Free AI Image Detector

AI Image Detector

Is this picture AI made? Drop one or many on the light box. A detector that runs on your device gives each one a reading, and a loupe lets you look closer yourself.

DROP A PICTURE ON THE LIGHT BOX

Add one picture or many: photos, art, screenshots. Each gets a reading from a detector that runs on your device, then you can look closer with a loupe.

    Pictures never leave your device. No account, no credits.

    The detector is about 103 MB and loads once from this site, the first time you check a picture; your browser keeps it for next time. Every reading is an estimate, never proof. How well it does, measured.

    91%right on 184 of 203 test pictures
    85%of 98 AI made pictures caught
    96%of 105 human made pictures cleared

    What this AI image detector does

    You give it a picture: a photo someone sent you, a piece of art you want to buy or share, a screenshot from a feed. It gives back a reading from AI made to human made, with how sure the detector is, in about 5 to 15 seconds per picture. You can put several pictures on the light box at once and see every reading side by side, then tap any picture to look closer with a 4x loupe and two lamps that bring out smears, noise and texture.

    The detector is an open model that its author trained on a large set of AI made and human made pictures; we measured how it does on photos, art and anime style pictures before putting it here. It runs inside your browser, so the picture is never uploaded anywhere. There is no account to make and no credits to run out of.

    How well it does, measured

    We tested the detector on 203 pictures with a known origin from Wikimedia Commons: AI pictures that their uploaders made with named generators (Midjourney, DALL-E, ChatGPT, Gemini, Imagen, Flux, Stable Diffusion, Firefly, Ideogram, Grok and others), and photos, paintings and drawings made by people, most of them from before AI pictures were common. It was right on 91% of them: it caught 85% of the AI pictures and cleared 96% of the human made ones. The full table, every picture and every miss are on How we test (measured 2026-11-19).

    A reading is an estimate, never proof. It can be wrong in both directions, it was not trained on every generator, and heavy editing, filters, screenshots and squeezing by chat apps all make it less reliable. Use it as one sign among others, never to accuse someone.

    What it can judge, and what it cannot

    It is made for still pictures: photos, digital art, paintings, drawings, anime style art and screenshots of them. It looks at the whole picture at once, shrunk to 256 x 256 pixels, and gives one reading for the whole picture. That means:

    How to use it

    1. Press Choose pictures, drop pictures on the light box, or copy a picture or screenshot and press Ctrl+V (Cmd+V on a Mac). On a phone, choose from your photos or screenshots.
    2. The first time, the detector loads once (about 103 MB). On a fast connection that takes under a minute; on a slow one it can take several minutes, so Wi-Fi helps. After that it starts at once on this device, even for the next visit.
    3. Each picture gets a tab: AI MADE, LEANS AI, LEANS HUMAN or HUMAN MADE, with how sure the detector is.
    4. Tap a picture to look closer: the reading on a scale from human made to AI made, any AI label hidden inside the file, the loupe and the lamps.
    5. Tick two pictures and press Compare 2 to see which one the detector rates as more likely AI made.

    Look for yourself too

    People who check pictures by hand zoom in and look for things that are hard for generators: text and lettering, hands and teeth, earrings that do not match, patterns that repeat, edges that melt into the background, light that comes from two places. The loupe shows any spot 4 times closer. Colour boost pushes colour and contrast far up, which makes the soft colour noise some generators leave easy to see; Edges shows only the fine detail, so smooth areas that should have texture stand out. Both also show ordinary photo noise and JPEG blocks, so they are clues, not a verdict. The guide how to tell if a picture is AI walks through each check.

    Labels hidden inside the file

    Some AI apps write a label into the picture file: the IPTC "trained algorithmic media" source type, Content Credentials (C2PA), or the prompt and settings in a PNG. When the detector finds one, it says so under the reading. Most social apps remove these labels when a picture is posted, so a missing label means nothing. More in AI labels in picture files.

    Why on your device

    Most AI image detectors upload your picture to their server, and many give a handful of free checks before asking you to pay. Here the work happens in your browser with an open model and an open runtime (onnxruntime-web), so it costs nothing per picture and the picture stays with you. The price is the one-time download and about 5 to 15 seconds of work per picture; a computer or a recent phone does it best.

    Questions

    Is this AI image detector free?

    Yes. It runs on your own device, so there are no credits, no daily limit and no account.

    Are my pictures uploaded?

    No. The picture is read and checked inside your browser. The only downloads are the detector itself (once) and the program that runs it.

    How accurate is it?

    On our test of 203 pictures of known origin it was right on 91%: 85% of AI pictures caught and 96% of human made pictures cleared. Every reading is still an estimate, not proof.

    Can it check screenshots?

    Yes, and that is often the only way to check a picture from a feed. Screenshots and pictures squeezed by chat apps give somewhat weaker readings.

    More questions