Fake news: you ain’t seen nothing yet
▻https://www.economist.com/news/science-and-technology/21724370-generating-convincing-audio-and-video-fake-events-fake-news-you-
Mr Klingemann’s experiment foreshadows a new battlefield between falsehood and veracity. Faith in written information is under attack in some quarters by the spread of what is loosely known as “fake news”. But images and sound recordings retain for many an inherent trustworthiness. GANs are part of a technological wave that threatens this credibility.
Audio is easier to fake. Normally, computers generate speech by linking lots of short recorded speech fragments to create a sentence. That is how the voice of Siri, Apple’s digital assistant, is generated. But digital voices like this are limited by the range of fragments they have memorised. They only sound truly realistic when speaking a specific batch of phrases.
Generative audio works differently, using neural networks to learn the statistical properties of the audio source in question, then reproducing those properties directly in any context, modelling how speech changes not just second-by-second, but millisecond-by-millisecond. Putting words into the mouth of Mr Trump, say, or of any other public figure, is a matter of feeding recordings of his speeches into the algorithmic hopper and then telling the trained software what you want that person to say.
When pressed for an estimate, he suggests that the generation of YouTube fakes that are very plausible may be possible within three years. Others think it might take longer. But all agree that it is a question of when, not if. “We think that AI is going to change the kinds of evidence that we can trust,” says Mr Goodfellow.
Yet even as technology drives new forms of artifice, it also offers new ways to combat it. One form of verification is to demand that recordings come with their metadata, which show when, where and how they were captured. Knowing such things makes it possible to eliminate a photograph as a fake on the basis, for example, of a mismatch with known local conditions at the time. A rather recherché example comes from work done in 2014 by NVIDIA, a chip-making company whose devices power a lot of AI. It used its chips to analyse photos from the Apollo 11 Moon landing. By simulating the way light rays bounce around, NVIDIA showed that the odd-looking lighting of Buzz Aldrin’s space suit—taken by some nitwits as evidence of fakery—really is reflected lunar sunlight and not the lights of a Hollywood film rig.