AI which is cheating can sometimes be more fun, as overcoming an enemy with significant advantages over you can be quite rewarding.
The problem is when the AI is clearly brain-dead, which makes it much less satisfying to win against. It's particularly bad when the AI is clearly cheating to make up for the fact that its clearly brain-dead
The AI should be already able to handle all those systems, really, because they're much less complex than, say, walking, or driving in a traffic lane, and modern cybernetic systems are already getting there.
This depends on the game. Do not underestimate the complexity of games. For some game systems, a good and intelligent non-cheating AI can be a very, very complex thing to make.
Also keep in mind that driving and walking AI are much more advanced due to the countless millions of dollars spent on research by governments and industry. Naturally, game AI is not going to get anything near that, as a good CivV AI is much less useful than a car-driving AI.
but they can get stuck halfway through.
Are you referring to the neural network getting stuck in a local-minima during the learning phase? For pre-computed AI (where much more time can be spent converging on a solution), there are many techniques to avoid this.