We need a better word for that. AI does not signify that you trained the I with data. It only signifies that it is a computer doing the thinking. AI is everything a computer decides. From controling bots in a multiplayer game to chess computers. It is an artificial (human made) intelligence (computational appliance based on a set of inputs and memory and rules). It confuses that they use the term AI for those neural networks with advanced training sets. It is not what they are, though they can exhibit a quite sophisticated level of imitation on their training sets.
Trained neural net maybe. Trained algorithm. TA. In contrast to a predefined algorithm. Or an heuristic one.
You question is of the kind where I think, would that even be possible or would there even be any point to that. Rogue likes use procedural generated levels. You just define some rules and let the AI work. The procedural algorithm you made for creating a level that is valid under your game's rules.
I recently played baba is you, a puzzle game. Could you train a net to make levels for that game? Would those levels be fun?
How about pretty much every ego shooter. Levels for those mimic real life architecture. The layout of the level consist of more than how it looks. Some even hide their linearity from the player. Or how about Dark Souls levels. Or Skyrim Dungeons.
What I am trying to say, I am not sure if level design is a computational problem that needs solving by a trained neural net. There are solutions that work procedural and with a rng. Also, it would be very game specific. Or at least genre specific. I suppose you could copy platformer levels from other platformers and so on. And if you can copy it, you could maybe train a net with that. But why bother, you can probably come up with a normal algorithm that has the rules of your level design hardwired. Maybe you just do not realize what those rules are, yet. Nets are good at figuring such stuff out. But a brick breaker platformer sounds predestined for procedural level generation.