The Ugly: A Downward Spiraling Race
Like any deep learning task, data collection
is crucial for realistic deepfakes. Kevin Roose, a tech columnist at NYTimes, tried the FakeApp by collecting video footage of himself with entertaining if unconvincing results
. For now, it implies that only public figures with plenty of high-quality videos will be the targets of convincing deepfakes. But deepfakes of politicians are potentially the most impactful forms of fake news and could negatively affect, for example, the 2020 US elections
This development also places a heavier weight on technology that detects deepfake content. That’s why we have curated 6 tools that can help you detect deepfakes:
But deepfakes are like computer viruses: as soon as someone finds a way to detect them, another will find a way around it
. The hide and seek between fraudsters and detectors spiral together downward as if the two constitute a gigantic GAN.
The Good: Humanlike AI for a new form of communication
Despite its dark side, deepfake also has positive uses that can help society, such as enabling new forms of communication. Take voice generation: Google Assistant can now speak like John Legend
by utilizing Wavenet
, a generative model for speech generation. Startups Lyrebird
can learn to talk like you with a few hours of speaking audio. Bernard Marr reports that Baidu’s technology takes only 3.7 seconds to clone a voice
. Soon we’ll have smart speakers that talk like our favorite singers or our own virtual selves who represent us when we are out of office.