AI debater uses arguments and emotions to challenge human opponent
Originally published on Feb. 15, 2019.
This past winter, World Debating Championships' grand finalist Harish Natarajan went toe-to-toe with IBM's latest artificial intelligence system in a debate — and won.
Held on Feb. 11, it was Project Debater's second public competition since it made its debut in June 2018, when it off against two humans.
"I felt I had a fairly good chance of winning, but … I knew I didn't understand its capabilities," admitted Natarajan, in an interview with Spark host Nora Young.
"And the history of man versus machine has been one where people start with a level of hubris, but in the end lose to machines."
Despite his victory, Natarajan had reasons to be worried.
As per the rules of their debate, both Project Debater and Natarajan were told the debate topic — subsidized preschool education — only 15 minutes before their competition.
Thankfully, Natarajan — who holds a world record for most debate competition victories — was accustomed to the 15-minute rule and prepared as if facing a human opponent.
"I was doing general reading, I was thinking a little bit about the topic, but I realized … you couldn't prepare for the topic itself, because you don't know what it is, and if I try to prepare for an opponent who I thought would be very fact-heavy, I realized I could definitely be surprised by the skills it had," explained Natarajan.
However, Natarajan also had to contend with Project Debater's gargantuan knowledge database.
According to Ranit Aharonov, manager of the Project Debater team, the AI computer — which takes the shape of a large, black monolith — can access a collection of almost 400 million newspaper and journal articles.
That amounts to approximately 10 billion sentences, all stored in the cloud.
"It also has another interesting source of information … and we call that a graph of the model of human dilemma," explained Aharonov.
This includes responses to ideological arguments like, "When is the right time for the government to intervene on the expense of personal freedom?"
The human art of persuasion
As a human versed in the art of debate, Natarajan did have one key area of expertise: the ability to persuade others through emotion and rhetoric.
However, the IBM team anticipated this issue and programmed the machine with Aristotle's rhetorical strategies in mind, including appeals to emotion and logical reasoning.
"A good speaker also integrates humour in her speech, so we try to give Debater some of these qualities as well," explained Noam Slonim, the principal investigator on the Project Debater team.
Project Debater flaunted its rhetorical skills throughout its debate against Natarajan, at one point even stating sarcastically, "I sometimes listen to opponents and wonder, 'What do they want? Would they prefer poor people on their doorsteps begging for money?'"
Natarajan was impressed with the machine's personality and acknowledged that he found the AI's emotional arguments quite interesting.
"Now obviously, an AI machine doesn't really have empathy, doesn't really have emotions, but it started being able to mimic that and try and play off the audience's intuitive emotions really well," conceded Natarajan.
Lessons learned from loss
In the end, Natarajan held his own against Project Debater and won, successfully convincing 17 per cent of the audience to move away from his rival's position.
In comparison, none of the audience that initially agreed with Natarajan changed their opinion at the end of the debate, meaning Project Debater failed to convince anyone to change their positions on subsidized preschool.
That being said, the audience overwhelmingly agreed that they learned more from Project Debater's arguments than from Natarajan.
As for what's next, Slonim explained that Project Debater still needs to work on its ability to generate rebuttals.
"[Natarajan] is capable of understanding very subtle nuances and [responding] to that instantly," said Slonim.
"The system is still not at that level. I believe the rebuttal [capabilities] that were demonstrated … by the system were very impressive and unique for an AI technology, but still inferior compared to Harish."