In this Costa Rican forest, monkeys come face to face with AI

One day last summer, a curious white-faced capuchin came across a strange thing in the woods. There, in the middle of the Taboga Forest Reserve in Costa Rica, sat a 15-inch touchscreen mounted in a wooden frame. As the monkey, an alpha male named Papi, began to investigate—poking the device here, poking it there—his fingers landed on the screen. Suddenly, a piece of dried banana fell into the tray under the frame.

It didn’t take long for Papi to learn the basic rules of the device: touch the screen and get a dried banana slice. He also provided a proof of concept for CapuchinAI, a new device designed to assess the cognitive abilities of monkeys in the wild. The test device the researchers described in the new paper used artificial intelligence facial recognition software to detect capuchins in real time and record their responses to a simple learning task.

The researchers hope that more sophisticated versions of the device, which they will begin testing in the coming weeks, will shed new light on the evolution and intelligence of primates and answer questions that could not be studied in the laboratory, such as how a monkey’s cleverness affects its success and survival.

“If we really want to understand how primates make decisions, if we want to understand how they use these big brains that they’ve developed, we really have to put it in the context of the world they’re moving around in.” said Marcela Benítez, a primatologist at Emory University and author of a new paper published Tuesday in the American Journal of Primatology. “But that’s a lot easier said than done.

Dr. Benítez has been studying white-faced capuchins in Taboz for years, recording their behavior, recording their vocalizations and measuring their hormone levels. For the new study, she and her colleagues used images of some of these monkeys to train an artificial intelligence model to identify capuchins and distinguish them from other animals moving through the forest. They then built a portable test station equipped with a touch screen, a web camera and a 3D printed food dispenser.

The station was designed to detect an approaching capuchin and present it with a simple associative learning test: Could the animal figure out that touching the screen meant getting a reward?

Last summer, scientists tested the device, setting it up in the forest and hoping the monkeys would engage. For the first two days, not a single capuchin appeared.

“And so, you know, in my typical first-time research, PhD fashion, I was there biting my nails,” said Federico Sánchez Vargas, a doctoral student at Emory University and author of the new paper. “Oh no, I’ll have to start from scratch, do something else.

But on the third day, Papi appeared. Eventually, other monkeys did as well. And over several test sessions, 16 capuchins explored or interacted with the device.

Not everyone understood the task. One monkey “would bite the area where the banana would come out,” Mr. Sánchez Vargas said. “He climbed on the box and slapped the box, but he just didn’t think to touch the screen.

But 10 of the 16 monkeys learned how to touch the screen to get a reward, and at least eight of them seemed to retain the knowledge.

This summer, researchers will pilot test stations with a more sophisticated facial recognition model that is trained to identify and distinguish individual monkeys. An earlier version of this capuchin identification model, trained on a smaller group of capuchins, was able to identify individual monkeys with more than 97 percent accuracy.

The machines will be loaded with a library of tasks designed to assess different cognitive skills. Tasks include impulse control tests, in which monkeys are rewarded for touching green squares and ignoring red ones, and cognitive flexibility, in which the rules of the green square, red square task are reversed. The test stations are designed to track each monkey’s progress and pick up where they left off in the previous session.

Researchers are particularly interested in examining individual variation in cognitive skills—and both the causes and consequences of that variation. Does experiencing early-life adversity such as extreme drought affect cognitive development? Does good short-term memory help monkeys find food?

“Connecting knowledge with real-world challenges, results, and ultimately fitness—that would be the ultimate end goal,” said Dr. Benítez.

Still, the researchers don’t plan for touch screens to become permanent features of the landscape. “I don’t want it to be like walking into a casino when you walk into the woods,” said Dr. Benítez.

And for now, she added, the stations are not intended to be used without human supervision. “I’m not at the point where I want it to be completely remote,” she said. “But maybe ask me in 10 years.