The word blends deep learning and fake, and its author is anonymous: in late 2017 a Reddit user called "deepfakes" began posting videos in which he had pasted celebrities' faces onto the bodies of performers in pornographic videos. The use has not changed much: a 2019 report by the company Deeptrace estimated that 96% of the deepfakes in circulation were pornographic, nearly always of women who had not consented.
The original technique was modest: two autoencoders with a shared encoder and one decoder per face. A frame of person A is encoded into the common latent space and decoded with B's decoder: the expression is A's, the face is B's. Then came GANs, diffusion models and voice cloning. Buckner (2024) recalls the 2018 video in which Jordan Peele made Obama deliver a script of his own, and the "arms race" between detectors and forgers that followed (2024, p. 58). He also marks the difference from adversarial examples: those fool a classifier; deepfakes fool human perception (2024, p. 206).
The philosophical problem is epistemological. A recording counted as evidence because it was hard to fake. Fallis (2021) puts it in terms of information: if fake videos become cheap and convincing, each video carries less information about the world. Chesney & Citron (2019) added the reverse side, the "liar's dividend": when anything could be fake, whoever appears in an authentic video can deny it. The harm is not only that we believe what is false, but that we can no longer believe what is true.
Buckner reads the same models as an artificial imagination, able to compose scenes never seen. What, then, separates imagining from forging: the technique, or the claim that what is produced is a record?