A special event held by IBM in June of 2018 intended to show off their latest breakthrough AI program. They called it Project Debater. The name pretty much says it all. They claimed that the machine would be able to out debate even the best humans. During this event Project Debater went up against the 2016 Israeli debate champion Noa Ovadia. The topic was the statement “We should subsidize space exploration”.
Each side was to deliver a four minute opening statement, a four minute rebuttal, and a final two minute summary. The AI focused mainly on the benefits of space exploration as a whole while Noa Ovadia claimed that there were better uses for government subsidies for research here on Earth. The AI was obviously provided with training data that comprised of certain statistics to prepare for this similar to how humans prepare but its reasoning was unscripted. Afterwards it was clear from the exit polls that the AI won. How did the AI accomplish this and how does it work?
HOW DOES THIS PROGRAM WORK?
Like with many new AI programs this one delivers such amazing results that it may seem more capable than it actually is. It is true that its reasoning is unscripted however, it has access to a huge database comprising of hundreds of millions of fact based articles pertaining to a certain topic. This AI is designed to quickly construct an argument despite having no prior knowledge of it. This is where the massive database serves it well. Everything is also saved locally so no internet access is required.
DOES IT MERELY REGURGITATE THE FACTS?
A lot of people wonder whether Project Debater can actually form real arguments or can it just regurgitate facts from its database. It does rely heavily on facts but it also has certain qualities built into it to ensure the debate is entertaining for a human audience. For example the AI has been programmed to incorporate jokes here and there which is more effective than most believe. Also, when it has a chance at rebuttal it focuses on portraying its human opponent as either a liar or as someone who is misrepresenting reality.
The program used this tactic to a great degree when debating Noa Ovadia. It sidestepped her arguments about how space exploration was a waste of government subsidies by making it clear that the benefits outweighed any cost and far outweighed any research that could be carried out here on Earth.
Project Debator going up against Dan Zafrir another debate champion.
HOW PROJECT DEBATER CATEGORIZES INFORMATION
This system effectively trawls through its information database to find small snippets of arguments related to the debate question. It then searches for any evidence to back up each claim and if there is none it assigns its value 0 and if it does it assigns it a value of 1. Those statements marked with a 0 will generally be discarded during the times in which the program tries to make a fact based argument. But, those statements aren’t always completely discarded because we all know that winning a debate comes down to much more than just facts. The goal of this program is to convince people of its argument which only requires facts to a certain degree.
THIS AI CAN PLAY BOTH SIDES
The creators of this new program designed it to be the most unbiased debater in existence. Its ability to categorize statements allows it to draw clear lines between information that supports one side and information that supports the other side. The methods used to code this AI make it very objective and unbiased. IBM has done a good job in this regard. However, the fact that it can play both sides evenly suggests that it is utilizing certain tactics to win in the absence of a solid argument which means it can be overly persuasive. This is something to watch out for in the future when more of these programs are created. They could be used for immoral purposes and mass influence.
CLAIMS OF PLAGIARIZING
Critics of this program claim that this program is plagiarizing. It trawls for statements related to the question and then takes categorized statements and uses them to form arguments. This suggests that the AI is copying phrases and statements it gathers without actually doing any of its own reasoning. This might be true to an extent since the AI does not possess general intelligence and cannot reason at the level of a human. However, it still has to perform the categorization process all on its own and this requires a very powerful program. All in all the AI is doing a lot more than simply copying statements and therefore these claims of plagiarizing don’t hold much weight.
THE FUTURE OF DEBATE
It remains unclear whether the presence of debate AI will change how we attempt to sway people’s opinions in our favor but what is clear is that the research behind it shows no signs of slowing down. IBM has already perfected text-to-text speech programs like Watson and there is huge potential use for them in the future.