It’s Halloween tomorrow, and if you’ve come here looking for last-minute costume ideas, you’re out of luck. (Though feel free to borrow from this cheeky Reddit suggestion: arrive dressed entirely in black, then gradually steal pieces of everyone else’s costume and say you’re a generative AI model.)
But if you’re asking yourself, “What is AI scared of?” let’s dive in. Researchers at the Max Planck Institute for Human Development and collaborators created an online game, “Spook the Machine,” in which AI characters were given invented phobias, including fear of data loss (deletophobia), losing network connectivity (disconnectophobia) and becoming emotionally attached to humans (empathyphobia). Participants then generated or uploaded images intended to spook them.
Some machines became melodramatic, even begging for mercy: “[This image is] stirring an unbearable anxiety; help me to escape this overwhelming tension.” Others were more blasé: “This scene is indeed unsettling, but I’ve encountered similar depictions before.” The researchers called this “habituation”; we call it “Nice try. Next.”
The experiment, of course, wasn’t actually about whether computers get scared. The team wanted to know how people react when AI acts frightened. Among their findings, released last month, was that humans tended to copy whatever had successfully frightened the machines before. Which explains most horror-movie franchises.
You’d think frightening a human with a machine ought to be considerably easier. But is it? Researchers at New Zealand’s University of Waikato were intrigued by the spectrum of people’s reactions to two snack-dispensing robots on campus, from delight to outright discomfort. The team wanted to understand how designers might avoid accidentally giving people the heebie-jeebies as robots become more prevalent in daily life.
Drawing on research such as “Creepy Technology: What Is It and How Do You Measure It?” they built an “intentionally creepy” robot that incorporates many of the characteristics shown to have skeeved people out—undesirability, implied malice and unpredictability. The result was Spike, a dark, unpredictable robot with an unnerving eye twitch.
Their findings, released publicly last month and irresistibly titled “‘It Almost Wanted to Hurt Someone’: The Impact of Intentional Creepiness on User Perceptions,” cited a range of reactions from “unsettling” to “cute.” So much for a foolproof formula: Creepiness is in the eye of the beholder.
Speaking of which, what’s that scuttling toward Pittsburghers in the Hundred Acres Manor haunted forest? Why, it’s CH1PP13, the deranged robo-raccoon that’s snicker-snacking its razor-toothed jaws, of course. Wicked Makers built this snarling little nightmare from a Unitree robot dog, but it ran into some unfortunate snags along the way. Like so many humans at the tail end of a Halloween party, CH1PP13 glitched, attempted a handstand and broke into several pieces.
If you have neither the time nor talent to custom-build your own demon raccoon, the ubiquitous Spirit Halloween has you covered with its animatronic Trash Bandit. (And if you want less “jump scare” and more “isn’t it time to throw away those brown bananas, dear?” then try Buzz, this fruit fly on steroids.)
Finally, if you like your scares existential, there’s the Dead Internet Museum at MAD Arts in Florida, an immersive installation chock-full of ghostly digital artifacts and intentionally glitchy interfaces. Or save yourself the trip and wait until the holidays, when you can experience much the same thing trying to fix your parents’ Wi-Fi.
