Deepfake

A name for synthetic media

"Deepfake" is a portmanteau of "deep learning" and "fake". Wikipedia traces the term to late 2017, when a Reddit user going by the name "deepfakes" was behind its origin. ^[1] The word now covers a broad class of material: images, video or audio that has been edited or generated with artificial intelligence, AI-based tools or audio-video editing software. Such output is commonly grouped under the heading of synthetic media. ^[1][2]

How a face swap works

Most face-swap deepfakes rely on a generative adversarial network, or GAN — a pair of neural networks trained against each other. One network, the generator, produces images; the other, the discriminator, tries to tell generated images apart from real ones. Each round pushes the generator toward output the discriminator cannot distinguish from genuine footage, and the two improve together. In a face swap, the generator learns to render a chosen person's face in the poses, lighting and expressions of a target video, while the discriminator penalises results that look synthetic. Korshunov and Marcel, assessing the threat deepfakes pose to face recognition, described face replacement as being done with a pre-trained generative adversarial network. ^[2]

The technical turning point came in the mid-2010s, when generative adversarial networks made convincing synthetic imagery practical. Wikipedia treats that development as a key moment in the evolution of deepfakes. ^[1][2]

Earlier experiments

The technology did not appear from nowhere. Wikipedia traces the underlying research to academic work that began in the 1990s, including a 1997 system called Video Rewrite, led by Christoph Bregler. ^[1] Two later programs are often listed as forerunners: Face2Face, published in 2016, and Synthesizing Obama, published in 2017. ^[1]

What changed at the end of 2017 was access. Once the method was packaged so that non-specialists could run it, the output spread through online communities. Early community output included celebrity faces swapped onto the bodies of actors in pornographic videos, and many clips with Nicolas Cage's face swapped into films. ^[1][2] The non-consensual sexual material was prominent enough that Korshunov and Marcel cited recent scandals over celebrities' faces being swapped onto pornographic videos as the reason automated detection was needed. ^[2]

From one application to open-source tools

In January 2018 a proprietary desktop application called FakeApp launched, putting face swapping in front of general users. ^[1] By 2019 it had largely been replaced by open-source alternatives such as Faceswap, DeepFaceLab and DeepfakesWeb. ^[1] Research tools circulated alongside them: the same 2018 paper by Korshunov and Marcel released the first publicly available set of Deepfake videos, generated from the VidTIMIT database with open-source GAN-based software, so that detection methods could be tested on real examples. ^[2]

Why it matters

Deepfakes matter on two fronts that pull in opposite directions. The first is harm: the same methods make it possible to put a person's face into footage they never appeared in, and the early pornographic face swaps show what that means for the people depicted. ^[1][2] The second is detection. Because a deepfake is judged by eye against ordinary footage, researchers treat the problem as a moving target: every improvement in generation leaves new traces for detectors to learn from, and every improvement in detection sets a new test for generation to pass. ^[2]

The term itself has stretched with the technology. It began as the handle of one poster and a description of face-swapped video, and it now serves as the everyday word for AI-edited images, video and audio of many kinds. ^[1][2]