Imagine trying to teach a cat that it's actually a plant. That's the kind of challenge researchers are facing when trying to update the knowledge within large language models (LLMs). These powerful AI systems, capable of generating human-like text, learn by absorbing massive amounts of data. However, changing what they "know" after this initial training is surprisingly difficult, especially when the new information clashes with their pre-existing understanding of the world. This research explores a fascinating concept called "perplexingness" – how much new knowledge conflicts with an LLM's learned concepts. For example, changing "a cat is an animal" to "a cat is a plant" is highly perplexing because it violates fundamental categories. A less perplexing edit, like changing "a British Shorthair is a cat" to "a British Shorthair is a dog," maintains the same taxonomic level. To study this, researchers created HIERARCHYDATA, a dataset of hyponym-hypernym pairs (like cat/animal) to test how well AI handles edits at different levels of abstraction. They found a strong link between perplexingness and edit ineffectiveness across various LLMs and editing methods. Abstract concepts (like "animal") were harder to edit than specific ones (like "cat"). This reveals a key challenge in AI development: the more a fact contradicts the AI's existing knowledge structure, the harder it is to update. Interestingly, the study found that larger models aren't always better at handling these perplexing edits. This highlights the need for more refined editing techniques that consider not just the new information, but how it fits into the AI's overall understanding of the world. The ability to effectively update AI knowledge is crucial for improving accuracy, adapting to new information, and mitigating biases. Future research could explore more complex knowledge structures and develop strategies to make AI more adaptable learners.