AI edited photo detector

Edits it looks for

• Faces changed with an AI image editor: a new expression, a new face, a different age.
• Instruction edits, where someone types a request ("put him in a suit", "remove the sign") and an AI editor redraws part of the picture.
• Inpainting: one region of a photo erased and regenerated by AI.

The editors behind them

The editors people use for this change every few months: Nano Banana, GPT-image, Flux Kontext, Qwen-Edit, Seedream and others. Checkobot is built to catch instruction edits from tools like these, and our errata page records the cases it gets wrong.

Where it's strongest

Edits to faces. A dedicated face analysis looks for the traces that face swaps and AI face edits leave behind, and it's where Checkobot is strongest.

For edits elsewhere in the picture, the whole-image analysis looks for traces the editor leaves behind. Small, local edits are harder to spot than edits that redraw the whole image.

AI editing or ordinary retouching?

Phones now retouch every photo a little, and editing apps have AI tools built in. Checkobot is looking for edits that change what the picture shows, not for exposure, colour or cropping. Strong beauty filters and heavy retouching sit in between: they can sometimes look like AI editing to the model.

Where it struggles

• Small edits away from any face, like a removed object in a corner, are the hardest to catch.
• Illustrated, cartoon, CGI or doll faces can trip the face analysis. It learned from photographs, so drawn faces can look odd to it.
• Images under 256 pixels on the short side, and heavily compressed JPEGs, are refused rather than guessed at.
• Documents, IDs and receipts aren't what it's built for. It can't tell you whether text on a document was changed.

Questions