Bottom line: LLMs are designed to model language --- not logic. It says so right in the name.