The word “deepfake” sounds like science fiction.
It isn’t.
It’s math. Specifically, it’s a set of mathematical operations so computationally intensive that they were impossible to run on consumer hardware until about a decade ago — and now run on a laptop in an afternoon.
Understanding what deepfakes actually are — not the fear-version, not the hype-version, but the real technical and human story — is the first step toward not being fooled by them.
Where the Word Came From
“Deepfake” is a portmanteau. Deep, from deep learning — the branch of artificial intelligence that uses layered neural networks to find patterns in data. Fake, from what the technology produces.
The term was coined in 2017 on Reddit, where a user named “deepfakes” began posting AI-generated videos that placed celebrity faces onto other people’s bodies. The technology wasn’t new — researchers had been working on face synthesis for years. But the accessibility was. For the first time, someone without a PhD in machine learning could produce a convincing fake video.
The genie had left the bottle.
The Technology: GANs
At the heart of most deepfake technology is something called a Generative Adversarial Network — a GAN.
A GAN consists of two neural networks competing against each other.
The first network is the Generator. Its job is to create fake images — faces, in the case of most deepfakes. It starts by generating random noise and gradually learns to shape that noise into something that looks like a real face.
The second network is the Discriminator. Its job is to look at images and decide: is this real or fake? It’s trained on thousands of real photographs so it develops a sense of what authenticity looks like.
Here’s where it gets interesting. The Generator and the Discriminator are locked in a competition. The Generator produces an image. The Discriminator evaluates it. If the Discriminator correctly identifies it as fake, the Generator gets penalized and adjusts its approach. If the Generator fools the Discriminator, it gets rewarded.
This competition runs for millions of iterations. Each cycle, both networks get better. The Generator gets better at producing convincing fakes. The Discriminator gets better at detecting them. And because they’re competing against each other, they push each other toward the limits of what’s possible.
The result, after enough training, is a Generator that can produce faces indistinguishable from photographs of real people — because it has been optimized, over millions of iterations, specifically to fool the most sophisticated detector it could be trained against.
How a Face Gets Synthesized
When you look at a face generated by a modern GAN, you’re looking at the output of a process that works roughly like this:
The Generator starts with a random point in what researchers call latent space — a mathematical space where every possible face exists as a set of numbers. Move in one direction through this space and faces get older. Move in another direction and hair color changes. Move in another and the face rotates.
The Generator maps a point in this space to a pixel-by-pixel image — a face — through a series of mathematical transformations. Early layers handle broad features: overall shape, skin tone, face geometry. Later layers handle fine details: pore texture, eyelash placement, the subtle shadows in skin folds.
The result is a face that has never existed. There is no person behind it. No childhood photographs, no DNA, no history. Just math — a set of numbers that happens to produce, when rendered, something your brain reads as a human being.
Face Swapping: A Different Beast
The face synthesis we just described creates entirely new faces — the thispersondoesnotexist.com model. But the deepfakes that cause the most concern aren’t new faces. They’re existing faces, transplanted.
Face swapping deepfakes take a real person’s face — extracted from existing video or photographs — and map it onto another person’s body in motion. This is technically harder than face synthesis because it requires:
Tracking the target face through every frame of video as it turns, expresses emotion, and changes lighting. Extracting the source face’s features — not just its appearance but its geometry, its texture, its response to light. Rendering the source face onto the target’s movements in real time, matching lighting, expression, and perspective. Blending the result seamlessly into the existing video so the boundary between real and synthetic is invisible.
Early deepfakes failed at most of these steps — edges were blurry, lighting was wrong, expressions lagged slightly behind the motion. The artifacts were visible to a trained eye.
Modern deepfakes fail at very few of them. The artifacts are still there — they’re just smaller, subtler, and require more specific knowledge to find.
Voice Cloning: The Sound of Someone Else
Video deepfakes are the ones that get the press. But voice deepfakes may be more dangerous.
Voice cloning technology can now replicate a person’s voice from as little as three seconds of audio. Given a short sample, AI models can synthesize speech in that voice saying anything — any words, any tone, any emotion.
The implications are significant. A phone call from what sounds like your CEO asking for an urgent wire transfer. A voicemail from what sounds like your bank asking you to verify your account. A message from what sounds like a family member in distress asking for help.
The human ear is even less equipped to detect cloned voices than the human eye is to detect fake faces. We’re wired to respond to voices we recognize. The emotional response — trust, urgency, concern — triggers before the analytical response has a chance to engage.
The Democratization of Deception
Here is the number that matters:
In 2019, creating a convincing deepfake video required a high-end workstation, several days of compute time, and significant technical expertise.
In 2024, it requires a consumer laptop, a few hours, and a free or low-cost application that handles the technical complexity automatically.
The barrier to deception dropped from high-expertise to near-zero in five years. It will continue to drop.
This is not a prediction. It’s a description of what has already happened.
What This Means for You
The existence of deepfakes doesn’t mean you should trust nothing. That’s paralysis, not protection.
It means you need to upgrade your perception.
The same way the invention of the printing press made literacy essential — the same way the invention of photography made visual literacy essential — the invention of convincing synthetic media makes perceptual literacy essential.
Not paranoia. Literacy.
Knowing what to look for. Knowing when to verify. Knowing the difference between a face that looks right and a face that is real. Building the cognitive habits that let you move through a world of synthetic content without being fooled by it and without being paralyzed by suspicion.
That’s trainable. That’s the skill.
The technology of deception advanced. Your perception needs to advance with it.