Provenance
Is this image AI? What the file itself says
Guessing from the picture no longer works. Hands, text and reflections have been fixed for a while. What still works, when it is there, is the record the generator wrote into the file. Most major generators now sign their output, and many leave the prompt. The Metadata tab reads all of it. This page is about what to look for and, just as important, what an empty result does and does not mean.
The signs, strongest first
- A C2PA manifest. Content Credentials signed by the generator, naming the tool and the action that produced the picture. A valid manifest is close to proof, because it is cryptographically signed. The report shows the manifest and whether the signature checks. The C2PA guide explains what a manifest contains.
- IPTC digital source type. A single field whose value says the image was created by a trained algorithm, or was a photo edited with one. Several generators and some editors write it.
- XMP provenance fields. Creator tool names, generation identifiers and history entries that name an AI feature.
- A prompt in a text chunk. Locally run image models commonly write the full prompt, seed, sampler and model name into a PNG text chunk. Read it and there is no doubt.
- A software name. Weakest on its own. Many ordinary editors also write one, and it is trivially changed.
How to check
Open the tool, pick the Metadata tab and drop the image in. The provenance block sits near the top of the report when it is present. Content Credentials are read and verified. A prompt shows in the text or description fields. Download the report if you need to keep a record.
What an empty result means
Nothing. This is the part worth reading twice. A screenshot of a generated image carries no tags. A generated image saved through a social network carries no tags. A generated image that went through a metadata cleaner, including this one, carries no tags. The absence of a provenance record is not evidence that a picture is a photograph. It is only evidence that there is no record.
The reverse is more reliable. A signed manifest that names a generator is hard to fake and easy to verify, so its presence means what it says.
Watermarks in the pixels
Some generators also put a signal into the picture itself, invisibly. SynthID is the best known. It is not a tag and this site cannot read it: only the company that made the watermark runs a detector for it, and the detector is theirs. If you need that check, use their detector. The SynthID guide covers where that line sits.
About percentage detectors
Sites that return "83% likely AI" from the pixels alone are making a statistical guess with a classifier, and classifiers are wrong often enough that a percentage should not decide anything that matters. A signed record in the file is a fact. A score is an opinion.
Which generator wrote what
Generators differ in what they leave behind and it changes with each release. The generator metadata guide lists the common tools and the tags each one is known to write.
Common questions
Can metadata prove an image is AI generated?
A valid, signed C2PA manifest naming a generator is close to proof. Other tags such as an IPTC digital source type or a prompt are strong signs. A software name alone is weak.
Can metadata prove an image is a real photo?
No. Any picture, generated or not, can have its tags removed by a screenshot, a social network or a cleaner. An empty report means there is no record, nothing more.
Does this site detect SynthID?
No. SynthID is a watermark in the pixels, not a tag, and only its maker's detector can read it.