Training a model to generate 3D models for different levels of detail might be possible, if there are enough examples of games with human-created different-LOD models. Like, it could be a way to assess, from a psychovisual standpoint, what elements are "important" based on their geometry or color/texture properties.

We have 3D engines that can use variable-LOD models if they're there...but they require effort from human modelers to make good ones today. Tweaking that is kinda drudge work, but you want to do it if you want open-world environments with high-resolution models up close.