AI-generated image metadata checker
AI image tools increasingly embed provenance information — and so do modern cameras. This tool reads an image in your browser and scans its metadata for signals that it was AI-generated: C2PA content credentials, generator names like DALL-E, Midjourney, or Stable Diffusion, Adobe Content Authenticity tags, and the software field that records the producing application. It also notes when an image carries essentially no metadata, which can be a weak signal in context.
Understanding the signals
Not all metadata indicators carry the same weight. Here is how to interpret what the checker finds:
C2PA content credentials are the most reliable indicator. C2PA (Coalition for Content Provenance and Authenticity) is an open standard supported by Adobe, Microsoft, and several AI image providers. A C2PA manifest is cryptographically signed, embeds a chain of custody, and names the tool that produced the image. It is difficult to fake convincingly. If the checker finds a C2PA manifest naming an AI generator, that is strong evidence.
Generator strings in Software/Creator EXIF fields are medium-confidence signals. Tools like AUTOMATIC1111 write their name into the Software field of JPEG EXIF. This is easy to strip (any image editor that re-saves will remove it), but if present it is a clear indicator.
AI-related XMP fields such as Iptc4xmpExt:DigitalSourceType with a value
of trainedAlgorithmicMedia are part of the IPTC Photo Metadata standard and
are embedded by some responsible publishers and tools.
Missing metadata entirely is a weak signal on its own. Most social platforms strip all metadata when you download an image, so absence means very little without other context.
How it works
You choose a JPEG or PNG file. The tool reads the raw bytes locally and searches the EXIF, IPTC, and XMP text regions for known markers: the C2PA / JUMBF manifest signature, generator strings, AI-related XMP fields, and the Software/Creator tags. Each match is listed with what it implies. The image is never uploaded — all parsing happens in JavaScript on your device. The result is a summary of indicators found plus a cautious verdict, since metadata can be removed or forged.
When to use it
This checker is useful when:
- You receive an image from an external source and want to check for embedded provenance before publishing.
- You want to verify whether your own publishing pipeline is correctly attaching or stripping metadata.
- You are a journalist, researcher, or fact-checker evaluating whether an image may be AI-generated as part of a broader investigation.
It is not a classifier and does not analyse pixel patterns — it reads only what is embedded in the file. For high-stakes assessments, combine it with visual analysis and dedicated AI detection classifiers.
Tips and notes
- Evidence, not proof. Metadata is easy to strip and possible to fake; weigh it alongside visual cues.
- Absence is weak. Platforms routinely strip metadata, so “no metadata” alone means little.
- C2PA is the strongest signal. A valid content-credentials manifest is hard to fake and usually names the producing tool.
- Cross-check. Combine with reverse image search and a dedicated detector for anything high-stakes.