IBM AI Project Debater Successfully Debates Humans

During a small event in San Francisco, Project Debater AI successfully participated in two debates hosted by IBM

The goal that was set for the AI was to be able to engage in a series of reasoned arguments, respecting at the same time the standard rules of debate: no information regarding the debate topic prior to the event, no pre-canned responses. Each side shared its introductory 4-minute speech, a 4-minute rebuttal to the other’s side arguments, and a 2-minute closing statement.

AI Project Debater was able to hold its own, just like the two human participants did.

This represents a big step forward if we think about the previous demonstration that IBM held, when Watson crashed its competition at Jeopardy. Project Debater, just like

Watson, was able to analyze the contents of all the available data and respond with relevant answers; however, this time, the answer was related to the four minute speech that was previously held.

Surprisingly, Project Debater was able to cite sources, pander to the audience’s love for children and veterans and even crack a joke or two in the process.

This is quite impressing, considering that it successfully put together information that it received only a few minutes before, with no prior preparation. However, the system has millions of articles stored, assumed to be accurate, in almost 100 areas of knowledge. As soon as it receives a debate topic, it takes a couple of minutes to go through them, decide what would make for the best argument and then create a speech enforcing those points.

On the other hand, the AI’s job was to respond directly to the point, whereas it didn’t quite rebut them directly. Even though it talked in the same area, it did not directly engage in the criteria set forth by the human opponent. The question is: did Project Debater not understand the criteria, or it simply chose not to engage on those terms?

Tell us what you think in the comments section below.