What Is a Deepfake?
A deepfake is a real photograph or video in which someone's face has been replaced with another. That definition matters more than it sounds: the scene is genuine, the body is genuine, only the face is synthetic — and everything about detecting one follows from that split.
Where the word comes from
“Deepfake” joins *deep learning* and *fake*. It surfaced around 2017 on Reddit, where a user of that name posted face-swapped video. The technique escaped the research world and became consumer software within about two years.
The term now gets stretched to cover any manipulated media, but the specific meaning is face replacement — and keeping it specific is useful, because different fakes need different checks.
How it actually works
In broad strokes: a model is trained on many images of a target face, learning to reconstruct it from any angle and under any lighting. It then regenerates that face onto the frames of an existing video, matching pose and expression frame by frame.
What used to take days of training on a large photo set now runs from a handful of images, and in some cases live during a video call. That collapse in cost is the whole story of the last two years.
Deepfake vs AI-generated image
These get confused constantly, and the distinction is practical rather than pedantic.
An AI-generated image was never photographed. There was no camera, no scene, no light. Everything in it is synthetic, so detection looks for global signals: missing sensor noise, flat frequency profile, absent capture metadata. That is what our AI image detector measures.
A deepfake is a real photograph with one region replaced. The noise, lighting and metadata of the original are still there — which means global checks are weak, and you instead look for *inconsistency between regions*: the seam, the mismatched noise floor, lighting that disagrees. Hence a separate deepfake detector.
What deepfakes are actually used for
Public discussion focuses on political disinformation. The measured reality is different: the overwhelming majority of deepfake material is non-consensual intimate imagery, targeting private individuals far more often than public figures.
The second-largest category is fraud. Fake video calls in hiring, voice and face cloning of executives to authorise transfers, and romance scams where the “person” can now appear live on camera.
Political fakes exist and matter, but they are a smaller share than the coverage suggests — and they are the category most likely to be caught quickly, because many people are looking.
What this means for you
Two habits are worth forming. First, a video call is no longer proof of identity — real-time swapping exists. Ask for a profile turn or a hand passed in front of the face; occlusion is where these systems still break.
Second, a photo of someone holding a sign with your name proves nothing. That was decent verification in 2020 and is trivially faked now.
If you need to check a specific image or clip, our guide to spotting deepfakes covers what to look for by eye, and the detector handles what the eye cannot reach.
Key points
- A deepfake is a real image with a synthetic face — not a fully generated picture.
- That split is why deepfakes and AI images need different detection approaches.
- Training now takes a handful of images, and can run live on a video call.
- Most deepfake material is non-consensual intimate imagery, not politics.
- Video calls and sign-holding photos are no longer proof of identity.
Frequently asked questions
Check an image yourself
Reading about it only goes so far. Our AI image detector runs twelve independent methods and shows you every score, not just a verdict.
More from the blog
Try it on your own image
Free account, 3 checks included, no card required.