Is This Image AI-Generated? A Four-Step Check

To answer "is this image AI-generated?", go through four checks in order and stop as soon as one gives a clear answer: look for Content Credentials, read the metadata, run a detection model, then use your own eyes. The checker on this page runs the first three for you; the fourth is yours.

The order matters because each step is either cheaper or more certain than the one after it. Credentials and metadata cost nothing and never leave your browser. A detection model costs a call and returns a probability. Your eyes are free, but on newer generators they are the least reliable of the four. It is the same order our AI image detection tool on the home page follows.

Content Credentials, built on the C2PA standard, are a signed record that some tools attach to the files they create. If the record says the image was created by a generative AI model, that is the strongest single signal there is, and our checker gives its answer from it without calling a model.

In our test, credentials identified 24 of the 50 AI images, every one an original file, and never pointed at a real photo. The signers were OpenAI, Google, Black Forest Labs and BytePlus (ByteDance); Midjourney files carried none. A missing credential is not evidence either way: screenshots, re-saves and most social platforms strip them. The content credentials checker shows the full record when there is one.

EXIF is the camera's notebook: make, model, lens, exposure, time. In our set, 37 of 50 real photos named a camera and none of the 50 AI images did. That makes camera details a useful sign of a real capture, but a weak one: metadata is easily stripped, plenty of real photos lose it when shared, and ordinary tools such as ExifTool can rewrite it.

Metadata can also point the other way. The IPTC vocabulary includes "Created using Generative AI" and "Edited using Generative AI", and Google Photos writes that kind of marker into pictures changed with tools like Magic Editor. When our checker finds an AI-edit marker, the answer is Edited with AI.

When the first two steps don't settle it, the image goes to a model that judges only the pixels. Our main engine, Sightengine, identified 49 of 50 AI images and flagged none of the 50 real photos. Hive, our backup, identified 45 of 50, also with no false alarms on real photos.

Those are good numbers from a small test, not guarantees. You get a range rather than a single figure, and a Can't tell when the evidence is thin. The accuracy report has the details and the limits.

Hunting for six fingers or garbled text used to work well. Newer generators make far fewer of those mistakes, so a clean-looking image proves nothing, and odd details turn up in real photos too. Our visual checklist shows where to look. Use what you find as context for the other three steps, not as a verdict.

Most of the time one step settles it. When the steps point different ways, the checker combines them into a single answer with a range, and these are the four patterns you will see most often.

• Credentials say an AI tool generated it → Likely AI-generated, with the signer named. No model call needed.
• Camera metadata present and the model says real → Likely a real photo, with the camera listed among the reasons.
• A screenshot with no metadata and a middling model score → Can't tell, with a wide range. That is an honest result, not a failure.
• Credentials or metadata say it was edited with AI → Edited with AI: a real capture that an AI tool changed.

Not necessarily, and not necessarily real either. More than half of the AI images in our test had no credentials, including every Midjourney file and every screenshot. Move on to the metadata and the detection model.

It may be, but steps 1 and 2 have nothing to read: a screenshot carries no Content Credentials or camera metadata. The answer rests on the detection model, which still caught 9 of 10 AI screenshots in our test (Sightengine). Give it a little less weight than a result on the original file.

Yes. EXIF fields can be edited with free tools, which is why camera details count as a supporting reason, never as proof. Signed Content Credentials are much harder to forge, because the signature breaks if the record is changed.

Signed credentials that say AI, but only some original files have them. For everything else, the detection model does most of the work, with a range that tells you how sure it is.

If the credentials already say an AI tool generated the image, a model call would add nothing. Skipping it is faster and keeps the image in your browser.