Can You Detect AI Images From Metadata?
Metadata can confirm an image was generated. It can never confirm that one was not — and that asymmetry matters.
The short answer
Metadata can prove that an image was generated, when the generator left a marker and nobody removed it. It can never prove that an image was not generated, because absence of a marker is the normal state of a file that has been through almost any processing.
That asymmetry is the whole story, and most confusion about AI detection comes from ignoring it.
What generators actually leave behind
The markers vary enormously by tool, and knowing which is which tells you how much to read into a result.
- Stable Diffusion interfaces (Automatic1111, Forge, InvokeAI) write the full prompt, seed, sampler and model into PNG text chunks — the most complete record of any tool.
- ComfyUI embeds the entire node workflow as JSON, which is even more detailed.
- Adobe Firefly and recent Photoshop attach signed C2PA content credentials.
- Hosted services such as Midjourney and DALL·E typically deliver files with little or nothing embedded, because the image is re-encoded and served by a web pipeline.
- Some tools write only a Software or CreatorTool tag naming themselves.
Why a clean file tells you nothing
By the time an image reaches you, it has usually been through several stages that discard metadata without anyone intending to: a platform re-encode, a screenshot, a resize, an export from an editor, a messaging app's compression.
So a file with no AI markers might be a photograph, a generated image that was cleaned deliberately, or a generated image that simply passed through Instagram. Nothing in the metadata distinguishes those three cases, and any tool that claims otherwise from metadata alone is guessing.
Where watermarks fit
This is precisely the gap pixel watermarks are meant to fill. Google's SynthID encodes a signal into the pixel values themselves, so it survives the re-encoding, resizing and screenshotting that destroy metadata. Because it lives in the picture rather than the container, no metadata tool removes it, and no metadata tool detects it either — reading it requires the detector built for it.
The same distinction applies to the least-significant-bit watermarking on this site: it is in the pixels, not the metadata, but it is fragile rather than robust, which is why it is documented as a way to mark files you control rather than as a provenance system.
What this means in practice
If you are checking an image someone sent you, metadata is worth reading: a positive result is strong evidence, and prompts and workflows are unambiguous. Treat a negative result as no information at all.
If you are publishing your own generated work, assume the opposite direction: the markers are real, they are readable by anyone, and if you do not want your prompt published then remove it before the image leaves your machine.
Check an image for AI markers
See exactly which AI markers an image carries before you publish it. This tool only reads — nothing is changed, and nothing is uploaded.
Open AI Metadata CheckerFrequently asked questions
Is there a reliable AI image detector?
Visual classifiers exist and are useful in aggregate, but they produce false positives and false negatives on individual images and their accuracy shifts as generators change. Provenance data, when present and signed, is more dependable than any classifier.
Does removing metadata make an AI image undetectable?
No. It removes one signal. Pixel watermarks and visual classifiers are unaffected, and platforms apply their own labelling using systems they do not publish.
Can I remove SynthID?
Not with a metadata tool. SynthID is encoded into pixel values rather than stored as metadata, so no metadata cleaner touches it — including this one.