Horses to Zebra's - AI is Killing our Grasp on Reality

On the AI and us beat, Sandra Upson in Wired tells us how "Artificial intelligence is killing our sense of reality". She outlines how image, video, text and sound modification is getting easier to automatically produce in any style by AI and wonders how long before we will not be able to tell real from fake images and sound.

The impetus comes from research (and code cycleGAN) at UC Berkeley by Jun-Yan Zhu* Taesung Park* and colleagues which tries to take and image in one class, and manipulate it to look like an image in a second class. They use two learning tools - one to try to do the image changes, and a second to try to tell if the image belongs in one category or not, and let the tools compete to train them for better and better results. The original usage was making computer game streets look like google maps for better training of self driving cars.

The initial results look like a lot of fun - with zebra horse manipulations: (There is also a nice 2 minute papers video on this, and that whole series is great)

Not a perfect horse -> zebra change but survives a quick glance.

Upson also notes how this has already been used for modifying porn videos to have the faces of famous actress's by a Redit user Deep Fakes (See article by Samantha Cole "AI-Assisted Fake Porn Is Here and We’re All Fucked"), and that people can now generate videos of famous people saying things they did not say. Fake Amazon, Yelp and other reviews are now relatively easy to produce, and getting more and more believable. One possible worry is a "near future in which a convincing fake video of President Trump ordering the total nuclear annihilation of North Korea goes viral and incites panic, like a recast War of the Worlds for the AI era.".

Fundamentally, the question Upson gets at is will be able able to distinguish real from fake information in the future, given the speed at which fakes are getting better. - The answer currently appears to be No - it is just a matter of time.