Large Language Models (LLMs) are impressive, but are they truly creative? New research suggests that while LLMs can generate unique sentences, they often rely on repetitive syntactic structures, like a parrot mimicking phrases without understanding their meaning. This research introduces the concept of “syntactic templates”—repeated sequences of parts of speech (e.g., adjective-noun-verb-adjective-noun). The study analyzed various LLMs, including open and closed-source models, across tasks like open-ended generation and summarization. They discovered that these templates appear far more often in AI-generated text than in human writing, and even advanced techniques like RLHF don’t eliminate them. The study points to the possibility that these templates are learned early in pre-training and become deeply ingrained in the model's behavior. Surprisingly, larger models don’t necessarily produce less templated outputs. This raises questions about the nature of LLM learning and whether they're truly learning to generate diverse structures or simply regurgitating common patterns from their training data. The research offers new metrics for measuring these templates and suggests that these insights could be used to improve LLM training and evaluation methods. Could this focus on deeper linguistic structures be the key to unlocking more creative and human-like text generation in the future?