Deepfake
Fake video or audio where AI copies a real face or voice
- Deepfakes come in three forms: face swapping, lip syncing a mouth to new dialogue, and voice cloning from a short recording.
- The method began with compressing a face and rebuilding it, moved through two networks competing against each other, and now leans on rebuilding a picture out of noise, which is what sharpened the quality.
- Blinking, fingers and other visual giveaways no longer work. In detection studies, people score close to a coin flip.
- So the response has shifted from catching fakes to proving what is real, layering provenance records, watermarks and detection models.
- The technology carries no guilt on its own, but making and spreading this material without consent is a crime in many countries. Your part is to hang up and call back on a number you already had.
Contents
1The analogy
When a skilled impressionist takes the stage, listeners with their eyes closed lose track of who is speaking. The impressionist has listened to that person hundreds of times and carried across everything, not just the pitch: where the breaths land, the little rise at the end of a sentence, the pause before a hard word.
A deepfake is a machine doing that same trick. It studies thousands of photos, video frames and recordings, and it copies more than the voice. The face, the expressions and the shape of the mouth come along with it, all at once and in sync.
On a stage the impression is a delight. The same skill, used to imitate a family member's voice on the phone and ask for money urgently, is a scam. What separates the two is not the skill. It is who agreed to it and what it is aimed at.
2In detail
The imitation splits three ways
Say deepfake and most people picture a swapped face, but there are three branches. The first is face swapping: the face of the person in a video is replaced with someone else's, with the body and setting untouched. The second is lip syncing, where the face stays and only the lips and jaw are redrawn to match new dialogue. Dubbing a film into another language sits here. The third is voice cloning, where a few seconds of recording is enough to produce a voice carrying the speaker's rhythm and breathing.
The damage today comes mostly from the three combined. Put a borrowed face on a video call and a borrowed voice over it, and from the other end of the line it is not distinguishable from talking to someone you know.
How the imitation is learned
Like the impressionist, the machine starts by gathering material. For a face, photos and video from several angles. For a voice, a short recording will do.
Next comes pulling out the features. Early methods trained a system to compress a face and then restore it, which left a compact summary holding the positions of the eyes, nose and mouth and the shape of an expression. Then came GAN (Generative Adversarial Network), where one side makes fakes and the other tries to catch them, each pushing the other to improve. Results grew far more convincing.
What lifted the quality to where it is now is diffusion, which starts from a screen of noise and restores the picture a little at a time. Because it begins blurred and refines over many passes, the places that used to fall apart, the ends of hair and the edges of shadows, come out natural.
Why your eyes cannot catch it
There used to be tricks. They rarely blink. The fingers are wrong. The skin on the neck does not match the face. None of them lasted. Once a giveaway is publicized, the next model is trained to close that exact gap.
There is a second problem. What actually reaches you is not a clean original. It is a clip that has been forwarded through four group chats and squeezed, or a voice flattened by phone-call quality. The awkward details get buried in compression, which makes them harder to catch than in any lab test. And the more shocking or infuriating the content, the more likely it spreads before anyone verifies anything.
Prove the real instead of catching the fake
So the response turned around. Rather than stamping the fakes, stamp what is genuine.
The first layer is provenance. The capturing device and the editing history are signed into the file, and later you can check whether that signature is still intact. The second is watermarking, where a mark that eyes and ears cannot pick up is embedded in AI-generated images and audio. The third is detection models, machines hunting for traces too fine for a person to see.
None of the three is complete. A signature falls away if someone films the screen. A watermark is fragile against cropping and recompression. A detection model loses accuracy against a generation method it has never met. Layering them is the baseline, not the ideal.
The line between the stage and the crime
The technology carries no guilt. It brings a late actor back into a film, gives a voice back to someone who lost theirs to illness, and dubs a lecture into a dozen languages. Those uses share two traits: the person involved agreed, and the result is labeled as synthetic.
The line is crossed when consent is missing. Making or spreading sexual images of someone, imitating a relative or a manager to demand a transfer, faking a candidate's statement during an election: these are punishable in many countries, and the laws are tightening. Your own part is simple. When a message or a call presses you about money or personal information, do not verify inside that same call. Hang up and dial a number you already had. A short family code word makes it faster still, and posting fewer photos and voice clips publicly means less raw material to work from.
3More precisely
The impressionist comparison holds for the result and not the process. A performer hears a sound and moves their own vocal cords. A deepfake learns the numerical distribution formed by the dots on a screen and the shape of a sound wave, then computes new dots and new waves. It is closer to redrawing from a learned distribution than to cutting and pasting the original. The imitation also has no limit of stamina or range: a performer can hold a voice for a few minutes, while a model can generate hours of it in any language it was trained on.
In terms, deepfake joins deep learning with fake. Face swapping began with autoencoders, which compress and restore the features of a specific person. Realism was lifted by the adversarial training of GAN, and diffusion models have since become the mainstream. Voice cloning extracts only the timbre from a short sample and lays it over new sentences, so the original speaker never has to have said those words. Detection tools return a probability rather than a verdict, which is why their output arrives as a percentage and not a yes or no. Last checked: 2026-09
Last verified: 2026-09
4Try it yourself
- Detect Fakes (spot the deepfake) ailearn.space Try to tell the real clips from the fake ones. Discovering how low your own score is happens to be the point of the exercise
- Which Face Is Real? ailearn.space Pick the AI-generated face out of a pair, and hunt for the giveaways yourself in backgrounds and earrings
- Content Credentials (check the origin) ailearn.space A tool for checking whether an image file carries a signed record of how it was shot and edited
5Common misconceptions
It's easy to think a deepfake gives itself away if you look closely, but actually in detection studies people score close to a coin flip.
It's easy to think only celebrities are targets, but actually a few photos from a public account and a short call recording are enough to build one of an ordinary person.
It's easy to think a detection tool settles the question, but actually the result is only a probability, and accuracy drops against a generation method the detector has never seen.
7One-line summary
In shortA deepfake is an AI impressionist copying a face and a voice whole, so the reliable defense is no longer your eyes but hanging up and calling back on a number you already had.
Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02