What image metadata can tell you
Image files can carry several layers of information besides pixels: EXIF fields written by cameras and apps, IPTC and XMP fields used by publishers and editing tools, PNG text chunks, and C2PA manifests. Some of these layers can record that an AI model was involved. All of them describe a file's history as written by software, so they are clues to examine, not facts to accept.
IPTC Digital Source Type
IPTC maintains a Digital Source Type vocabulary for describing how media was made. The code trainedAlgorithmicMedia means media created by an AI model trained on sampled content, compositeWithTrainedAlgorithmicMedia covers augmentation such as inpainting or outpainting, and compositeSynthetic describes a composite that includes at least one generative AI element. digitalCapture describes media captured from a real-life source with a digital camera or recorder. Platforms and tools can read these values when deciding how to label content.
Source: IPTC Digital Source Type vocabulary (opens in a new tab)
Generation parameters in PNG files
Some local image generators write prompts and settings into PNG text chunks. ComfyUI, for example, can embed the full workflow in saved PNG images so the file can be reopened to recreate the setup. These fields are strong evidence when present, but they often disappear when an image is converted to JPEG, screenshotted or uploaded to a social platform.
Source: ComfyUI examples: workflow metadata in images (opens in a new tab)
C2PA manifests and EXIF software tags
A C2PA manifest is embedded data that records claims about a file, such as the tool that generated it, and it is signed so that later changes can be detected. Its value depends on validating that signature. By contrast, an EXIF Software tag or an XMP creator-tool field is plain text that any editor can write or remove. A software field naming an AI app is a lead worth checking, not proof of origin.
How to inspect metadata safely
Use a trusted viewer such as ExifTool on a copy of the original file, and avoid uploading private images to unknown websites. Compare what you find with the claimed source, for example whether camera fields match the device someone says they used. Note whether each field is signed, standardized or free text, and record what is missing as well as what is present.
Source: ExifTool (opens in a new tab)
Why metadata can mislead
Metadata is easy to strip and easy to forge. Messaging apps and social platforms often remove it, screenshots never carry the original fields, and editing software can copy fields from one file to another. A file with camera EXIF can still be AI-generated, and a photo with no metadata can still be genuine. Weigh metadata alongside provenance signals, the source and the claim.
Can you tell if an image is AI-generated from its metadata?
Sometimes. A validated C2PA manifest from a generator, or an IPTC Digital Source Type of trainedAlgorithmicMedia, is a strong clue. But most shared images have had their metadata removed, and fields can be edited, so metadata alone can confirm a lead but never prove that an image is real.
Does AI-generated art have EXIF data?
It varies by tool. Some generators write IPTC or XMP fields, some embed C2PA manifests, some write generation parameters into PNG text, and some write nothing. AI images rarely carry genuine camera EXIF such as exposure settings, but such fields can be added afterwards, so their presence is not proof of a camera.
Do social media platforms remove image metadata?
Many platforms re-encode uploads and strip most metadata from the copies they serve, though some read provenance fields first in order to apply AI labels. Assume that an image saved from a feed has lost useful metadata, and ask for the original file when it matters.