An IBM computer debates humans, and wins, in a new, nuanced competition

We saw computers beat humans at chess in 1997, beat humans at Jeopardy in 2011 and vanquish the world’s best human players of the ancient game of Go in 2017. On Monday, a computer edged out a victory over people in a far more nuanced competition: debate.

IBM created a system called Project Debater that competes in what the company calls computational argumentation knowing a subject, presenting a position, and defending it against opposition. At a press event, IBM pitted the system against two humans with a track record of winning debates.

In one debate, Noa Ovadia overall nudged two people among a few dozen in a human audience toward her perspective that governments shouldn’t subsidize space exploration. But in the second, Project Debater soundly defeated Dan Zafrir, pulling nine audience members toward its stance that we should increase the use of telemedicine.

You’re not going to lose your job yet to Project Debater’s commercial spin-off right away as Big Blue tries to profit from the IBM Research project. Project Debater betrayed its inhuman nature several times over the course of its 20 minutes of off-the-cuff speech. But Debater did demonstrate that artificial intelligence can handle some complexities of human interaction, not just the clear-cut rules and victories of a board game or game show.

“Our life is not black or white. It’s ambiguous, it’s subjective,” said Ranit Aharonov, director of Project Debater. “AI will have to navigate in that territory.”

Project Debater was trained in advance on debating methods, but not the details of the debate itself, which it found out about only moments before the debate started. To formulate its argument, it had at its disposal a collection of 300 million news articles and scholarly papers, previously indexed for quick search results. But it had to find the information, package it persuasively, listen to its opponents’ arguments, and formulate a rebuttal.

“It’s amazing to see this technology pull from 300 million sources and distill it into what sounds like a conversational narrative in debate,” said Clea Conner Chang, chief operating officer of Open to Debate formerly known as Intelligence Squared U.S., an organization that runs its own debates with a similar argue-and-respond style.