← All guides Provenance

Why EXIF metadata cannot detect AI images

Metadata is stripped by almost every platform on upload, which means it fails on precisely the images people need to check. What survives, and what to use instead.

· 7 min read · Best AI Image Detector

Because it is gone. Platforms strip metadata on upload, so a picture that reached you through social media or a chat app has no EXIF left to read, whatever it was made by.

What EXIF is and what it holds

EXIF is a block of data a camera writes into an image file: make, model, lens, exposure, focal length, orientation, timestamp, and sometimes location. Editing software adds its own entries, so a file can also record which application last saved it.

None of it is signed. Every field is plain text that any tool can write, rewrite or remove. That is the first problem: EXIF was designed to be helpful to photographers, not to resist anybody.

The stripping problem

Almost every platform removes metadata when you upload. The reasons are sound: EXIF frequently contains GPS coordinates, and publishing somebody's home location because they posted a photo of their garden is a privacy failure. Stripping it also reduces file size.

The side effect is that provenance disappears at the first hop. An image that went from a camera, through an editor, onto a platform and into a message has been stripped at least once, usually twice.

Camera Edited Uploaded Screenshotted EXIF metadata Survives Survives Lost Lost Content Credentials Survives Survives Lost Lost Invisible watermark Survives Survives Lost Lost Pixel signal Survives Survives Survives Survives

A screenshot is a re-render, so it also discards the file structure entirely and starts a new one with no history.

What is left at each stage of an ordinary journey. Only one row is still readable by the time an image reaches a group chat.

The three failure modes

  1. Absent. The common case. No EXIF at all, which tells you the file has been through a platform and nothing about how it was made.
  2. Present but meaningless. A camera make and model written by an editing tool, or a timestamp reflecting the last save rather than the capture.
  3. Present and false. Any field can be written by hand. Setting a camera model on a generated image takes seconds and defeats every check that trusts EXIF.

The third one is why metadata-based verdicts are actively unsafe rather than merely weak. A tool that reports an image as genuine because it carries camera EXIF can be defeated by anybody who knows the field exists.

When EXIF is still worth reading

None of this makes metadata useless. It makes it useful in a narrow set of situations, and knowing which ones saves time.

Where metadata earns its place
SituationUseful?Why
A file supplied directly by a photographerYesNothing has stripped it, and it cross-checks the account
An original from a claimant or a sourceYesDevice and timestamp can be checked against the story
A picture downloaded from social mediaNoStripped on upload
A screenshotNoA new file with no history
An image forwarded through a chat appRarelyMost apps strip and recompress
A file where EXIF claims a cameraWith careThe field is unsigned and trivially written

The pattern is that EXIF works when you are close to the source and fails as soon as an image starts travelling. It is a chain of custody tool rather than a detection tool.

What replaced it

Two things, doing different jobs. Content Credentials are the signed successor to unsigned metadata: same idea, with cryptography attached so the record cannot be rewritten without detection. Where they survive, they are stronger than EXIF ever was.

Pixel analysis is the other half, and it exists because credentials also get stripped. Reading the image data itself is the only method that still works after an image has been through a platform, which is where the difficult questions actually arise.

How to preserve metadata when you want it

  • Send files as documents, not as photos. Most messaging apps have a file option that transfers the original bytes.
  • Ask for the original, from the device that took it, rather than a copy that has been through anything.
  • Avoid screenshots entirely when the provenance matters. A screenshot discards the file and creates a new one.
  • Archive the raw capture, not just the export, if you are the photographer.
  • Check your export settings. Some tools strip metadata by default on save for web.

The reverse mistake

There is an error that runs the other way and it is becoming more common. Somebody finds camera metadata on an image and treats it as proof the picture is genuine.

Writing EXIF is trivial. Free tools will set a camera make, model, lens and timestamp on any file in seconds, including on an image that came straight out of a generator. A checker that reports genuine because it found a camera name is reporting what somebody typed.

The asymmetry is worth stating plainly. Absent metadata means nothing. Present metadata is weak evidence at best, and it is weakest exactly where somebody has a motive to fabricate it. Neither direction supports a confident conclusion on its own.

Content Credentials exist precisely to close this gap. The difference between the two is a signature, and that single difference is what separates a field anybody can write from a record that cannot be altered without detection.

Questions people ask

Can EXIF data tell me if an image is AI generated?
Only in the narrow case where a generator wrote a field saying so and nothing has stripped it since. That combination is rare. On a stripped image there is nothing to read, and on a file with hand-edited fields you are reading whatever somebody chose to put there.
Do AI generators add metadata?
Some do. Several tools write software identifiers or full Content Credentials into their output. That is genuinely useful when the file arrives intact, and it disappears the moment the image is uploaded anywhere, which is usually before you see it.
Why do platforms strip metadata?
Privacy first: EXIF often contains GPS coordinates, and publishing a user's home address alongside their photo would be a serious failure. File size is the secondary reason. Neither is done to frustrate verification, and both have that effect.
Is missing EXIF suspicious?
No. It is the normal state of almost every image on the internet. Treating an absent metadata block as a warning sign would flag nearly everything, including every photograph anybody has ever posted anywhere.
Can I trust a timestamp in EXIF?
Only loosely. It reflects the camera clock, which is often wrong, unset or in a different time zone, and it can be rewritten by any tool. Use it as a cross-check against a story rather than as an established fact.
What should I ask a source for?
The original file from the device, sent as a document rather than as a photo. That preserves whatever metadata and credentials exist and gives a detector the best version of the pixels to work with. Both kinds of evidence improve at once.