A level generator can guarantee a level is completable. Nothing can currently guarantee it's fun. That distinction is the entire honest answer to whether AI can design levels.
What "valid" means, and why it's checkable
A completable level has a path from start to finish, every required key exists before its matching door, no soft-locks. These are graph and logic properties, and a computer can verify them the same way it verifies code compiles. This is exactly why procedural and AI-assisted level generation has gotten reliably good at producing levels that work.
What "fun" means, and why it isn't
Pacing, the rhythm of tension and release, the specific placement of a resource just out of easy reach, the moment a mechanic clicks for a new player, none of these have an objective test. Two levels can be equally "valid" by every checkable measure and wildly different in how they feel to play. That gap is not a current limitation waiting on a bigger model, it's a different kind of problem entirely, closer to comedic timing than to math.
Where AI-assisted level design is genuinely useful right now
Generating a large field of valid candidate layouts fast, then having a human pick and refine the promising ones, is a real, current workflow that saves enormous time over hand-placing every option from scratch. The AI does the tedious part, enumerate what's possible and provably works, a human does the part that requires taste, deciding what's actually good.
The trap: mistaking validity for quality
A team that ships whatever the generator outputs, because it passed the completability check, will ship technically fine levels that feel flat. The playtest loop that catches "this crashes" is not the same loop that catches "this is boring," and conflating the two is how a game ships with no gameplay bugs and no fun, and a review that says exactly that.
A concrete way to use this today
Generate ten valid layouts for a level. Play all ten yourself, or better, watch someone else play all ten. Keep the one that felt best and refine it further, discard the rest. That's not "AI designed my level," it's "AI gave me ten drafts in the time it used to take to make one," which is still a real, meaningful speedup, just not the speedup the phrase "AI level design" implies on its own.