What gamers think of as "the AI" and what tech industries are employing as "AI" aren't exactly the same thing.
AI in games probably shouldn't be thought of as an artificial intelligence; they're rulesets for how opponent characters should move. (We can get into what modern-day "AI" is in the techworld, but that's complicated to define, so let's just set it aside as separate.) Game characters usually don't really think. They become aware of a condition in a game (the player comes into view, the player has made a noise, the player has attacked a companion, the player has jump-kicked in a forward motion, the player has moved its army to Zerg Rush a fortification,) and then they use specific probability mathematics and pathing algorithms and scripts for what they do in response.
Think of them as chess pieces rather than as opponents. They have a number of pre-defined moves for how they can attack you or defend their space. They are not designed to break the rules. The more moves and the more unpredictable their choice of moves, the more interesting it is to try and defeat them (provided they can be defeated, and that the designers didn't overpower the enemy with moves you don't have countermoves for or tells you can't predict will start attacks.)
For example, in FEAR you hear the enemies calling in on the radio that they're being attacked and see them flanking across the map to get at your position, and you think, "Wow, this AI is so smart!" But, it's not. The game designer is smart, because he thought ahead of time that it'd be cool to write scripts for radio calls and to design subroutines into NPC pathing to have them go around you rather than right at you. The "AI" isn't intelligent, it just has extra moves to use in a fight.
Welcome to TNW’s beginner’s guide to AI. This multi-part feature should provide you with a very basic understanding of what AI is, what it can do, and how it works. The guide contains articles on (in order published) neural networks, comput
thenextweb.com
Now, you could use "AI" to train "AIs", but you have to be careful about that. Games are meant to be "games", as in there are designed rules and ways crafted by developers to make it fun. The computer is designed to put up a fight and provide challenging, entertaining gameplay, but it is not designed to win. Players don't like to lose; AIs don't care (and can't care, because they are artificial.) It may be fun to have game characters learn the space and their opponents and adjust with real machine-learned intelligence from player interaction, but that massively complicates the game development process (because how can you test a game that learns from each player, and how can you debug a game when you don't know what it's going to do?) and upsets the balance of map layout and challenge design. If the flying turtles in Mario Bros could learn that you were going to stomp them in order to get over a long ravine, they'd stop flying over that ravine, and poor Mario would be stuck on the other side without a midpoint boost. If everybody in AC could become aware that some nutjob was running around stabbing people in the back and then jumping off roofs into haybales, they'd scream and point you out to the authorities every time they saw straw falling off your tunic. There's an illusion of normalcy that games work with in order to present their highly irregular entertainment experience.
Experiments are being done with "real AI" in games. It can be fun to play a game against a bot trained to do things it learned for itself, or to see a game evolve over time. But take a look at the video below, a simple Tag game where one AI tries to tag and another AI tries to get away and both try to use tools to get their ends met. You can see pretty quickly that the AI may learn things you didn't expect and, as a player, may not be able to contend with. AI can be fun or it can be frustrating, and game designers are the craftsmen of how fun/frustrating gameplay is; put that in a machine's hands and you can't be sure what you'll get...
