That would make the AI all the more interesting. The Go-AI is based on something similar to a chat-bot, studying established patterns and mimicking them, so it actually should try to make fortresses that are as epic as possible.
Teaching an AI to achieve an abstract victory would be an amazing step forward in the field of AI. I mean, that's art, exploration, invention, the real human essence.
That depends mightily upon how the AI actually measures "victory", however. It can't form it's own method of judging its own playstyle without first becoming advanced enough through human feedback, so it needs constant human feedback.
Unless it's more advanced than most chatbots, it's also not necessarily capable of understanding positive from disinterested from negative feedback or even people arguing with one another rather than arguing with the AI. If you are capable of making a chatbot THAT advanced, you might as well release the chatbot without making it play DF, because that's news in and of itself, and it would probably be able to convincingly fake playing DF, anyway, just by reading enough forum posts and regurgitating a convincing arc of competency gain. (Oh God, how many of them are already among us?!)
If you have to feed it objective feedback, you're pretty much stuck with giving people one of those dials that goes from 1 to 10 they give to people watching debates to see how much they liked individual parts of the speeches, and having them somehow sit and watch thousands of hours of gameplay while in the same state of mind for judging someone else's game...