The advent of machine learning in AI has the potential to disrupt – or even render obsolete – entire industries, and displace millions of jobs. From banking and finance, to advertising and even software development, any work that previously required the output of human thinking may one day be taken over by a self-learning (and possibly self-aware) artificial neural network.
However, machine learning is also being used for other purposes, with troubling implications. An article on Motherboard written by Samantha Cole tells us about how people are using AI to create fake celebrity porn videos, called deepfakes. Using stock photos, videos, and Google image search, they train a neural network – a group of interconnected computer nodes – to manipulate porn videos, superimposing celebrities’ faces into the bodies of the videos’ performers.
The above screenshot shows Star Wars actress Daisy Ridley’s face in a porn performer’s video. The video was created using FakeApp, a desktop tool based on the original deepfakes algorithm.
Early examples of deepfake videos were hilariously terrible. But with the AIs getting better at manipulating videos, and with more people joining in, deepfakes are increasingly becoming more realistic. The release of FakeApp has lowered the barriers to entry, by allowing users with no experience in writing machine learning algorithms to create their own deepfakes.
In a more wholesome vein, Abhimanyu Ghoshal wrote an article on The Next Web about Lyrebird, an experimental voice synthesis tool created by researchers from the Montreal Institute for Learning Algorithms. Now open for public beta, the technology resembles Adobe’s VoCo in that it attempts to recreate your voice using audio samples that you record. However, instead of needing twenty minutes’ worth of samples, this tool needs only one minute.
Here’s a snippet of my AI doppelganger’s voice after hearing me speak 30 phrases. You can still hear distortions in the generated audio, but give it enough phrases to learn (say, a hundred) and the quality of the artificial voice improves, until it becomes almost impossible to tell apart from the real one. Hearing my computer use my voice to say something I never said is rather creepy.
Granted, these examples of machine learning AI are fascinating, and both technologies do have useful applications. Imagine amateur filmmakers digitally resurrecting long-dead actors for their pet projects, or creating a personalized digital assistant whose voice is indistinguishable from your own. But in the wrong hands, the implications are disturbing. If you think about how many selfies and videos we have already uploaded to our social networks, it’s not implausible that voice synthesis and the deepfake algorithm may someday be (ab)used together to create fake, but convincing videos, to blackmail or discredit people.
One thing’s for sure: we’re getting closer to the future depicted in Black Mirror.