Deepfake Detector

A deepfake takes a real person and swaps or re-animates their face inside an existing photo or video: a celebrity's face on someone else's body, a politician "saying" something they never said, a friend's face on a dating profile that isn't theirs. The background, the body and the lighting are usually genuine; only the face is synthetic.

A fully AI-generated image is different: the whole picture is invented by a generator like Midjourney, DALL·E, Stable Diffusion or Grok, and the person in it does not exist. Both leave statistical traces a camera does not, so this detector flags either. The distinction matters for how to read the score, below.

The model was trained on millions of real photographs and AI-generated images. When you upload a picture it measures patterns in the pixels that generators and face-swap pipelines leave behind: the blending boundary where a swapped face meets skin and hair, texture that is smoother or noisier than the rest of the frame, and colour and lighting on the face that disagree with the body. It also reads the file's metadata for AI-generation tags or Content Credentials.

You get one probability. Above about 85% the model is confident the image is synthetic. Below 15% it is confident the image is a real photo. In between it is a lean, not an answer, and the checks below decide.

1. Edges. Soft, smeared or slightly blurred boundaries where the face meets the hairline, ears, jaw or neck, while the rest of the image is sharp.
2. Skin and light. A face that is lit from a different direction than the body, or skin tone that does not match the neck and hands.
3. Eyes and teeth. Pupils that are not the same size, reflections that differ between the eyes, teeth that are too even or merge together.
4. Accessories. Glasses, earrings and hair that warp, double or vanish near the face.
5. Expression. A mouth that does not quite match the emotion of the eyes, or a face that stays oddly still and smooth.
6. Resolution. A face that is crisper or softer than everything around it, a sign it was rendered separately and pasted in.

Full walkthrough: how to spot a deepfake. For edited but not swapped photos, the edited photo detector reads what the file records.

This tool analyzes still images, not video files. To check a video, pause it on a clear, front-facing frame, take a screenshot, and upload the frame. Then do it again on two or three other moments, ideally where the head turns or the mouth opens. A swapped face is rendered frame by frame and usually fails on at least one.

Screenshots strip metadata, so the score you get is from the pixels alone. That is fine for this purpose; it is exactly the case the model is built for.

• You get a probability score in seconds, a metadata read of the same file, and a second independent check via the metadata checker. Free, no account, image deleted after analysis.
• You don't get video or audio analysis, a named source for the fake, or a legal-grade verdict. A careful face swap on a real photo can score in the middle band because most of the image is genuine; run the visual checks when the score is not decisive.
• For faces that don't exist at all (fully generated profile pictures) the fake profile picture detector is the focused version of this tool.