AI Metadata Checker

An AI metadata checker opens the data stored inside an image file and shows you anything that says how the picture was made. Drop a file into the checker above and your browser reads the EXIF, XMP and IPTC metadata and any Content Credentials (C2PA) on your own device, without uploading the image. You see the camera, the software, the capture date, any field that mentions AI, and who signed the credentials.

It answers a narrower question than a detector: not "does this look AI-generated?" but "what does this file say about itself?" When the file has a signed AI record from a signer on the C2PA trust lists, that is close to proof and the AI metadata checker says so. When it doesn't, you get the facts and a link to run our AI image detector, which judges the pixels.

Four places in a file carry most of the AI metadata you will find, and this checker looks at each of them before it tells you anything.

• Content Credentials (C2PA). A signed record that some generators attach, such as OpenAI's GPT Image and Google's Nano Banana. It can say the picture was generated or edited by AI and names the signer. It is the only one of the four that is cryptographically signed.
• IPTC DigitalSourceType. A standard field whose value trainedAlgorithmicMedia means generated by AI, and compositeWithTrainedAlgorithmicMedia means partly AI. Anyone can write it.
• Credit and similar text fields. Google, for example, has written "Made with Google AI" or "Edited with Google AI" into pictures made or edited with its AI tools.
• Software and generation fields. The name of the generator in Software or CreatorTool, or the prompt and settings that some Stable Diffusion apps save in a PNG text field called parameters.

Everything except signed credentials is plain text in the file: easy to write, easy to delete, easy to copy into a real photo.

Camera details and AI labels split cleanly in our test, but most AI files carried no label in their metadata. 37 of 46 real photos (screenshots counted separately) had a camera make and model, and 0 of 40 AI originals did. The software field was the same: 33 real photos had one, no AI original did.

On the AI side, 24 of 40 originals carried AI Content Credentials, and only 4 carried a plain metadata label, all of them IPTC DigitalSourceType in Midjourney files. No real photo had an AI label, and 2 real photos, from a Google Pixel 9 and a Google Pixel 10, carried Content Credentials saying a camera took them. The table above has every row.

Missing metadata tells you almost nothing, because it disappears so easily. 12 of the 40 AI originals in our test had no metadata or credentials at all: 11 copies from public AI image datasets, which had been re-saved along the way, and 1 Midjourney picture we generated ourselves. So did 9 of the 46 real photos, mostly stock pictures and heavily recompressed copies.

Screenshots are the extreme case: 0 of our 14 screenshots, real and AI, had any metadata or credentials, because a screenshot is a new file drawn from the screen. Many social and messaging apps also strip metadata from uploads. If the AI metadata checker finds nothing, the next step is a model that judges the pixels; our screenshot guide covers what else you can try.

Camera make, model and exposure settings are a good sign but not proof, because they can be copied from one file to another with free tools. In our test no AI original carried them, so their presence pointed the right way every time; a forger only has to copy them once to break that pattern.

That is why our checker never calls a picture real on camera metadata alone. Camera Content Credentials are different: when a phone such as a recent Google Pixel signs the capture, the signature covers the image data, and editing the picture breaks it or adds a new entry to the record. The Content Credentials checker explains how to read those records.

Platform labels such as Meta's AI info on Instagram, Facebook and Threads are based largely on the same metadata. Meta says it labels images when it detects industry-standard AI indicators, naming the AI-generated information in the C2PA and IPTC standards, or when the person posting discloses AI use.

So if a real photo picked up an AI label, check the file first. An edit made with a generative AI tool, such as removing an object, can write an AI-edited marker into the metadata, and Meta has said that minor AI retouching triggered its earlier "Made with AI" label. The AI metadata checker shows which field is there. It won't remove anything, and if the picture really was edited with AI, the label is telling the truth.

Use the full detector whenever the metadata doesn't settle the question, which in our test was most of the time: trusted credentials settled only 14 of 50 AI images without a model. The AI image checker reads the same metadata first and then asks Sightengine to judge the pixels, so it also works on screenshots and stripped copies.

The full check uses one of your 3 free daily checks when it needs the model; this metadata checker is free and unlimited, because nothing leaves your browser. For pictures of people, see the AI photo detector guide; for the numbers behind both tools, the accuracy report.

Yes, with no limit and no sign-up. It runs entirely in your browser and never uploads the image, so it costs us nothing to run.

Only a signed one from a signer on the C2PA trust lists. Plain fields such as IPTC DigitalSourceType or Software can be written or deleted by anyone, so treat them as a claim to check, not a verdict.

Some generators don't write any, and screenshots, uploads, re-saves and edits remove what was there. 12 of the 40 AI originals in our test had none, most of them copies that had been re-saved before we got them.

Yes. It reads JPEG, PNG, WebP and HEIC files. Some formats simply hold less metadata, and a PNG saved from a screenshot usually holds none.