- A new crop of websites shows the disturbing potential of deepfake technology
- The sites present pictures of faces, cats and buildings that are completely fake but look incredibly real.
- One of the site's creators says even people without computer programming experience can use freely available tools to create fake pictures in a couple of hours.
- The Uber engineer behind another one of the sites says he made the site to "raise public awareness" about the new AI technology.
Deepfake technology has caused a stir with the eerily realistic but completely fake depictions it can produce of celebrities, such as Scarlett Johansson appearing in porn videos and former President Barack Obama calling Trump a "dips---."
Now, a crop of websites have emerged that highlight just how pervasive and consequential the technology is likely to become.
ThisPersonDoesNotExist.com serves up a rotating gallery of pictures of different faces — but each face is completely fake and computer-generated.
The site can create these AI-based faces using something called a generative adversarial network (a GAN). As The Next Web explains, these GANs pit two algorithms against each other — a generator and a judge. The generator creates fake depictions of something and attempts to fool the judging algorithm into believing it's legit. Each item that the GAN spits out is an iteration of where the generator was successful in beating the judge.
However, ThisPersonDoesNotExist.com uses a specific algorithm called StyleGAN, developed by AI company Nvidia. The code was first published in a research paper, but is publicly available for use on GitHub. (Nvidia declined to comment because the paper is currently under peer review, during which it can't talk about it with the media "according to submission rules.")
Several other sites have used StyleGAN to develop similar sites showing fake cats, fake anime characters and even fake Airbnb listings.
A developer behind one of these websites explained that he took on the project in order to demonstrate an important point about AI and neural networks: This technology can be used to easily fool people into believing fake and doctored images. Experts have raised concerns that these sophisticated tools could be weaponized for furthering fake news and hoaxes.
"This means that just about anyone with a couple hours to kill could create something just as compelling as I did," Chris Schmidt writes on his website, ThisAirbnbDoesNotExist.com. "[AI is] now sufficiently advanced that they can often fool folks, especially if they’re not looking very hard."
Check out all the different ways the technology is being used to create fake pictures that raise troubling questions about our perception of reality:
How it works: Every time you refresh the website, the StyleGAN creates a new AI-generated face. The generator uses a dataset of faces from Flickr.
Created by: Philip Wang, former Uber software engineer. He shared the website in a public Facebook group about artificial intelligence and deep learning.
"I have decided to dig into my own pockets and raise some public awareness for this technology," Wang wrote in his post in the Facebook group.
How it works: Nvidia's code on Github includes a pretrained StyleGAN model, and a dataset, to apply the code to cats.
Created by: Two websites have since emerged.
The other version, TheseCatsDoNotExist.com, was created by Australian developer Nathan Glover. He posted the link to his site on Twitter, and wrote he had generated over 30,000 fake cats. The website shows rows of these cats at the same time, but they change each time the page is refreshed.
How it works: Each time the page is refreshed, the website shows a new fake Airbnb listing, complete with AI-generated room pictures, name and face of the host, and listing description. "They are all fevered dreams of computers,"the website says.
Created by: Christopher Schmidt, an engineer working on open source code at Google. In the "About" section on his Airbnb listing's site, Schmidt writes that he was able to produce StyleGAN content without any "real experience with neural networks" or his own "fancy computing resources."
"This means that just about anyone with a couple hours to kill could create something just as compelling as I did," Schmidt writes. "While there are parts of the experience that are weak, overall, I think that it works: the listings are often dubious, but typically plausible enough that they would survive a quick glance."
See the rest of the story at Business Insider