching are known weak spots; and how you got the file — after a screenshot or a platform re-encode, the original camera metadata is gone and the model is guessing. When that happens, get the original file if you can: the camera make and shooting parameters in it are the strongest available rebuttal.

What is GB 45438-2025 and how does it relate to AI image detection?

GB 45438-2025 is a Chinese mandatory national standard on labelling AI-generated content, in force since 1 September 2025. It requires AI content to be labelled through generation and distribution, split into explicit labels (inspectable markers in metadata) and implicit labels (digital watermarks embedded in the data for tracking across the distribution chain). Its relevance to detection is that it supplies hard evidence independent of any pixel judgement: a file carrying a compliant AIGC label confirms its AI provenance outright. The converse does not hold — finding no label does not prove an image is genuine.

Related articles

Check your own image

Upload an image and the tool reports metadata clues and pixel-level AI probability separately, so you can cross-check the two yourself.

Open the detector