Charles Stross is a highly dependable source of nightmare fuel in his SF/horror writings. He’s just as disturbing when he points out real developments about to go mainstream:
AI assisted porn video is, it seems, now a thing. For those of you who don’t read the links: you can train off-the-shelf neural networks to recognize faces (or other bits of people and objects) in video clips. You can then use the trained network to edit them, replacing one person in a video with a synthetic version of someone else. In this case, Rule 34 applies: it’s being used to take porn videos and replace the actors with film stars. The software runs on a high-end GPU and takes quite a while — hours to days — to do its stuff, but it’s out there and it’ll probably be available to rent as a cloud service running on obsolescent bitcoin-mining GPU racks in China by the end of next week.
(Obvious first-generation application: workplace/social media sexual harassers just got a whole new toolkit.)
But it’s going to get a whole lot worse.
What I’m not seeing yet is the obvious application of this sort of deep learning to speech synthesis. It’s all very well to fake up a video of David Cameron fucking a goat, but without the bleating and mindless quackspeak it’s pretty obvious that it’s a fake. Being able to train a network to recognize the cadences of our target’s intonation, though, and then to modulate a different speaker’s words so they come out sounding right takes it into a whole new level of plausibility for human viewers, because we give credence to sensory inputs based on how consistent they are with our other senses. We need AI to get the lip-sync right, in other words, before today’s simplistic AI-generated video porn turns really toxic.
(Second generation application: Hitler sums it up, now with fewer subtitles)
There are innocuous uses, of course. It’s a truism of the TV business that the camera adds ten kilograms. And we all know about airbrushing/photoshopping of models on magazine covers and in adverts. We can now automate the video-photoshopping of subjects so that, for example, folks like me don’t look as unattractive in a talking-heads TV interview. Pretty soon everyone you see on film or TV is going to be ‘shopped to look sexier, fitter, and skinnier than is actually natural. It’ll probably be built into your smartphone’s camera processor in a few years, first a “make me look fit in selfies” mode and then a “do the same thing, only in video chat” option.
