What Will IBM’s AI Debater Learn from Its Loss?
The utility of IBM’s latest man-versus-machine gambit is debatable. At the very least its Project Debater got us thinking about the potential uses of artificial intelligence as a way of helping humans sift through all the noise, to make sense of information, misinformation and, increasingly via the hijacking of social media, disinformation.
There have been other (fictional) debates between humans and machines. Among the most memorable was the argument between astronaut Dave Bowman and HAL, the brains of the spacecraft Discovery One in the 1968 film 2001: A Space Odyssey. Dave lost the debate over opening the spacecraft doors, but outwitted HAL by coming in through the emergency airlock and then disconnecting HAL’s “higher functions.”
IBM’s debater is of course more benign. And one could make the case that a set of algorithms combined with HAL-like speech synthesis could help humans sharpen their thinking through debate. Indeed, part of IBM’s pitch in promoting an AI-based debater is that it promises to rescue us from “superficial thinking.”
IBM’s AI debating system dubbed “Miss Debater” took on Europe’s champion debater, Harish Natarajan of the University of Cambridge. They sparred over whether preschools should be subsidized. Miss Debater drew on nearly a decade of research by IBM’s Watson team, resulting in a machine that reportedly scanned more than 300 million newspaper articles and scientific journals to identify arguments relevant to specific topics.
Ultimately, Natarajan’s command of the facts and the way he was able to articulate them swayed the crowd in his favor during IBM’s Think conference in San Francisco.
Presumably, IBM’s Miss Debater “learned” much from the exchange with a champion human arguer. That was one of the takeaways from Noam Slonim, the IBM researcher heading Project Debater. He was quoted after the event as acknowledging that debate algorithms still must learn how to connect with and convince humans. It’s all in the delivery, Slonim told Fortune magazine, and “this is human territory.”
That takeaway underscores other AI realities, namely, that narrow and general AI remain “brittle,” as in lacking in the common sense most of us develop early in life. Hence, the Defense Advanced Research Projects Agency is funding investigations into machine “common sense” reasoning. Making sense of the world is currently beyond the capabilities of current AI constructs, and DARPA is hoping to find a middle ground between narrow and general AI that delivers adaptable AI technologies. The effort addresses AI’s current lack of semantic understanding while addressing the mystery of how machines learn.
If it’s true we learn through experience, perhaps Miss Debater’s loss and the insights gained in its encounter with a world-class debater will shed light on these and other AI mysteries. That would go a long way toward using machines to augment human activities.
Few would argue that point.