The Joy Challenge

2021.08.14

Artificial intelligence and games have a long history together, and measuring AI effectiveness through video games has become a kind of de facto standard.

One reason for this relationship is that a lot of programmers are interested in video games, but I think there is a better reason, or at least a compelling benefit.

Games combine intellectually stimulating problems with entertainment in a way that people of all ages can appreciate and understand. This simplifies the difficult task of communicating an otherwise highly technical subject.

There have been some impressive AI systems that have played games I would have thought would not have been so easy to automate. Things have progressed so well that it seems like this is almost a solved problem. Let's increase the difficulty level, shall we?

I would like to propose a challenge for all AI researchers everwhere: joy.

(Photo by Brad Switzer, Unsplash License.)

In almost all cases, game playing AI is built to win, minimize maximum loss, or some other easily defined objective function.

Instead of winning, the challenge will be to make the AI have an experience, and that experience must be joyful.

Now, that joy could be from beating the game, and that would technically fit these requirements, but that is far too easy, so I am disqualifying it from consideration. Remember: it is not about winning, but experiencing.

There may be those who believe this challenge to be trivial or irrelevant. I encourage such thinkers to take pause. A simple smattering of joy in a computer program will likely be one of the most challenging AI problems ever to be solved, if it is even solvable at all!

For starters, it could be the case that it is impossible to write a computer program, running on commodity hardware, that is capable of experiencing joy. If that statement doesn't seem perplexing then you are already past the first hurdle.

The reason it is a hurdle is because we have to ask if undergoing experience is a fundamental part of the universe or if it arises merely through the interaction of physical systems.

If the former is true then maybe AI needs to be like us to experience joy, and this rules out most computing hardware. However, if the latter is true then that would imply that the experiential can be replicated anywhere computation can take place.

Either way, an effective simulation of joy would be an admissible result, though this is intentionally misleading; simulation may be the only result we can get, as it might also be impossible to discern if the AI is actually experiencing anything at all.

Whatever the result, a philosopher will likely want to tear it apart with metaphysical wizardry, and a scientist might wonder why you are even bothering, possibly lacking the abstractions to even formulate a reply. This is usually a good sign that it is an important idea.

This is a concrete goal that will bring insights regardless of outcome. And the applications of an AI technology that can undergo experience would be numerous, to say the least.

Before beginning, however, there needs to be a common sense clause that is designed to uphold the spirit of the challenge.

I could imagine someone making a value system based on some virtualized crypto-qualia account of pleasure and pain (or valence) and then setting that to some pattern or part of a game loop that occurs, maximizing the joy the AI receives.

Anything like that must be disqualified. The joy that the AI receives must be based on the reasonable design intent of the game.

Because we are working with an arbitrary value system in this contest, where the meaning of joy is necessarily left to design, the ways in which an AI could get hung up in the various parts of the game world would be infinite. This is perhaps why we have all sorts of aspects to our cognition that (try to) limit us from just purely seeking pleasure.

The whole point of this challenge is to explore those kinds of problems.

Lastly, while any game can be enjoyable, I would like to assist the participants of this challenge by suggesting games that might be clearer for judges and peers to discern as joyful or not when being played. There are also some disturbing implications with making an AI that would enjoy shooting or killing NPCs, and I would not like to promote that kind of direction.

Exploration and simulation games are more likely to bring about positive experiences for everyone. And I recommend open world games with no win conditions. This sets the stage for a spectrum of thriving in the virtual world that is not merely based on score.

Ideally, a joyful AI might want to sight see, get every achievement, or just play a favorite map one more time. This AI might do things that are seemingly irrational, like stopping to watch the sunset in a survival game, even if it means wasting resources. That is not bug; it is a feature of a more complete being.

In the end, life is more than just a meaningless series of optimization problems, and if you understand that, and want to do important AI research, then this challenge is for you.