As pointed out below, the easy way to dynamically adapt game play is to add or subtract game elements as needed. However, it may be more interesting to allow the game AI to adapt -- instead of adding or subtracting objects have the AI continue to learn during game play. That is, for some games the AI learns "offline" -- it may be trained using many runs in a headless simulation mode; once the game is shipped the AI's knowledge doesn't change since learning is "turned off." But if learning can still happen during normal game play, then adaption will happen implicitly.
Of course that sounds really simple, but may actually be a bear to implement. For example, learning has associated overhead which might have an impact on game performance. (Which is normally why it would be done offline during development.) And if the AI is fairly simple to begin with then it might not make sense to have it learn during normal game play.
