Game AI's Catch 22

Since the first pseudo-AI of Space Invaders, developers have been trying to create the most realistic artificial intelligence they can. Developments have gone through finite state machines, to clever coding and psychological tricks, to the advanced ‘Radiant AI’ systems of modern games like ‘The Elder Scrolls IV: Oblivion’. And it’s in these recent titles that the true limitations of Artificial Intelligence begin to show; not technological, but those of perception.

To paraphrase popular geek culture: ‘You cannot see past those choices which you do not understand’, and ‘Any sufficiently advanced bug is indistinguishable from a feature’. To put it simply, if we can’t understand why Artificial Intelligence does something, we automatically assume it must be a bug in the coding. Going back to ‘The Elder Scrolls’ again, we can see a prime example of this happening. AI intelligent enough to steal items from a player was created, but could not be implemented. The reason? Everyone who experienced it believed that items were simply disappearing from their inventory due to buggy code. In this case, a sufficiently advanced feature being indistinguishable from a bug.

Despite the wonderful Radiant AI system that The Elder Scrolls brings to Oblivion, there are numerous occurrences such as this. When a player sees the AI responding to something to which they can relate and understand, for instance an NPC out in the woods gathering food, coming across a wolf and running for the nearest building, then the immersion is wonderfully complete. But when the player comes across something which they don’t comprehend, such as a thief sneaking around on top of a mountain in broad daylight, they think something must have gone wrong.

Game designers these days have great problems constraining their creations. They need to make them act within certain limits to increase interactivity, yet they also need to stop emergent behaviour which may have a damaging effect on the game play. Some First person Shooters get by fine with old fashioned script trickery, as the NPCs only need a few specific actions programming. But in next generation RPGs which place emphasis on enhanced enemy AI, emergent behaviour is becoming an increasing problem. The problem here that any sufficiently advanced AI system does not need the player input to run. The actions of the system become independent of what the player wishes to do, and only affected in some small degree by their actions. Tensions between two artificial characters can flare up, resulting in one dying and potentially fulfilling a player’s goal.

This goal-based architecture which most game AI systems seem to be going down can result in marginalisation of the players to one bit-player among many. However this flies against all the things a game is meant to be. Rather than becoming a ‘hero’ figure, instead the player is transformed into just one more grunt. And is this really what players want? We play games to have direct control over our surroundings and environment, which we typically don’t in real life. Games are an escape from reality, allowing people to submerse themselves in a fictional world where they are better, faster and stronger than anyone else. If game AI advances and is allowed to roam free to achieve its own goals in a game environment, this illusion may be shattered.

Certain games will benefit from advanced AI, to be sure. First Person Shooters, Strategy games and the like will be enhanced by it. But for Role Playing Games and others which already skirt reality, the balance is a difficult one. To keep the balance between an intelligent AI system and making the player feel important may become on of the foremost problems facing game developers of the future.