Blog: Robotic Debator

Robotic Debator
Robotic Debator
Date: 2019-Mar-03 08:14:24 EST

A few thoughts on IBM Project Debator's recent IQ2 Debate on subsidising preschool, available on youtube here.

First, on the debators:

  • Harish is a masterful debator who can make sharp points that echo throughout the debate, and it'd take careful work for a skilled opponent to dismantle them and establish countervailing points that would do the same. I would love to see him in action some more
  • Project Debator (which I'll call PD for the rest of this post, and because I think of gender in terms of genetics, I will refer to as an it) was able to put together a reasonable argument on its own (and it even went for emotional pleas, if not particularly well). It failed utterly in offering adequate responses to Harish's points because, like Eliza the psychotherapist, it didn't actually understand them. Or at least, that's my impression. It provided a lot of structure to its arguments, but overcommunicated that structure, wasting valuable time
I'm a tough judge of arguments - I've done debates and thought about philosophy a lot over the years. I'm probably tougher on PD than I need to be, as I suspect a fair number of people would, if those lines were said by a human debator without any "AI project" background, would find them reasonable if a bit cluttered. Harish opened strong, but I think he backed off a bit when he saw that PD lacks enough understanding to offer a truly competitive (or intelligent) response. The opening statements were, in my view, the only truly solid part of the debate, and Harish's key argument - that subsidies are not useful to shape behaviour if they only go to people who would already do the intended action, and that any reasonable subsidies would not likely be sufficient to get the poor to send their children to preschool - never was countered (or even fully addressed). There are at least two ways to make a response (suggest subsidies sufficient to get poor families to do it, or suggest that lowering the burden for families in their childbearing years at the cost of everyone in society paying over all their years is worthwhile) that a human debator could have offered.

There was an interesting post part to this debate where they talked about what PD can do now. It can already pull large datasets together into something that looks like an argument (or possibly just find existing premade statements with the right content and press them together with some summarisable structure). That's potentially useful - if I wanted a perspective on a matter that differs from my own, it'd be nice to have something that can try to make an argument. A decent editor could shape PD's first draft into something coherent (even though right now it'd likely be of mediocre quality). If this were a feature on my phone or Google Home devices I'd probably use it a lot (I often just ask for random facts of my Home devices). I look forward to seeing this technology develop - I am not convinced that it has much deep understanding of what it's doing, but even Eliza sees a certain amount of use and it's a program from the 1960s. I also wonder about using things like this, with human proofreading, for content generation for a lot of areas (games, storytelling) where humans are the most expensive part.