Can AIs Ever Play Dungeons&Dragons?

I slew a dragon, saved a town, and built a castle, all the while having a barbecue together with my friends, in real-time. With Dungeons and Dragons, you can live in two worlds at once, one of which amounts to the existence of your characters in a role-playing game.

Even so, with the right friends and the right imagination, all of it can feel like the perfect escape. I have been playing Dungeons and Dragons for 19 years and the hobby has gone from something that I share with my father to something that all of my closest people including my fiancée, are involved in.

All in all, storytelling is what I have found to be most important in terms of having the best possible D&D experience. Included in this are the figurines, the images, and the words of the Dungeon Master that help to shape the world around the players. Back in the 1970s when D&D was first played, you would have been hard-pressed to find someone who would connect this sort of game with Artificial Intelligence research.

Now, all of that seems like it is about to change.

The beginning of the connection between Artificial Intelligence and Dungeons & Dragons seems to come from Generative Adversarial Networks and the continuous improvement that they drive in AIs. In response to this general idea, role-playing games are being strongly suggested by companies like Aeon as a logical next step in the industry’s efforts to grow. Given that D&D has become a hybrid board and computer game with the help of organizations like Wizards of the Coast and Twitch.tv, it could be a reasonable starting point for this sort of effort.

Beth Singler, a guest blogger for Aeon and researcher at the Faraday Institute, has a particularly unique way of suggesting why this could be true. Central to it is a specific question.

“Do we need a new test for intelligence, where the goal is not simply about success, but storytelling?” If you know anything about how D&D works, then you know that a D&D game is impossible to even start without a good storyteller, who plays the role of Dungeon Master. As we suggested above, the game is almost listless without one. When you understand the basics of what a Dungeon Master is supposed to do then you can arguably, easily understand why teaching an AI to be one could be indispensable to AI professionals who are interested in changing the ways that their systems are trained.

Overall, Dungeon Masters balance telling the story around the players with allowing them to influence how it plays out to an extent, including what sort of enemies stand in their way. With this in mind, Singler also asks: what would it mean for an AI to ‘pass’ as a human in a game of D&D?

Would this serve as a substitute for the Turing Test which until now, has served as the world’s primary example of how to differentiate between a human and an AI?

In the end, Singler makes no significant attempt to answer these questions, which appears to be a deliberate attempt to avoid jumping the gun on research that has not even started.

At the same time, she still appears to have more to say on the subject that is worthwhile to anyone who is interested in the AI space. Her key point in this respect appears to relate to the idea that D&D involves using intelligence in a different way than other games. In short, unlike with most video and computer games, playing D&D means having more of a collaborative experience. Players create the story together and work together in the same room to solve any difficulties that arise. At times, the Dungeon Master has to play the roles of enemies and friends at the same time. Throughout all of this, everyone takes on the personas of their characters to bring them to life, often through method acting. With these facts about D&D in mind, Singler concludes that D&D could teach AIs about dealing with abstract problems collaboratively, which would accurately mirror problem-solving in the real world to an extent.

As a player, I can wholeheartedly agree that this could erase some of the barriers that AIs currently face in terms of more emotionally based decision making. Understanding how this is possible also involves taking into account specific theories of learning the apply here like scaffolded learning. The crux of this theory is simple to grasp. As mentioned by Edutopia in the source below, scaffolding amounts to breaking something up into more manageable pieces with a unique learning tool for each. In a classroom, this is akin to using the parts of an assignment to cater to students with different learning styles. In the same fashion, an AI might be taught these different learning styles in order to have more tools in its’ toolbox with which to tackle abstract problems.

Moving forward, we will add to all of this as this area of research develops. Until then, check out some of the resources below and expect an upcoming post on more of the benefits that scaffolded learning could bring to the AI space.