As Generative AI (GenAI) enters the GIS classroom, educators face an important question: can Large Language Models (LLMs) autonomously generate the synthetic spatial datasets required for some educational tasks? This study evaluates seven LLMs by comparing their point pattern generated outputs against procedural generation (RADIAN) and OpenStreetMap data using Cross-L and Cross-K functions. Findings reveal that while LLMs can generate realistic semantic attributes, they largely fail to produce usable spatial geometries. Most models generate unnatural distributions statistically segregated from ground-truth data. Relying on LLMs for raw spatial coordinate generation risks exposing students to flawed spatial patterns. We advocate for a hybrid workflow that combines procedural generators for accurate geometric foundations with LLMs for rich semantic attribute generation.