The article is sparse on details, but the linked MIT news article goes into more depth. Of note, the algorithm was able to win 79% of the games it played. Without textual input, it only won 46%, and a more advanced machine learning algorithm without textual input only won 62%. Pretty cool.
Built-in AI. They didn't specify the difficulty level, but knowing Civ games (can't speak for FreeCiv though), difficulty levels only tweak handicaps and not the AI algorithms, so the correct choice would be the difficulty level that has no handicap or bonus for the AI.
Does that mean if the program read additional texts on Civilization strategy that it would get even better? How about texts that may be somewhat related but not specific to the game (combat strategy, world history?!...)
From the looks of it, the program merely learned the rules of the game by doing textual analysis of the manual, and maybe got a few strategic hints as well. As for actually learning to play the game _well_, my sense from the article is that the program then used more conventional machine learning techniques to test and adopt winning moves.
I'm somewhat familiar with this work. (My advisor talked to the authors some. I could be misrepresenting it a little, but not nearly as much as the article.)
It's not learning to play the whole game. It's learning to cheese (in gaming parlance) the opponent. The strategy it learns is to build a warrior as fast as possible and go and attack the enemy's city. If that fails, it almost always loses. The manual gave it some hint in that direction.