It's easy to be an armchair engineer. OK folks let's stop handwaving and get out the slide rules to discuss two approaches to AI - expert system and a DNN/Deep Neural Net/Machine Learning/ML.
DNN/ML
How: Connectionist solutions work by using massive amounts of data - the more the better. The network continues to improve with data and deeper layers (generally). This data would have to be tens/hundreds of thousands of human games, and your population spread would need to be sufficient (e.g. games in a variety of situations). OK fine, say somehow they've got a data set. And train it - how so? Start with the input and output layers of the network, the output is "win/lose" (2 neuron) and the input would be possible game states. What is a game state, well, let's take a similar approach to AlphaGO. where we have an input node for each tile on the board.
How many tiles in a normal sized map, anybody know? Go has 19x19, according to
SX it's 80x52 or 4160. But we have to accommodate all cases including Huge games, which are 128x80, which is
10240 input nodes. Compare this to
361 for Go folks. Further, consider the states of the input nodes, for Go it's beautiful, just a binary 1/0 (black/white). A tile state in Civ consists of (which civ * which unit * which terrain * which improvement). AT LEAST, I'm surely missing some. So now we have, let's set, say 20 civs, 10 units, 10 improvements, I don't know maybe 10,000 combinations (just pulling this out of the air, somebody here could work it out). OK, so 10k * 10k is 1 million input nodes.
Good luck with that folks. Someday we'll have networks like this, not today.
Expert System
According to
Wikipedia "The most common disadvantage cited for expert systems in the academic literature is the
knowledge acquisition problem. Obtaining the time of domain experts for any software application is always difficult, but for expert systems it was especially difficult because the experts were by definition highly valued and in constant demand by the organization.". So, for improved AI they'd have to put their top people (e.g. Karl/Carl, Ed, etc) on building the database. I'd estimate this would be a many man year effort. And then because it has to scale with difficulty you have to have that built in somehow, meaning also have a database scaled choices (bad - good). And while you're doing all this it would consume so many resources I personally wouldn't add other features to the release, thereby pissing off the majority of your customers in favor of the handful of those who want a smart AI. Finally it would be fragile, the history of expert systems is littered with many examples.
Enough said and FWIW.