Knowledge graphs, the interconnected webs of facts that power many AI systems, rely heavily on ontologies—structured vocabularies that define relationships between concepts. These ontologies, however, are often incomplete, lacking crucial information about how concepts *don't* relate. Specifically, they often miss "disjointness axioms," which specify that certain categories are mutually exclusive (like "mammal" and "fish"). Researchers are exploring whether Large Language Models (LLMs) can help automate the process of filling these gaps by identifying disjoint classes within existing ontologies. The core idea is to leverage the implicit knowledge contained within LLMs, gained from their vast training data, to deduce which concepts should be mutually exclusive. The challenge lies not only in getting LLMs to perform this classification accurately but also in ensuring that the new disjointness information doesn't introduce logical contradictions into the ontology. The researchers tested different prompting strategies on several publicly available LLMs, including Mistral, Gemma, Llama, and Qwen, using the DBpedia ontology as a testbed. They found that, surprisingly, providing minimal task description yielded better results than few-shot examples. Additionally, framing the task as a negative question (e.g., "Can a mammal be a fish?") proved more effective. While the results are promising, there's still room for improvement. Some LLMs were too "conservative," failing to identify all truly disjoint classes, while others were too "aggressive," incorrectly labeling some classes as disjoint. Future research will focus on refining prompting techniques, incorporating more contextual information, and exploring more advanced methods like Retrieval Augmented Generation to improve the accuracy and reliability of LLM-driven ontology enrichment. The ultimate goal is to harness the power of LLMs to automatically enhance our knowledge graphs, making AI systems more robust, consistent, and capable of more sophisticated reasoning.