AI Picture Detector: What It Catches and What It Can't

This AI picture detector takes whatever you have: a meme, a product shot, a poster, a group photo, a screenshot. First your browser reads the file's Content Credentials and metadata. If the credentials say an AI tool generated the picture, you get the answer straight away and no detection model is called. If they say it was edited with AI, you get Edited with AI.

Otherwise the picture goes to a detection model (Sightengine, with Hive as a backup if Sightengine fails), the same engines as our AI image checker on the home page, and you get Likely AI-generated, Likely a real photo or Can't tell, with a range and the reasons.

It catches whole images made by the major generators. Sightengine detected 49 of 50 AI images in our test, and 41 of 42 from the newer generators (everything in the table above except Stable Diffusion XL and Midjourney v6). Hive detected 45 of 50.

The table lists every model we tested, including recent ones such as GPT Image 2, Nano Banana Pro, FLUX.2 [max], Midjourney v8.1 and Seedream 5.0 Pro. Those are good numbers on a small test, not a promise about your picture. Each answer comes with a range for that reason.

No AI picture detector is equally good at everything. In our test the weak spots were one generator family, copies that had been screenshotted or reposted, and pictures that are only partly AI; fakes made without AI are outside what it can judge at all.

• FLUX was the hardest family for both engines in our test: Sightengine caught 8 of 9, Hive 6 of 9.
• Screenshots and reposts lose detail and every bit of metadata. See checking a viral picture or screenshot.
• Partial edits and composites. Sightengine describes its model as telling whether an image was entirely generated by AI or is a real image. A real photo with one AI-added object, or an AI face pasted into a real scene, sits in between, and our test didn't include any.
• Fakes made without AI. A picture doctored the old way in an editor, or a real photo with a false caption, isn't AI-generated. A detector is the wrong tool for those.

Memes are usually screenshots of screenshots with a caption on top, so there are no credentials left and everything rests on the pixels. An AI picture detector scores the whole picture, caption band included. If the question is really about the photo inside the meme, look for a copy without the caption and check that too. We didn't test memes, posters or graphic designs, so a Can't tell on them is a fair outcome, not a malfunction.

Shops and support teams sometimes receive pictures that might be generated: a damaged item for a refund, an unboxing, a product shown in a home. If a buyer's picture comes back Likely AI-generated, ask for a short video or a second photo from another angle before you decide, and don't turn the claim down on the score alone. A real customer with a really broken item is exactly who a false alarm hurts.

If you need a record, share the result to get a link. A shared result can show a small thumbnail of the picture (you can turn that off); if you don’t share, we keep only a hash of the picture and the verdict.

Each engine draws its own line. In our test we counted a Sightengine score of 0.5 or more as AI, and for Hive we used the vendor's recommended 0.9. A score just below a line and one just above it are almost the same evidence, which is why our checker shows a range instead of one number that flips between yes and no.

Sometimes, but don't count on it. Sightengine describes its model as telling fully generated images from real ones, and we didn't test composites. If the file's credentials or metadata record an AI edit, you get Edited with AI.

FLUX. Sightengine caught 8 of 9 FLUX images and Hive 6 of 9. Sightengine caught every image from every other family.

Not by itself. Renders and digital paintings made without a generator are human work, but they don't look like camera photos either. We didn't test them, so treat answers on them with extra caution.

The detection model doesn't: Sightengine says its model looks only at the pixels. Our checker reads Content Credentials and metadata separately, in your browser, and lists what it finds as reasons of their own.