The US government may be concerned about China getting its hands on its artificial intelligence technologies, but Beijing may be thinking twice about trying to acquire them after reading about how easily a few creative US Marines defeated one.
The episode in question appears in a forthcoming book by Paul Scharre, a former Pentagon expert in unmanned and autonomous systems and emerging weapons technologies, titled “Four Battlegrounds: Power in the Age of Artificial Intelligence,” and was recounted in US media.
The Defense Advanced Research Projects Agency (DARPA), the Pentagon’s private bureau of mad scientists, was testing out a new AI developed as part of its Squad X program. The program aimed to integrate unmanned assets with dismounted ground troops by providing them with many of the resources that vehicle-based units enjoy, such as multi-directional threat detection and both increased autonomy and coordination.
In plainer terms, it included a camera the AI used to identify human beings moving toward it.
Unfortunately for DARPA, the US Marines working with the program developed a cheap and easy way to defeat the AI. Fortunately for readers, it was really hilarious.
The Marines were called in to drill around the camera-equipped robot for a week so it could learn what it was like to see a squad of Marines moving around it. At the end, the Marines would switch sides and try to outsmart the computer.
“If any Marines could get all the way in and touch this robot without being detected, they would win. I wanted to see, game on, what would happen,” Phil Root, the deputy director of the Defense Sciences Office at DARPA, told Scharre in the book.
According to Scharre’s book, the AI struggled with how a human body changes shapes as it moves, as well as spotting a camouflaged human body. The Marines quickly figured this out and developed some novel ways of defeating the AI.
Some of them approached in ways disruptive to their body profiles, such as somersaulting for 300 meters to approach the robot undetected. Two others hid under a cardboard box and snuck up.
“You could hear them giggling the whole time,” said Root in the book.
Another Marine held the branches of a pine tree in front of him as he approached, and the AI couldn’t figure out there was a human behind them.
“Eight Marines – not a single one got detected,” Root noted.
According to Scharre, the problem is that computers are really good at doing what they’re taught to do, but really bad at figuring out something they haven’t been taught to do, even if it would be obvious to humans.
“Humans tend to have a much richer understanding of the world,” he said.
So rest easy for now, the Terminator revolution isn’t coming anytime soon – but if it is does, know that we can apparently beat the T-1000 by dressing up as trees and cardboard boxes.