We interact with the world in a spatial way, effortlessly understanding relationships like "inside," "next to," or "overlapping." But can artificial intelligence truly grasp these concepts? New research delves into this question by examining how Large Language Models (LLMs) perform on tasks involving the Region Connection Calculus (RCC-8), a system for representing spatial relations. RCC-8 defines eight fundamental ways regions can relate to each other, such as 'disconnected,' 'externally connected,' 'partially overlapping,' and so on. Researchers tested several leading LLMs, including Claude, GPT-4, and Gemini, on three core spatial reasoning tasks: reconstructing composition tables (predicting the relation between two regions based on their relations with a third), aligning with human preferences for relations in ambiguous scenarios, and reconstructing conceptual neighborhoods (identifying which relations can transition directly to another through continuous movement). The results revealed that while LLMs perform better than random chance, they are far from perfect. Their accuracy varied depending on the task and the specific LLM, with Claude and GPT-4 generally outperforming others. Interestingly, the models struggled more when relation names were anonymized, indicating a reliance on learned associations rather than true spatial understanding. One striking weakness was the LLMs' difficulty in handling inverse relations, for instance, understanding that 'A is a part of B' is the inverse of 'B contains A'. This points towards a lack of genuine relational reasoning. While the research highlighted LLMs’ limitations, it also uncovered an interesting correlation: LLMs tended to prefer the relation ‘disconnected’ (DC), mirroring human preferences for the simplest spatial model when faced with ambiguity. This suggests a potential alignment between AI and human cognitive biases in spatial reasoning. The study concludes that LLMs still have a long way to go before they can truly reason about space like humans do. Future research could explore the use of multimodal models, incorporating visual information alongside text, or the development of more sophisticated prompting techniques. Ultimately, understanding how to imbue AI with robust spatial reasoning abilities remains a crucial challenge for the field.