New study finds it’s harder to turn off a robot when it’s begging for its life

Robots designed to interact socially with humans are slowly becoming more and more common. They’re appearing as receptionists, tour guides, security guards, and porters. But how good are we at treating these robots as robots? A growing body of evidence suggests not good at all. Studies have repeatedly shown we’re extremely susceptible to social cues coming from machines, and a recent experiment by German researchers demonstrates that people will even refuse to turn a robot off — if it begs for its life.

In the study, published in the open access journal PLOS One, 89 volunteers were recruited to complete a pair of tasks with the help of Nao, a small humanoid robot. The participants were told that the tasks (which involved answering a series of either / or questions, like “Do you prefer pasta or pizza?”; and organizing a weekly schedule) were to improve Nao’s learning algorithms. But this was just a cover story, and the real test came after these tasks were completed, and scientists asked participants to turn off the robot.

A photo of the experiment’s setup. Participants had to complete a series of tasks with the Nao robot before being asked to turn the machine off.
Credit: Aike Horstmann et al

In roughly half of experiments, the robot protested, telling participants it was afraid of the dark and even begging: “No! Please do not switch me off!” When this happened, the human volunteers were likely to refuse to turn the bot off. Of the 43 volunteers who heard Nao’s pleas, 13 refused. And the remaining 30 took, on average, twice as long to comply compared to those who did not not hear the desperate cries at all. (Just imagine that scene from The Good Place for reference.)

When quizzed about their actions, participants who refused to turn the robot off gave a number of reasons for doing so. Some said they were surprised by the pleas; others, that they were scared they were doing something wrong. But the most common response was simply that the robot said it didn’t want to be switched off, so who were they to disagree?

As the study’s authors write: “Triggered by the objection, people tend to treat the robot rather as a real person than just a machine by following or at least considering to follow its request to stay switched on.”

A selection of reasons participants gave for not turning off the robot in the study.
Credit: Aike Horstmann et al

This finding, they say, builds on a larger theory known as “the media equation.” This was first established in a 1996 book of the same name by two psychologists: Byron Reeves and Clifford Nass. Reeves and Nass theorized that humans tend to treat non-human media (which includes TV, film, computers, and robots) as if they are human. We talk to machines, reason with our radios, and console our computers, said Reeves and Nass.

Various studies since have shown how this principle affects our behavior, especially when it comes to interactions with robots. We’re more likely to enjoy interacting with a bot that we perceive as having the same personality type as us, for example, and we’ll happily associate machines with gender stereotypes. We observe what’s known as the “rule of reciprocity” when interacting with robots (meaning we tend to be nice to them when they’re nice to us) and will even take orders from one if it’s presented as an authority figure.

“Now and in future,” wrote a group of scholars on the topic in 2006, “there will be more similarities between human-human and human-machine interactions than differences.”

And this isn’t the first time we’ve tested the “begging computer does not want to die” scenario. Similar research was carried out in 2007, with a robot resembling a cat that also pleaded for its life. Participants were forced to turn it off by observing scientists and all of them did — but not before going through a serious moral struggle.

In a video clip of the experiment, you can see the robot asking a volunteer: “You’re not really going to switch me off, are you?” The human says: “Yes I will!” — while failing to do so.

The new study, which was published July 31st, builds on this earlier work by using a greater number of participants. It also tested whether it made a difference if the robot was shown to have social skills before it asked not to be turned off. In some of the trials, Nao expressed opinions to the human volunteers, told jokes, and shared personal information. Surprisingly, this social behavior did not have a huge effect on whether the volunteers “spared” Nao.

So what does all this mean for our machine-filled future? Are we destined to be manipulated by socially sophisticated bots that know how to push our buttons? It’s certainly something to be aware of, says Aike Horstmann, a PhD student at the University of Duisburg-Essen who led the new study. But, she says, it’s not a huge threat.

“I hear this worry a lot,” Horstmann tells The Verge. “But I think it’s just something we have to get used to. The media equation theory suggests we react to [robots] socially because for hundreds of thousands of years, we were the only social beings on the planet. Now we’re not, and we have to adapt to it. It’s an unconscious reaction, but it can change.”

In other words: get used to turning off machines, even if they don’t appear to like it. They’re silicon and electricity, not flesh and blood.

This brain-controlled prosthetic will lend you a hand — and a whole arm

For years, scientists have been exploring how we can use signals from the brain to control prosthetic limbs. Usually, this work is focused on restoring motor function to people who have lost an arm or a leg, but new research from Japan shows how the same technology can also be used to augment existing human capabilities.

Engineers from Kyoto’s Advanced Telecommunications Research Institute have demonstrated how people can be taught to control a third robotic arm with their brains, even using the limb to multitask. As described in a paper published in the journal Science Robotics today, eight of 15 test subjects were able to successfully balance a ball on a board with their hands, while grabbing a water bottle with a brain-controlled robot arm.

Although this may sound like something out of science fiction, it’s important to stress that the functionality of this third arm is extremely basic. The prosthetic moved along a predetermined path and performed only a single gesture: closing and opening its hand. Similarly, the brain-machine interface used to control the arm is not some magical mind-reading device. It’s a cap fitted with electrodes that measure electrical signals produced by the brain. In this case, participants were asked to imagine opening and closing the robot hand. The scientists recorded this signal, and turned it into an instruction for the robot arm.

“What we’re measuring is leakage from the brain’s electrical activity,” Shuichi Nishio, one of the researchers involved in the study, tells The Verge. “We have to tune [the brain-machine interface] for each participant; selecting the right electrodes and frequencies.”

Even with these limitations, though, it is very interesting work. As Nishio and his colleague Christian Peñaloza point out in their paper, it seems to be the first time supernumerary limbs have been controlled using the human brain. Usually such prosthetics are operated using joysticks or, if connected directly to the human body, electrical signals from muscles.

Footage from the tests. A full video can be seen here.
Credit: ATR

Past research in this area gives some clue as to how extra limbs might be used if they became common in the future. This concept device from MIT, for example, gives users two extra robotic “arms” that are worn as a backpack. The creators thought these arms could be useful in manufacturing; helping to hold tools and parts, or bracing a worker’s body if they’re squatting in one place for an extended period of time. If such robot arms could also be controlled by the brain, they could become a seamless extension of the body.

For Nishio and Peñaloza, supernumerary limbs might augment more than just our physical capability — they might improve our brains, too. In their paper, the pair noted that during the multitasking challenge, participants’ performance split into two distinct groups: those who were good and those were bad, with hardly anyone in the middle. This was in contrast to performance when individuals were asked just to control the robot limb by itself. In this single-task challenge, all participants did about the same, performing pretty well.

This difference, say Nishio and Peñaloza, suggests some people are just better at multitasking than others. And if that’s the case, then maybe multitasking is a skill that could be improved by using these sorts of brain-controlled devices. “By operating this brain-machine interface, we have an idea that we may be able to train the brain itself,” says Nishio.

This has yet to be proven, but the pair hope to work on the idea in future tests. After all, if you can learn to control three arms, then surely controlling two is child’s play.