Zombies and aliens may not be a realistic threat to our species. But there’s one stock movie villain we can’t be so sanguine about: sentient robots. If anything, their arrival is probably just a matter of time. But what will a world of conscious machines be like? Will there be a place in it for us?
Artificial intelligence research has been going through a recent revolution. AI systems can now outperform humans at playing chess and Go, recognizing faces, and driving safely. Even so, most researchers say truly conscious machines — ones that don’t just run programs but have feelings and are self-aware — are decades away. First, the reasoning goes, researchers have to build a generalized intelligence, a single machine with the above talents and the capacity to learn more. Only then will AI reach the level of sophistication needed for consciousness.
But some think it won’t take nearly that long.
“People expect that self-awareness is going to be this end game of artificial intelligence when really there are no scientific pursuits where you start at the end,” says Justin Hart, a computer scientist at the University of Texas. He and other researchers are already building machines with rudimentary minds. One robot wriggles like a newborn baby to understand its body. Another robot babbles about what it sees and cries when you hit it. Another sets off to explore its world on its own.
No one claims that robots have a rich inner experience — that they have pride in floors they've vacuumed or delight in the taste of 120-volt current. But robots can now exhibit some similar qualities to the human mind, including empathy, adaptability, and gumption.
Beyond it just being cool to create robots, researchers design these cybernetic creatures because they’re trying to fix flaws in machine-learning systems. Though these systems may be powerful, they are opaque. They work by relating input to output, like a test where you match items in column ‘A’ with items in column ‘B’. The AI systems basically memorize these associations. There’s no deeper logic behind the answers they give. And that’s a problem.
Humans can also be hard to read. We spend an inordinate amount of time analyzing ourselves and others, and arguably, that’s the main role of our conscious minds. If machines had minds, they might not be so inscrutable. We could simply ask them why they did what they did.
“If we could capture some of the structure of consciousness, it’s a good bet that we’d be producing some interesting capacity,” says Selmer Bringsjord, an AI researcher at the Rensselaer Polytechnic Institute in Troy, N.Y. Although science fiction may have us worried about sentient robots, it’s really the mindless robots we need to be cautious of. Conscious machines may actually be our allies.
Robot, Know Thyself
Self-driving cars have some of the most advanced AI systems today. They decide where to steer and when to brake by taking constant radar and laser readings and feeding them into algorithms. But much of driving is anticipating other drivers’ maneuvers and responding defensively — functions that are associated with consciousness.
“Self-driving cars will have to read the minds of what other self-driving cars want to do,” says Paul Verschure, a neuroscientist at Universitat Pompeu Fabra in Barcelona.
As a demonstration of how that might look, Hod Lipson, an engineering professor at Columbia University and co-author of a recent book on self-driving cars, and Kyung-Joong Kim at Sejong University in Seoul, South Korea built the robotic equivalent of a crazy driver. The small round robot (about the size of a hockey puck) moves on a loopy path according to its own logic. Then a second robot is set with the goal of intercepting the first robot no matter where the first one started, so it couldn’t record a fixed route; it had to divine the moving robot’s logic.
“People expect that self-awareness is going to be this end game of AI when really there are no scientific pursuits where you start at the end.”
Using a procedure that mimicked Darwinian evolution, Lipson and Kim crafted an interception strategy. “It had basically developed a duplicate of the brain of the actor — not perfect, but good enough that it could anticipate what it’s going to do,” Lipson says.
Lipson’s team also built a robot that can develop an understanding of its body. The four-legged spidery machine is about the size of a large tarantula. When switched on, its internal computer has no prior information about itself. “It doesn’t know how its motors are arranged, what its body plan is,” Lipson says
But it has the capacity to learn. It makes all the actions it is capable of to see what happens: how, for example, turning on a motor bends a leg joint. “Very much like a baby, it babbles,” Lipson says. “It moves its motors in a random way.”
After four days of flailing, it realizes it has four legs and figures out how to coordinate and move them so it can slither across the floor. When Lipson unplugs one of the motors, the robot realizes it now has only three legs and that its actions no longer produce the intended effects.
“I would argue this robot is self-aware in a very primitive way,” Lipson says.
Another humanlike capability that researchers would like to build into AI is initiative. Machines excel at playing the game Go because humans directed the machines to solve it. They can’t define problems on their own, and defining problems is usually the hard part.
In a forthcoming paper for the journal "Trends in Cognitive Science," Ryota Kanai, a neuroscientist and founder of a Tokyo-based startup Araya discusses how to give machines intrinsic motivation. In a demonstration, he and his colleagues simulated agents driving a car in a virtual landscape that includes a hill too steep for the car to climb unless it gets a running start. If told to climb the hill, the agents figure out how to do so. Until they receive this command, the car sits idle.
Then Kanai’s team endowed these virtual agents with curiosity. They surveyed the landscape, identified the hill as a problem, and figured out how to climb it even without instruction.
“We did not give a goal to the agent,” Kanai says. “The agent just explores the environment to learn what kind of situation it is in by making predictions about the consequence of its own action.”
The trick is to give robots enough intrinsic motivation to make them better problem solvers, and not so much that they quit and walk out of the lab. Machines can prove as stubborn as humans. Joscha Bach, an AI researcher at Harvard, put virtual robots into a “Minecraft”-like world filled with tasty but poisonous mushrooms. He expected them to learn to avoid them. Instead, they stuffed their mouths.
“They discounted future experiences in the same way as people did, so they didn’t care,” Bach says. “These mushrooms were so nice to eat.” He had to instill an innate aversion into the bots. In a sense, they had to be taught values, not just goals.
In addition to self-awareness and self-motivation, a key function of consciousness is the capacity to focus your attention. Selective attention has been an important area in AI research lately, not least by Google DeepMind, which developed the Go-playing computer.
“Consciousness is an attention filter,” says Stanley Franklin, a computer science professor at the University of Memphis. In a paper published last year in the journal "Biologically Inspired Cognitive Architectures," Franklin and his colleagues reviewed their progress in building an AI system called LIDA that decides what to concentrate on through a competitive process, as suggested by neuroscientist Bernard Baars in the 1980s. The processes watch for interesting stimuli — loud, bright, exotic — and then vie for dominance. The one that prevails determines where the mental spotlight falls and informs a wide range of brain function, including deliberation and movement. The cycle of perception, attention, and action repeats five to 10 times a second.
The first version of LIDA was a job-matching server for the U.S. Navy. It read emails and focused on pertinent information while juggling each job hunter's interests, the availability of jobs, and the requirements of government bureaucracy.
Since then, Franklin’s team has used the system to model animals’ minds, especially behavioral quirks that result from focusing on one thing at a time. For example, LIDA is just as prone as humans are to a curious psychological phenomenon known as “attentional blink.” When something catches your attention, you become oblivious to anything else for about half a second. This cognitive blind spot depends on many factors and LIDA shows humanlike responses to these same factors.
Pentti Haikonen, a Finnish AI researcher, has built a robot named XCR-1 on similar principles. Whereas other researchers make modest claims — create some quality of consciousness — Haikonen argues that his creation is capable of genuine subjective experience and basic emotions.
The system learns to make associations much like the neurons in our brains do. If Haikonen shows the robot a green ball and speaks the word “green,” the vision and auditory modules respond and become linked. If Haikonen says “green” again, the auditory module will respond and, through the link, so will the vision module. The robot will proceed as if it heard the word and saw the color, even if it's staring into an empty void.
Conversely, if the robot sees green, the auditory module will respond, even if the word wasn’t uttered. In short, the robot develops a kind of synesthesia.
“If we see a ball, we may say so to ourselves, and at that moment our perception is rather similar to the case when we actually hear that word,” Haikonen says. “The situation in the XCR-1 is the same.”
Things get interesting when the modules clash — if, for example, the vision module sees green while the auditory module hears “blue.” If the auditory module prevails, the system as a whole turns its attention to the word it hears while ignoring the color it sees. The robot has a simple stream of consciousness consisting of the perceptions that dominate it moment by moment: “green,” “ball,” “blue,” and so on. When Haikonen wires the auditory module to a speech engine, the robot will keep a running monolog about everything it sees and feels.
Haikonen also gives vibration a special significance as “pain,” which preempts other sensory inputs and consumes the robot’s attention. In one demonstration, Haikonen taps the robot and it blurts, “Me hurt.”
“Some people get emotionally disturbed by this, for some reason,” Haikonen says. (He and others are unsentimental about the creations. “I’m never like, ‘Poor robot,’” Verschure says.)
A New Species
Building on these early efforts, researchers will develop more lifelike machines. We could see a continuum of conscious systems, just as there is in nature, from amoebas to dogs to chimps to humans and beyond. The gradual progress of this technology is good because it gives us time adjust to the idea that, one day, we won’t be the only advanced beings on the planet.
For a long while, our artificial companions will be vulnerable — more pet than threat. How we treat them will hinge on whether we recognize them as conscious and as capable of suffering.
“The reason that we value non-human animals, to the extent that people do, is that we see, based on our own consciousness, the light of consciousness within them as well,” says Susan Schneider, a philosopher at the University of Connecticut who studies the implications of AI. In fact, she thinks we will deliberately hold back from building conscious machines to avoid the moral dilemmas it poses.
“If you’re building conscious systems and having them work for us, that would be akin to slavery,” Schneider says. By the same token, if we don’t give advanced robots the gift of sentience, it worsens the threat they may eventually pose to humanity because they will see no particular reason to identify with us and value us.
Judging by what we’ve seen so far, conscious machines will inherit our human vulnerabilities. If robots have to anticipate what other robots do, they will treat one another as creatures with agency. Like us, they may start attributing agency to inanimate objects: stuffed animals, carved statues, the wind.
Last year, social psychologists Kurt Gray of the University of North Carolina and the late Daniel Wegner suggested in their book “The Mind Club” that this instinct was the origin of religion. “I would like to see a movie where the robots develop a religion because we have engineered them to have an intentionality prior so that they can be social,” Verschure says. ”But their intentionality prior runs away.”
These machines will vastly exceed our problem-solving ability, but not everything is a solvable problem. The only response they could have to conscious experience is to revel in it, and with their expanded ranges of sensory perception, they will see things people wouldn’t believe.
“I don’t think a future robotic species is going to be heartless and cold, as we sometimes imagine robots to be,” Lipson says. “They’ll probably have music and poetry that we’ll never understand.”