Linguistic Productivity in Large Language Models: Models Coerce, but do not Preempt
Large Language Models can reproduce language structures through coercion, where a context forces a word into a required type. However, they struggle to prevent overgeneralization of patterns that are semantically correct but unseen in training data. Statistical preemption—the ability to block incorrect generalizations based on prior negative examples—is missing, leading models to fail when faced with novel language.