These people don’t actually exist

But barring a few similar case studies, the world of AI deepfakes has remained largely underground, which is precisely why ThisPersonDoesNotExist.com is trying to bubble it to the surface.

The resemblance is uncanny. With every hit of the refresh button, a new human face is generated. The spectrum is endless; all ethnicities, ages, hair, and makeup styles are represented. Even small personality traits are somehow captured through the facial expression the person chooses to put on while being captured by the camera. Wrinkles around the eyes suggest this person might be a smiley one—crevices between the brows hint to this person’s hardship life. There is only one thing that really threads these faces together, which is that they are absolutely fake.

The website’s founder, a software engineer at Uber named Philip Wang, relied on research done by Nvidia, an American tech company focused on the gaming industry, in order to generate this infinite pool of fake faces. Nvidia has in fact been rendering its AI production of fake faces over the past several years and in 2014 released a paper showing some of its first attempts (the image of the left-hand side). Four years later, Nvidia’s algorithms can create an absolutely undetectable human face within milliseconds. These images are created using millions of real images that are processed through a neural network called generative adversarial networks (GAN), to then manufacture new examples. Hair movement is studied, nose shapes, eyebrows, lips and cheeks. All to fabricate faces that are realistic and accurately represent certain DNA traits.

The images forged on ThisPersonDoesNotExist.com are not being displayed from a pool of readymades every time the page is refreshed. Instead, the person is quite literally being made on the spot every single time. In a recent Facebook post, Wang wrote, “Each time you refresh the site, the network will generate a new facial image from scratch”.

There are obviously positive and negative implications of this fast-developing system. On the plus side, it can help advance virtual worlds in the ability to imitate and slightly fabricate real-life things, like humans or natural beings, plants and even objects, aiding creatives as they develop the next generation of virtual reality and anchor it in something relatable and realistic. And as reported by the Fast Company last week, GAN is already helping artists like Mario Klingemann generate portrait paintings using his algorithm ‘Memories of Passerby’, which are currently selling at Sotheby’s for $50,000.

On the more negative side, the ability to generate realistic images of people at an unprecedented rate poses a risk for both the development of deepfakes such as Jordan Peele’s Obama video and the creation of more undetectably realistic images of bots, tweeting, posting on Instagram and potentially being used as a tool for propaganda. Instead of imitating real people’s accounts, as bots often do today using profile pictures and changing the handle ever so slightly (a lower case instead of upper, an underscore instead of a dash), GAN images open possibilities for the creation of an infinite army of real fake people online.

Just like any other groundbreaking innovation, GAN’s latest achievement is here to remind us that how we chose to utilise technology is up to our discretion and that our approach is key. For starters, education around deepfakes desperately needs to exit the underground and enter the surface level. Tools that can be used to detect fake images must be created and awareness campaigns, such as Wang’s ThisPersonIsNotReal.com need to become part of the everyday.