Deepfake detector for photos
Drop in a photo. Checko looks closely at the face and at the whole picture, then tells you AI or No AI detected, with a one-line reason. It squints at pixels, never at people.
Show Checko an image
What it checks
Checkobot detects AI-generated and AI-manipulated images, with a dedicated face-forensics model for deepfakes. In a photo, a deepfake usually means one of three things:
- Face swaps: one person's face placed onto another person's head.
- AI face edits: a real face changed with an AI image editor, from a new expression to a different age.
- Fully AI-generated pictures, with or without people in them.
How the verdict works
Two detectors look at every image. The face analysis runs when there's a face at least 64 pixels across; the whole-image analysis always runs. If either one crosses its threshold, the verdict is AI and the result says which detector fired. Otherwise the verdict is No AI detected.
No AI detected means neither detector found signs at its threshold. It doesn't prove a photo came straight from a camera.
How we measure it
We test on frozen benchmarks the model never saw in training, including face-swap engines held out of training entirely, and we check that wrong flags don't pile up on any one skin-tone group.
We'll publish measured catch and false-flag rates, each with its dataset and thresholds, once our current test round is finished. Until then we'd rather say nothing than quote a number we haven't confirmed.
Photos from the wider web are messier than identity selfies, and both detectors make more mistakes on them. Confirmed mistakes go on our errata page.
Photos now, video later
Most deepfakes people worry about are videos. Checkobot checks still images today. A frame from a video often carries the same traces a swapped photo does, so you can pause on a sharp frame of the face, save it and check that. The video deepfake check page explains how, and what a single frame can and can't tell you.
Where it struggles
- Talking-head avatars and lip-sync from new engines are our weakest area: it misses many of them.
- 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, and it doesn't check audio, live video or liveness (printed photos held up to a camera, screen replays).
Questions
What is a deepfake?
A picture or video in which someone's face, expression or voice has been made or changed with AI so it looks like they said or did something they didn't. Checkobot checks the picture side of that: swapped, edited and generated faces in still images.
Is it free?
Yes. You can check 3 images a day without signing up, or 10 a day with a free account.
Do you keep my photo?
Not on free checks: the image is checked in memory and then discarded, and we keep a fingerprint of the file and the verdict. Paid plans keep the images you check in your private history, with a switch in Settings to turn that off. The privacy page has the details.
Is the verdict proof?
No detector gives proof. The verdict is a signal from two models, each with a known error rate. Use it alongside the usual checks: where the image came from, who posted it first, and what else is in the frame.
Does it check videos?
Not yet. In the meantime you can check a still frame from the video.