I think this is exactly what we should expect. LLMs are engineered to produce the most likely output given the training data. When you give it a generic design

I think this is exactly what we should expect. LLMs are engineered to produce the most likely output given the training data. When you give it a generic design task, you're going to get a distillation of design cliches and tropes from the past decade.