To check image metadata for AI, open the original file (not a screenshot) in a tool that shows its full metadata and look for three things: Content Credentials (C2PA), an IPTC “digital source type” field, and a software or generator tag naming an AI tool. On a computer, ExifTool or the Content Credentials Verify site show all of this. On a phone, your gallery’s info panel shows the basics, and an AI detector that reads metadata can surface the rest.
Metadata answers only one question: did the file declare that AI was involved? If it did, that’s strong evidence. If it didn’t, you’ve learned very little, because most images online have had their metadata stripped along the way.
What AI metadata looks like #
| Field | Where it lives | What an AI marker looks like |
|---|---|---|
| Content Credentials (C2PA) | A signed manifest embedded in the file | An assertion that the image was created or edited with generative AI, plus the tool name |
| IPTC Digital Source Type | XMP/IPTC metadata | trainedAlgorithmicMedia or compositeWithTrainedAlgorithmicMedia |
| Software / Creator Tool | EXIF or XMP | The name of a generator or AI editor |
| PNG text chunks | Inside PNG files | Prompts, seeds and settings written by some Stable Diffusion interfaces |
| Camera fields | EXIF | Make, model, lens, exposure; present on camera photos, absent on most generated images |
The IPTC codes to know #
The IPTC, the standards body for news photo metadata, publishes a list of “digital source type” values (IPTC). The ones that matter here:
- trainedAlgorithmicMedia: created by an AI model trained on captured content. A fully generated image.
- compositeWithTrainedAlgorithmicMedia: a real image augmented or corrected with generative AI, such as inpainting or outpainting.
- compositeSynthetic: a composite of several elements, at least one of them generative AI.
- digitalCapture: captured from a real scene with a camera.
- screenCapture: a screenshot.
Google Photos uses IPTC metadata to mark photos edited with its AI tools such as Magic Editor and Magic Eraser (Google), and Meta reads the “AI generated” information in C2PA and IPTC fields to label images from other companies’ tools (Meta).
How to check metadata on a computer #
Content Credentials Verify #
The quickest way to read C2PA data is the free Content Credentials Verify page at contentcredentials.org/verify, run by the Content Authenticity Initiative. Drop an image in and it shows whether the file carries credentials, who signed them, which tools were used, and whether AI was involved. If the file has no credentials, it tells you that too.
ExifTool #
ExifTool is a free command-line program that prints every metadata field it can read, including EXIF, IPTC and XMP. Run exiftool image.jpg and scan the output for “Digital Source Type”, “Software”, “Creator Tool” and any PNG “Parameters” text. It’s the most thorough option, and it shows fields that friendlier viewers hide.
Built-in viewers #
- Windows: right-click the file, choose Properties, then the Details tab. You’ll see camera fields and the program name, but not C2PA or most XMP.
- Mac: open the image in Preview and choose Tools, then Show Inspector. The EXIF and IPTC tabs show camera and source details when present.
These are fine for a first look. A photo that supposedly came from a phone but has no camera make or model is a small point against it, though it isn’t proof of anything.
How to check metadata on a phone #
Phone galleries show less, but they’re a start:
- iPhone: open the photo in Photos and swipe up, or tap the info button. You’ll see camera and lens details when they exist.
- Android with Google Photos: swipe up on the photo. Photos edited with Google AI tools carry a note about it in these details.
Neither shows the full C2PA record or raw XMP. For that, Expose AI runs a metadata pass on every image you give it. It reads embedded Content Credentials, XMP and generator tags, and recognizes tags from tools including Adobe Firefly, FLUX, Midjourney, Stable Diffusion, DALL·E, Google Imagen, Gemini, Ideogram, Leonardo.Ai and NovelAI. A credible AI claim in the metadata decides the verdict on its own. Camera and editor metadata show up in the “How we decided” card as evidence, never as proof. When the metadata has nothing to say, the on-device model makes the call from the pixels, and the image never leaves your phone.
We walk through phone-specific steps in checking a photo for AI on iPhone and on Android.
Why metadata is often missing #
You’ll find AI metadata far less often than you might expect. It disappears in ordinary ways:
- Uploads. Many social platforms and messaging apps re-compress images and drop most metadata.
- Screenshots. A screenshot is a new file with only your phone’s own data.
- Re-saving. Some editors and converters write a fresh file without carrying the old fields across.
- Deliberate removal. Anyone can strip metadata with free tools in seconds.
And some generators never write an AI marker in the first place.
Can metadata be faked? #
Plain EXIF, IPTC and XMP fields can be edited by anyone. Someone could add fake camera details to a generated image, or write an AI tag into a real photo. That’s why a camera make and model is only weak evidence of a real photo.
C2PA is harder to fake. Its manifests are cryptographically signed, and the specification uses hashes so that any change to the image or its provenance record can be detected (C2PA). A valid signature from a known issuer is strong evidence of where a file came from. But credentials can still be removed, and C2PA itself says Content Credentials don’t make value judgments about whether content depicts the truth. Our explainer on what C2PA and Content Credentials are goes further.
How to read what you find #
| What you find | What it tells you |
|---|---|
| Valid Content Credentials with an AI assertion | The file was made or edited with AI |
IPTC source type trainedAlgorithmicMedia |
Very likely AI-generated, unless someone added the tag |
| A generator name in Software or PNG text | Very likely AI-generated |
| Full camera EXIF, no AI fields | Leans real, but EXIF can be copied or faked |
| No metadata at all | Tells you nothing either way; check the pixels |
When the metadata comes up empty, move on to the visual checks and a detector. How to tell if an image is AI-generated covers the full routine.
Frequently asked questions #
Can you see if a picture is AI from its properties? #
Sometimes. If the file still has its metadata, the properties may show an AI tool in the software field or a digital source type marking it as generated. Most images that have been shared online have lost those fields, so empty properties don’t mean an image is real.
Does WhatsApp remove metadata from photos? #
Messaging apps generally re-compress photos sent in the normal way, and most of the original metadata doesn’t survive. If you need the original file, ask the sender to share it as a document or file attachment where the app allows it.
What does trainedAlgorithmicMedia mean? #
It’s an IPTC digital source type value meaning the image was created by an AI model trained on captured content. In plain terms, it’s the standard metadata label for a fully AI-generated image.
Is a photo with camera EXIF data definitely real? #
No. EXIF fields can be copied from another photo or typed in by hand. Full, consistent camera data makes a real photo more likely, but it isn’t proof, so combine it with a visual check and an AI detector.