How does the AI image detector work?
An open AI model (SwinV2, trained on AI made and human made pictures) looks at your picture, shrunk to 256 x 256 pixels, and gives a score from 0% to 100% AI made. It runs in your browser with onnxruntime-web, so the picture never leaves your device.
How accurate is it?
On our test of 203 pictures of known origin it was right on 91%: it caught 85% of AI made pictures and cleared 96% of human made ones. See How we test for every number and every miss.
Is a reading proof?
No. Every reading is an estimate. Detectors make mistakes in both directions, and new generators appear all the time. Never accuse anyone on a detector reading alone.
What kinds of pictures can it judge?
Still pictures: photos, digital art, paintings, drawings, anime style art and screenshots. It cannot check video, voice or text, and it gives one reading for the whole picture, so a real photo with one AI edit may read as human made.
Which generators can it detect?
Its author trained it on a large private set of pictures but does not list the generators. In our test it caught pictures from Midjourney, DALL-E, ChatGPT, Gemini, Imagen, Flux, Stable Diffusion, Firefly, Ideogram and others, though not every one.
Why is the first check slow?
The first time, your browser downloads the detector (about 103 MB) from this site, in 5 parts; if the line drops, Try again carries on where it stopped. After that it is kept in the browser and starts at once. Each picture then takes about 5 to 15 seconds on a computer and longer on a phone.
Does it work on a phone?
Yes, on recent phones. It needs several hundred MB of free memory while it works; on older phones the page may say the detector could not be opened, and a computer will work instead.
Are my pictures uploaded or stored?
No. Pictures are read in your browser and forgotten when you close the page. See Privacy.
What do the lamps and the loupe do?
The loupe shows any part of the picture 4 times closer. Colour boost pushes colour and contrast up to reveal colour noise and smears; Edges shows only fine detail. They help you look; they do not change the score.
Why did my real photo read as AI?
Filters, beauty modes, AI upscaling, strong noise removal and very smooth digital painting can all make a real picture look like AI to a detector. See why detectors get it wrong.
Can it read the AI label in a file?
Yes. It looks for the IPTC "trained algorithmic media" label, Content Credentials (C2PA) and generation settings in PNG files, and tells you when it finds one. Most posted pictures have had these removed.
Is there a limit or a sign up?
No limit, no sign up and no credits. It is free because the work happens on your own device.