In 2011, IBM's Watson system squared off against two human champions, Brad Rutter and Ken Jennings, on the game show Jeopardy! It beat them both so handily that for his final response Jennings simply wrote, "I, for one, welcome our new computer overlords." It was an awesome display, unlike anything anyone had seen before.
The implications went far beyond the company or the game show. Watson's triumph kicked off an arms race in artificial intelligence. Later that same year, Apple launched Siri, its personal assistant. In 2015, Google's AlphaGo computer beat a human champion at the famous Asian board game and Amazon launched its Echo smart speaker.
This summer, IBM raised the stakes again with its Project Debater, a system that can compete with skilled humans arguing about controversial topics. Much like Watson, Debater's objective is not to launch a new product, but to expand horizons. While the full ramifications aren't exactly clear yet, what is becoming clear is that we are embarking on a new era of possibility.
A History of Grand Challenges
In the technology industry, IBM is unique for its longevity. While others seem to rise and fall with each new cycle, the giant of Armonk has somehow managed to to stay on the cutting edge for over a century. It was a leader in tabulating machines, then mainframes, then PC's, The Internet and now artificial intelligence and quantum computing.
A key to its success has been its history of grand challenges such as the Deep Blue project which defeated world champion Garry Kasparov at chess and the Blue Gene project which created a new class of "massively parallel" supercomputers and, more recently, Watson and Debater. These are pursued without any immediate business applications in mind, but are meant to stretch the boundaries of technology.
"A successful grand challenge is one that people, even experts in the field, regard as an epiphany and changes assumptions about what's possible," Bernard Meyerson, IBM's Chief Innovation Officer, told me. "The commercial value comes in applying those new possibilities to business problems."
Project Debater is very much in the same vein. Nobody really knows how it will affect IBM's products or its competitive position. Rather, it was a task undertaken to pursue problems that were, until now, considered to be unsolvable. If history is any guide though, it will drive the business forward for years to come.
Going Beyond Games
What makes Project Debater unique is that it attempts to answer questions that have no definitive answers. With today's personal assistants, we can ask questions like "What is the weather going to be today?" or Where is the nearest Starbucks?" but we can't ask them things like, "Should I invest my money in stocks or in bonds?" and expect to get a cogent answer.
"When AI started back in the 50's they used games as a test, first checkers, then backgammon, then chess and eventually Alpha Go." Noam Slonim, a researcher at IBM told me. "It's clear at each stage of the game what the options are and you can approach it like a search problem, which can be solved largely with computational power and clever algorithms."
While making clear that teaching computers to play -- and win -- those games was a major and worthwhile achievement, he stressed that solving the much more enigmatic problems of debate presents new and very different challenges. "Games represent the comfort zone of AI," he says. "With Project Debater we wanted to move out of that comfort zone."
Yet to do that takes more than just a vision. The reason that nobody has taught a machine to debate is not that nobody ever thought of it before or were aware of the potential, but because it presents unique problems that are devilishly hard to solve.
Solving the Unique Problems of Debate
AI systems are generally developed the same way. There is a mountain of data, called a learning corpus, that the system analyzes to solve problems and answer questions. Much like a human, with each attempt, the system learns and gets better at the task it's being trained for. The major difference between machines and humans is that machines can do it much faster.
With a debate, however, the process is not so simple. "You can't just run 100 debates a minute, see how the system does, come up with a quantitative score and make adjustments, since there is no simple, automatic way to determine the debate result," Slonim points out. "You can't win a debate in a way that can't be understood by humans." That makes the training process inherently different.
Another issue is that the system needs to be trained to learn very subtle distinctions. For example, it needs to understand the difference between a definition and an argumentative statement. When somebody says "racism is discrimination against someone on the basis of race," it is a definition. But when someone says, "racism leads to mass incarceration," they are making an argument. It is often difficult, even for humans, to separate between the two.
These are just a few of the problems that the team needed to solve. Nevertheless, as you can see in this video, the system is able to take complex, ambiguous issues and make a clear, cogent argument.
Notice how the system is able, about a minute and a half into the video, to not only argue its own case, but to summarize its opponent's, discuss its significance and explain why the view it is presenting better matches the facts. If it had been given the other position to argue, it would have done the same for the opposite side.
A Machine Partner for Human Decisions
A crucial difference between Project Debater and a human is that a computer system has no emotions. While that may seem to be an advantage, scientists have long established that emotions are critical in decision making. In fact, patients with brain injuries that caused them to lose the ability to emote also lost their ability to make decisions. They can process information, but are unable to weight it to decide what's important and what isn't.
Yet humans don't need a brain injury to disregard the moral outcomes of their decisions. For example, in the Ford Pinto scandal back in the 1970s, the company produced and sold a car that it knew was unsafe because it believed that it was more profitable to sell a defective product than to fix the problem.
One reason why the scandal happened is that there was nobody to provide the other side of the argument. Humans are social creatures, which often leads us to fail to challenge a consensus viewpoint. Project debater raises the possibility that machines can help keep us honest by always providing a partner to argue the other side of the case.
Machines will never be able to make our decisions for us. Only we can decide whether, say, a human life means more to us that quarterly profits. But by showing us the other side of the argument, machines can sharpen our thinking, provide alternatives and, hopefully, free us to pay more attention to our own ability to weigh emotional and moral content.
The future of technology is always more human.