Last month, in almost real time and in front of a live audience in San Francisco, two humans debated facts and ethics with an AI system, IBM Project Debater.
The AI platform offered points and rebuttals during its first public debate, going first against Noa Ovadia, the 2016 Israeli national debate champion, on the issue of subsidizing space exploration, and then against Israeli professional debater Dan Zafrir on the topic of telemedicine. Debaters, human and machine, were not made aware of the subjects ahead of time.
According to a snap poll of the audience after each session, audiences felt that Project Debater, at least on the topic of space exploration, enriched their knowledge more than its human counterpart.
The humans, however, were largely found to be better, more persuasive speakers. Indeed, despite its digital prowess, some debate experts have noted that the AI debating system lacks a certain ability to deploy tonal effects, such as irony and sarcasm.
The debating system isn't IBM's first foray into machine-human jousting. In 2011, IBM's Watson supercomputer beat trivia stars in a game of Jeopardy, and in 1997, IBM's Deep Blue chess computer bested world chess champion Garry Kasparov.
The Project Debater engineers appear to be taking a more nuanced approach this time.
Ranit Aharonov, manager of the IBM Project Debater team, said members of the project team didn't think of the system simply as something that could win a debate.
"When we look at a debate, we don't only look at who swayed the audience more. There's a lot more to it," Aharonov said.
Building a debating machine
Developed over the past six years at the IBM Research lab in Haifa, Israel, IBM Project Debater uses sophisticated machine learning algorithms and millions of newspapers and articles to identify and organize facts relevant to a debate topic.
The AI is able to cluster that information into themes based on the topic of debate, and, using what IBM calls data-driven speech writing, delivers the information in a coherent sentence.
IBM Project Debater is essentially trained in the art of debate -- to have a general idea of when and how to use factual and ethical arguments to support or dispute a point. To be able to debate in real time, the system employs natural language processing to identify the main components of an opponent's speech and then give a rebuttal.
According to Noam Slonim, principal investigator for IBM Project Debater, the technology could have a number of applications. The most immediate one, he said, is advancing the field of science.
"While pursuing this researching project, during the process, we are actually finding ourselves facing new problems we haven't faced before," Slonim said
Beyond advancing the field, Slonim said he sees "the underlying technology with the Debater being very, very aligned with technologies that help people make a better-informed decision," noting that it could eventually have uses in the fields of politics or business.
"Just imagine giving Debater a topic and asking it to find everything of relevance to that topic and what that could mean," he said.
Also, Slonim said he sees the technology being useful in the education field. Having a debating AI system "can help kids learn how to build better arguments and become more informed in a topic."
"IBM Project Debater, while still in development, could be brought into specific use cases, and is slated to be released in some form next year," Slonim said. He declined to say what incarnation the technology might take.
As for bringing some of the technology behind IBM Project Debater to IBM's well-publicized AI system, Watson, Slonim said: "The implication is that these will be incorporated into Watson and enhance its capabilities."
Mixed review from an analyst
Adrian Bowles, vice president of research and lead analyst for artificial intelligence at Aragon Research, was at last month's live debate in San Francisco.
Bowles, who said he first spoke with an IBM representative about the IBM Project Debater four years ago, said he was struck more by how IBM Project Debater identified arguments than how it expressed them.
"The natural language generational software is not nearly as impressive to me as what they've done with natural language understanding," he said, adding that the AI system presented arguments more on a high school or college level than a professional one.
"Finding and representing the logical position and being able to identify the opposite of that is where the magic happens, if you will," Bowles continued.
Bowles agreed that the technology could be useful when applied in a classroom setting, but noted that he would also like to see it used to help extract provable facts from bodies of text, like multiple sources of news.
Specifically, Bowles cited fake news and the political bias reflected by news sources. Technology in IBM Project Debater could be used to analyze multiple news sources on the same topic and help separate facts from bias or misreporting.
"What I would like to see is it being able to identify practical arguments and being able to map those out," he said.
Due to the vast number of documents Debater has access to, Slonim said he would expect technology to be separated out before the system is commercialized, partitioning it out to get to the basic underlying technology and allowing users to input their own data to be analyzed.
"I think that would get people using it and experimenting in novel ways," he said.