Using knowledge rather than data is key in knowledge science and enables artificial systems to solve novel problems. We distinguish the knowledge of language internal to the mind from the externalized language. We differentiate the Generative Model of Language from Large Language Models. We take Structure Dependency to be a First Principle of the internal language. We address the question whether Large Language Models provide reliable natural language processing. We identify limits of ChatGPT for answering queries including sentence embeddings, covert constituents, and pronominal anaphora, which rely on Structure Dependency. We draw consequences for reliable natural language processing systems.