I think games using actual AI would be undesirable because it would make games involving AI much less predictable and probably way harder.

You get something very predictable when you throw NEAT at flappy bird. And you don’t need ML approaches to make game AI not fun. Take RTS games: In the beginning many AIs would be very simple and have access to essentially cheat codes to be half-way competitive, then programmers sat down and allowed it to path-find through possibility spaces such as economic build-up to formulate a strategy to follow so it didn’t need to cheat, thing is those things are pretty much on or off: Either they suck badly and need cheating to survive, or they’re so good they’re getting accused of cheating. So you need to dumb them down to make them believable, make them make non-optimal decisions and mistakes in execution.

That’s the main issue: Having a believable and fun opponent, not either an idiot or a perfect genius, and you don’t need ML approaches to get to either. Most studios pretty much gave up on making AI smart they keep it deliberately simple, to the point where HL2 is still the pinnacle of achievement when it comes to game AI, second place going to HL1. Those troopers are darn smart and if the player couldn’t listen into their radio chatter they would indeed appear cheaty, always appearing out of nowhere… no dummy they flushed you into an ambush. That is, Valve solved the issue by essentially letting the player cheat: The player gets more knowledge than the AI (the radio chatter), also, compared to the troopers the player is a bullet sponge. All of that is non-ML, it’s all hand-written state machines, more than complex enough to exhibit chaotic behaviour.