People have long feared what super-intelligent machines could mean for society. But can AI also scare us in a more immediate sense, evoking the kind of terror one might feel watching a horror movie?
That's the question three current and former MIT Media Lab scientists set out to answer—just in time for Halloween. They created Nightmare Machine, a website displaying images of people and places that a machine-learning algorithm has completely zombified. The project suggests a truly frightening future of Halloween, one in which artificial intelligence taps into your individual fears—and exploits them.
Using 200,000 images of celebrities, the team generated a data set of fake faces. Then, they borrowed a framework that can learn the artistic style of a given image—in this case, a generic zombie image—and applied it to the generated faces to maximize the fear factor. They used a similar technique on famous landmarks.
Then they launched a site that asks people to rate the faces "Hot or Not?"-style. A week after the project launched, the researchers collected 400,000 individual evaluations of their computer-generated images, and each response helps the algorithm learn what terrifies people. "[We are] hoping to learn more insights about what makes these faces scary," says Manuel Cebrian, who was a founding member of the Scalable Cooperation research group at the MIT Media Lab and now works as a research scientist at CSIRO Data61, a digital innovation group in Australia.
While some people regard the faces as very scary, others don't—but the Nightmare Machine algorithm deems them all equally creepy. "So that reveals that there is extra information in how humans perceive horror that can be exploited to make even scarier faces," says Pinar Yanardag, a postdoc at the MIT Media Lab.
Artificial intelligence could have some intriguing implications for the business of horror. "Maybe in the future, we can generate personalized horror images where we to tailor the generation process to the individual data," Yanardag says.
The three scientists are not the only ones contemplating how AI might make Halloween creepier than ever. One of the most recent episodes of the television show Black Mirror, which appeared the same day as Nightmare Machine, explored these ideas directly.
"Coincidentally, one of their spookiest episodes ever, dubbed 'Playtest,' imagines a technology that uses advanced AI and neural networks to immerse a person in visceral, augmented-reality horror," says Iyad Rahwan, an associate professor in the lab. "Maybe this technology is closer than we think."
The researchers consider the project a fun and silly way to get in the Halloween spirit. But at its foundation is the desire to understand the obstacles to human and machine cooperation. "Psychological perceptions of what makes humans tick and what make machines tick are [an] important barrier for such cooperation to emerge," Rahwan says.
The trio built the Nightmare Machine in just two weeks, but didn't have quite enough time to let people zombify photos themselves. There's always next year.