How to Prove You Took a Photograph

Why arguing with the detector fails

When a photograph is accused of being AI-generated, the instinct is to challenge the tool that said so. It rarely works, because the person deciding usually cannot evaluate that argument and has no reason to trust your reading of it over the score in front of them.

The stronger move is to change what is being examined. A detector produced a probability from pixels. You can produce material it never saw, which no generator could have produced, and which is tedious enough to fake that its presence is persuasive on its own.

The strongest evidence, in order

Not all of it carries equal weight. If you can produce the first item, the rest is usually unnecessary.

• The RAW file. Generators do not output RAW. It is sensor data in a camera-specific format, with its own embedded thumbnail and capture settings, and it is the single most convincing artefact you can present.
• The untouched original as it came off the camera or phone, before any export or edit.
• The frames around it. Bursts, brackets and the shots immediately before and after, with their own sequential filenames and timestamps.
• Editing history. Intermediate saves, layered project files, or an editor history showing the image being built from the original.
• The wider context: other images from the same session, and the original memory card contents if you still have them.

Why sequence is more convincing than any single file

One file can be constructed. A coherent sequence is a different problem entirely.

A real shoot leaves a trail: sequential frame numbers, timestamps minutes or seconds apart, consistent camera and lens settings drifting the way real conditions drift, near-identical frames differing in small ways. Fabricating one plausible file is easy; fabricating forty that agree with each other is not something anyone does casually.

This is why the advice to keep everything matters more than any single precaution. The value is in the corpus, not the file.

Metadata is your evidence here

This site removes metadata, so treat this as the deliberate exception it is: do not strip metadata from files you might need to defend.

The EXIF block on a camera original is your provenance record. Camera make and model, lens, exposure, ISO, focal length, timestamps, and on some bodies the shutter count. None of it is proof alone and all of it is forgeable by someone determined — but it is internally consistent in ways that are laborious to fake, and it is what a reviewer will actually ask to see.

There is a second reason it matters. A file that has been stripped looks the same as a file that never had anything, and you cannot demonstrate afterwards which one yours was.

Content Credentials, working the other way round

Provenance metadata is usually discussed here as the thing that gets you wrongly labelled. It has a second use that runs the other direction.

A C2PA manifest can record a capture and editing history rather than a generative one. Where your camera or editor supports it, that produces a signed chain describing how the image reached its current state — evidence that asserts something, rather than merely failing to assert anything.

The distinction is worth holding onto. Removing a manifest makes a file silent about its history. It does not make the file claim camera origin, and nothing about an absence supports a positive claim.

The workflow that keeps both

Privacy on published files and provable authorship look like conflicting goals. They only conflict if you keep one copy.

Archive originals untouched — RAW where you have it, camera JPEGs otherwise, metadata intact, backed up somewhere you will still have access to in two years. Then remove metadata from the copies you publish, which is where the privacy exposure actually lives, since the published file is the one strangers can download.

Those two halves do not interfere with each other. What creates the conflict is editing in place and publishing the only copy you own.

What does not work

A few approaches come up often and are worth ruling out before you spend effort on them.

• Re-running the image through other detectors until one returns a favourable answer. Reviewers discount this immediately, and it concedes that detector output is the right basis for the decision.
• Adding EXIF to a file after the fact. It is trivially detectable, and being caught doing it is far worse than the original accusation.
• Removing metadata to make the image look less machine-made. Detectors never read it, so nothing changes except that you have discarded your own evidence.
• Posting the image at higher resolution as proof. Resolution is not origin, and generators produce large images.