Get hold of the original
Metadata only survives in the file as it left the camera or generator. Screenshots, re-saves and images pulled from social feeds have usually been re-encoded and lost most of it. Ask for the original export — the one sent “as a file”, not as a compressed photo — whenever you can.
Clue one: Content Credentials
Tools that support C2PA attach a digitally signed record. Its claim generator names the app or service that produced the image, its c2pa.actions list steps such as “created” or “edited”, and a digitalSourceType can state AI generation directly.
Because the record is signed, a verifier can also tell whether the image changed after signing. A manifest from an AI service is the strongest metadata evidence available.
Clue two: IPTC DigitalSourceType
Many generators and editors fill in a standard IPTC field describing how content was made. The values you will meet most:
- trainedAlgorithmicMedia — produced by a generative AI model.
- compositeWithTrainedAlgorithmicMedia — a real photo with AI-generated or replaced areas, as with Generative Fill.
- algorithmicMedia — made by an algorithm without training data, such as procedural art.
- digitalCapture — taken with a camera, no generative content declared.
Clue three: prompts and tool names
Stable Diffusion front-ends and ComfyUI save prompt, seed, model and workflow as text in PNG files. Other tools leave their name in XMP CreatorTool or EXIF Software. Any of these points clearly to AI involvement — and often tells you exactly how the image was made.
Checking if an image is AI-generated
Lay the file on the table at the top of this page. The “AI traces” tab gives one of three verdicts — AI traces found, worth a second look, or no AI traces — and lists each signal with its source. “Raw metadata” shows the underlying values. Nothing is uploaded. To remove what it finds, use the AI metadata remover.
If the file is empty
Absent metadata proves nothing either way: editors, chat apps and social networks strip it constantly, and missing camera EXIF is a weak hint at best. Next steps are the provider’s own watermark checker (for example for Google’s SynthID), a reverse image search to find the earliest copy, and a look at where the image first appeared. Pixel-based AI detectors report probabilities, not proof.
About checking images for AI
Can I tell if an image was made with ChatGPT or Gemini?
If the file still has its Content Credentials, the claim generator names the service. For Gemini images, Google’s SynthID check can also detect the watermark in the pixels.
If AI Label Remover finds no AI traces, is the photo real?
Not necessarily. It means the file holds no AI metadata. Metadata is easily lost or removed, so read a clean result as “unknown”, not as proof.
Can I check something saved from Instagram or Facebook?
You can, but platforms usually strip the published copy, so expect little or nothing. The platform’s own “AI info” label is a separate signal.
Is the check free and private?
Yes — free, no account, and the file is read locally without being uploaded.