Recent concerns from tech luminaries about a robot apocalypse may be overblown, but artificial intelligence researchers need to start thinking about security measures as they build ever more intelligent machines, according to a group of AI experts.
The fields of AI and robotics can bring huge potential benefits to the human race, but many AI researchers don't spend a lot of time thinking about the societal implications of super intelligent machines, Ronald Arkin, an associate dean in the Georgia Tech College of Computing, said Tuesday during a debate on the future of AI.
"Not all our colleagues are concerned with safety," Arkin said during the debate, which was hosted by the Information Technology and Innovation Foundation (ITIF) in Washington, DC "You cannot leave this up to the AI researchers. You cannot leave this up to the roboticists. We are an arrogant crew, and we think we know what's best."
While human-like intelligence in machines should still be a long time away, it's not too early to start thinking about policies and regulations to prepare for that future, Arkin and other AI researchers said.
Long-held fears of a robotic takeover of the world, voiced in science fiction stories for decades, have gained new traction in recent months, with tech thinkers including Bill Gates, Stephen Hawking and Elon Musk raising concerns about the dangers of AI.
Meanwhile, recent advances like Apple's Siri, Google's self-driving cars and DeepMind's Deep-Q AI that has mastered dozens of Atari video games make some people believe that human-like machine intelligence is coming soon.
But it's hard to predict when human-like machine intelligence will happen, and it could still be decades away, said Nate Soares, executive director of the Machine Intelligence Research Institute. AI is now capable of "deep learning" involving specific tasks, but researchers need several more breakthroughs before they can design machines that can learn to accomplish a broad range of activities, like humans do, he said.
Super human intelligence from machines will happen "somewhere between five and 150 years, if I was going to be bold" about a prediction, Soares said.
Soares said he falls on "both sides" of the debate about the danger of super intelligence machines. "AI's going to bring lots and lots of benefits and if we do it poorly, it's going to bring lots and lots of risks," he said.
It's important not to overstate the risks, countered Robert Atkinson, ITIF's president. Some policymakers and members of the media will latch onto visions of a robot apocalypse when AI experts express concerns about the downsides of intelligent machines, he said.
Those fears, in turn, could lead to limits on government AI funding and stunt the growth of the technology, Atkinson said. Musk's recent statement suggesting AI is "summoning the demon" is demonising the technology, he said.
Few other technologies generate the same level of fear, he said. "It's very different to say, 'Look, we are a community of responsible scientists who are building safety into this thing, and we're pretty sure it's going to work,'" Atkinson said.
The good news is that humans are still in control over how AI and robots will develop, but a more robust discussion about AI's future is needed, said Stuart Russell, a professor of electrical engineering and computer science at the University of California, Berkeley.
Even though Atkinson suggested that the danger is limited because it's still impossible to design a robot with intentionality, Russell suggested intentions aren't necessary for there to be a risk.
"If the system is better than you at taking into account more information and looking further ahead into the future, and it doesn't have exactly the same goals as you, then you have a problem," Russell said. "The difficulty is that we don't know what the human race's values are, so we don't know how to specify the right goals for a machine so that its behavior is something that we actually like."
In some cases, AI developers might think they're giving the right instructions to an intelligent machine, but the results aren't what they expected, like in the legend of King Midas, Russell said. "What happens when you don't like what they're doing?" he said. "You could say, 'Shut them down,' but a super intelligent system ... knows that one of the biggest risks to it is being shut down, so it's already outthought you."
With many AI researchers working on a small piece of the general-purpose intelligence puzzle, policymakers and scientists should talk about the potential negative implications instead of "keeping our fingers crossed that we'll run out of gas before we run off the cliff," Russell added.
Some people are more optimistic about super intelligent machines coexisting with humans, said Manuela Veloso, a computer science professor at Carnegie Mellon University. Service robots now escort visitors at Carnegie Mellon to Veloso's office and surf the Web to learn new information, she noted.
Robots are reaching a point where they will provide benefits to many people, she said. Research on coexistence will teach intelligent machines "not be taught to be outside of the scope of humankind but to be part of humankind," she said. "We will have humans, dogs, cats and robots."