Artificial intelligence has gotten a bad rap lately. Stephen Hawking, Bill Gates and Tesla Motors CEO Elon Musk are just a few notable people who have warned about the perilous consequences associated with developing AI systems.
"The development of full artificial intelligence could spell the end of the human race,” Hawking cautioned last year. Musk has even donated millions of dollars to the Future of Life Institute (FLI) to fund a program with the goal of making sure humans manage AI so it doesn’t destroy us. But artificial intelligence in itself isn’t really dangerous, Tom Dietterich, president of the Association for the Advancement of Artificial Intelligence, said at DARPA’s "Wait, What?" conference on Thursday. Rather, the real threat stems from making AI systems completely autonomous, he said.
AI is basically smart software that enables machines to mimic human behavior. For many people, it is already a part of daily life. Apple's Siri, Google Now, and Skype's Real-Time Translation tool are all examples of artificial intelligence. Some AI systems incorporate many different components like computer vision, speech recognition, tactile feedback and touch systems. All of these sensory modalities give computers the ability to sense as well as, or even better than humans. The collected data can then be used to plan or take action.
For example, the autopilot systems used on commercial aircrafts are AI systems that help safely fly the plane. But when people like Musk or Hawking warn about AI, they are cautioning against giving AI systems complete autonomy — which isn’t something that happens naturally, Dietterich said. “A misconception I think people have is that somehow these systems will develop free will on their own and become autonomous even though we didn’t design them with that in mind... AI systems will not become spontaneously autonomous, they will need to be designed that way,” Dietterich said. “So I think the dangers of AI is not so much in artificial intelligence itself, in its ability to reason and learn, but in the autonomy. What should we give computers control over?” In other words, computers won't just take over the world someday, unless we design them to.
The Future of Life Institute proposes that for safety critical AI systems, like in weapons and vehicles, it may be worth considering some form of human control. Dietterich goes even further and said he believes we should never build fully autonomous AI systems. Google is one company investing in AI technology to develop its self-driving cars. “By definition a fully autonomous system is one that we have no control over,” Dietterich said. “And I don't think we ever want to be in that situation.” But given that many companies are already investing heavily in this technology, it seems inevitable that these systems will come into existence.
The potential applications of autonomous AI are already being proposed and many of them involve high-stakes decision making. Think self-driving cars, autonomous weapons, AI hedge funds, and automated surgical assistants. The risks associated with these AI systems must be addressed before it's safe to completely implement them, Dietterich said. As Elon Musk said in August during an AI panel in Silicon Valley “It's definitely going to happen. So if it’s going to happen, what’s the best way for it to happen?”