For the past 30 days in the frigid port city of Duluth, Minnesota, a colony of houseflies has been kept alive by a piece of software.
The computer takes care of their needs, giving the insects water and nutrients in the form of powdered milk and sugar. The flies, of course, are unaware that their ultimate fate depends on whether or not a machine correctly identifies blobs of pixels flitting across a camera—if it fails, they die.
“We should be smart about how we plan for artificial intelligence, because one way or another it’s coming,” said David Bowen, the 41-year-old artist and professor behind the installation, which he calls FlyAI. “It could either be a great benefit to humanity, or it could be quite tragic.”
At the core of the installation is a type of machine learning software known as “deep learning,” which essentially trains itself to perform a task: in this case, correctly identifying flies and feeding them. As the flies move across a camera installed in their enclosure, the computer makes its best guess as to what it’s looking at, and if it’s a fly, a pump releases life-giving water and nutrients for them to sip on.
Like most AI today, it’s not a perfect system. Due to limitations in its photo training, the software regularly incorrectly identifies the insects as light switches, or even egg nog, Bowen said. When this happens, the flies go hungry and thirsty. But even so, Bowen insists, the colony is doing fine after a month, and if something went wrong, he’d intervene to rescue his “friends.”
“They die, I suppose, a natural death,” he said. “They’re getting geriatric in there.”
Image: Flickr/Mike Licht
Bowen’s inspiration for the installation was partly philosopher Nick Bostrom’s recent book Superintelligence: Paths, Dangers, Strategies, which ruminates upon the risks posed by a god-like artificially intelligent being.
But Bowen’s own installation shows that we don’t need superintelligent beings to have computers that can take care of a living thing over time. The system is based on Google’s open source machine learning library, TensorFlow, and off-the-shelf image recognition training databases. It all runs on a Raspberry Pi, a computer for tinkerers so simple that it’s almost a toy.
Perhaps the greatest risk to humanity when it comes to machines isn’t an omnipotent AI, but that, like Bowen’s flies, we become oblivious to the critical software working in the background of our lives—and thus, powerless.
“One of the dangers is that we get so far removed from the technology that it’s not even something that we understand anymore,” Bowen said. “That seems like it might be kind of scary.”
Image: David Bowen
Bowen brought up the example of Air France flight 447, which crashed on its way to Paris from Rio De Janeiro in 2009 after an autopilot failure left its pilots unable to fly, killing all 228 people aboard.
But Bowen’s installation also unintentionally presents a bright spot for our future with machines. The flies are rebelling; they’re shitting all over the encasement and screwing up the AI’s camera. They might die, but they’ll be free.
But then again, they’re flies. They don’t know what’s happening.
Get six of our favorite Motherboard stories every day by signing up for our newsletter.