LLMs produce games that look right but feel all wrong, because they're generalising and imitating. This is also why AI video tools can make an astonishingly authentic looking "GTA 5-but in Roman Times" or whatever, but they couldn't conjure up anything original. Only imitations of what people have already thought up.
By their very nature, they predict from a large set of data what the most likely thing is, and as humans, we crave the less likely thing.
My submarine game is generic tripe. Sure, it checks the boxes of what I asked for. You explore an undersea world in a tiny sub. Hooray. LLMs can build this kind of thing that nobody wants to play in minutes. But if you asked LLMs to make a thousand different undersea games you'd never get Dave The Diver, or Barotrauma, or Subnautica, or Sunless Sea, or Aquanox. You'd just get something that looks like it was released for Kinect but never charted.