The Creepiest Thing Online This Week Is An AI That Creates Digital Ghosts
We often think of the paranormal as something born from the natural world. Malevolent spirits, mysterious beasts and cosmic horrors all terrify us because they're part of a natural order that humans can scarcely understand, control or even escape from. But what if the technology we created — the impossibly complex and powerful algorithms running out lives — began to manifest their own phantoms? What if the ghosts were coming from inside the machine?
This is the idea behind the latest project from Deep Angel co-creators Matt Groh and Ziv Epstein: AI Spirits. Here, Groh and Epstein trained a deep neural network not to erase things from images, but to insert phantasmal, glitched-out ghosts into perfectly normal images. The end result is, well, take a look for yourself:
The digital creepiness on display here — something our society has only just begun to explore as glitches and errors become the new bumps in the night — is similar to that of creepypasta like Slender Man or BEN DROWNED, where a dark, unknowable evil injects itself into something familiar and comforting.
"In AI and robotics, we talk a lot about "the uncanny valley," where stuff is human-like enough to make it seem plausible, but small deviations from our expectation of humans lead to a very creepy result," Epstein writes in an email to me. "AI Spirits in particular explores this uncanny valley by actually leaning into and appreciating the erroneous, distorted and bizarre output of these deep neural nets."
While learning more about something creepy usually makes it less so, delightfully, understanding how AI Spirits is able to conjure up digital phantoms in otherwise normal images only serves to make it more unsettling.
Where Deep Angel is tasked with removing people from images, Groh and Epstein fed all of that image data into AI Spirits so it could learn how to insert rough, but believable, approximations of people. Spirits, you might say.
"This model maps scenes with missing people to scenes with people. So, it is not exactly detecting humans and altering them to glitched out spirits. Instead, it's a process of disappearing humans and then reimagining humans from what it knows about people in images," writes Groh in an email. "The algorithm adds in phantasms where appropriate based on understanding of size and location of people in images. These phantasms are colors and texturized based on the AI's understanding of the scenery in rest of the image."
In short, Groh and Epstein trained an AI whose sole purpose is to creep out humans.
Of course, what's truly terrifying here isn't AI, but rather, humanity itself. It's all too easy to think of AI as independent from humanity, a set of impartial rules that could eventually lead to a Skynet-like future that ultimately dooms the human race. But Groh and Epstein believe that the true terror of AI is how it can hide and obfuscate the worst impulses of human nature.
"We encode our own value systems and physical reality into our technology, and especially AI, which learns from the past. Thus AI is often times malevolent because it is simply regurgitating something structurally problematic," writes Epstein. "Just as facial recognition software and risk recidivism algorithms fail on racial and gender minorities, AI can be quite evil when its only sees a small but specific part of the world."
"I don't think malevolent forces lurk in [AI Spirits], but if machine learning models are applied indiscriminately and used inappropriately, there are certainly many unjust, malevolent things that can come out of it," writes Groh.
The true monster, as always, is us.
