Checking AI images on iPhone and Android

Why there is no app

An app would be a worse version of the same thing. The detector is a model file and a runtime, and both work in a mobile browser without any of the ceremony an app store adds: no install, no permissions dialogue, no update to forget.

It also avoids the trade most detection apps make. An app that sends your camera roll to a server for scoring has taken custody of your photographs, and the privacy policy governing them is somebody else's. Running in the browser removes that question rather than answering it.

The practical cost is the first load. The model is a real download, so the first check on a new device takes longer than the ones after it. Once cached it stays available, and subsequent checks start immediately.

Checking a photo from your camera roll

1. Open the tool in Safari or Chrome

   Any current mobile browser works. Adding it to your home screen gives you an icon that behaves like an app without being one.

2. Tap the upload area

   The picker offers your photo library and the camera. On iOS it also offers Files, which is where an attachment you saved from Mail or Messages will be.

3. Choose the original, not a screenshot

   If the image arrived in a chat, save the attachment rather than screenshotting the conversation. A screenshot re-encodes the picture and throws away much of what the check reads.

4. Wait about four seconds

   Longer on the first run while the model downloads. The phone does the work, so a newer handset finishes faster.

5. Read the region map

   Turn the phone sideways or pinch to zoom. The map is where a small edit inside an otherwise real photo shows up.

Checking something you were sent

This is the common case and the one where people lose the most information. Messaging apps recompress images aggressively on send, and a picture that has passed through two apps has been re-encoded twice before it reaches you.

Where the app offers it, save the original attachment rather than the preview. On iOS, tapping through to the full-size image and saving to Files preserves more than sharing directly into another app. On Android, the download from the conversation is generally better than the thumbnail cache.

If a link is all you have, paste the URL instead. The tool can fetch a direct image link or pull the pictures out of a public page, which usually retrieves a better copy than a screenshot of the same page would.

What phone cameras do to a score

A modern phone photograph is not a straightforward record of light hitting a sensor. It is a composite assembled from several exposures, denoised, sharpened and tone-mapped by software before you ever see it.

That processing is machine learning too, and it leaves its own texture. The detector is reading for signs that pixels were produced by software rather than sampled from the world, and computational photography sits awkwardly close to that line.

None of those five images is fake. The spread is entirely an artefact of how much software touched the pixels between the sensor and the file, which is why a middling score on a phone photograph is weak evidence of anything on its own.

Battery, data and what actually runs

The model download is the only significant data cost, and it happens once per device until the browser clears its cache. After that a check costs no network at all, which means the tool works on a plane or with the connection off.

Running the model uses the phone's processor for a few seconds per image, so a long batch will warm the handset and use some battery. A batch of twenty-five is comfortable on a recent phone and slower on an older one, where splitting it into two smaller batches is the better approach.

• Works offline after the first load, because nothing is sent anywhere.
• No permissions beyond the file picker you tap yourself.
• Same result as on a desktop, since it is the same model and the same settings.
• Add to home screen for an icon, without installing anything.
• Private browsing works, but the model downloads again each session.

Comparing against your own camera

The most useful thing you can do on a phone costs one extra check. Photograph something ordinary in the same mode as the picture you are investigating, and run it. That gives you the number your device produces for a photograph you know is real.

That baseline is worth more than any published threshold, because it accounts for your specific handset, your specific camera app and the processing both apply. A suspect image sitting near your own baseline is unremarkable. One sitting thirty points above it is a question.

It also protects you from the most common mistake people make on mobile, which is treating a middling score as a soft accusation. Most phones produce middling scores on ordinary photographs, and knowing your own number removes that whole category of false alarm.

Where a phone check falls short

Small screens hide region maps. A single hot tile in a twelve-tile grid is obvious on a laptop and easy to miss on a handset held at arm length, so zoom into the map rather than glancing at it.

Batches are also harder. Selecting twenty-five images from a photo library is fiddly, and comparing them across a narrow screen loses the thing that makes a batch informative, which is seeing the spread. Where a set matters, do it on a larger screen.