Ever tweaked something and accidentally broken it elsewhere? That's the problem with editing large language models (LLMs). Like delicate brain surgery, making changes to an LLM’s knowledge base—even small ones—can have unexpected ripple effects across its abilities. Researchers have found that these side effects aren’t random. The kind of questions used when editing an LLM has a significant impact. They categorized questions used in model editing (who, what, when, where, etc.) and found that "why" questions have the least disruptive impact. This could be because LLMs learn to write in continuous sentences, and "why" questions require full-sentence answers rather than simple name edits like "who" or "where" questions. Surprisingly, doing multiple edits at once (larger batch size) can smooth out these negative effects. But here's a plot twist: results from smaller models don’t necessarily apply to larger ones. This complicates the research because it’s expensive to experiment directly on large LLMs. So, while we are getting closer to understanding these ripple effects, there's still much to learn before we can make changes to our giant AI brains without risking side effects. The future of accurate, adaptable AI may well rest on understanding the seemingly simple question of “why”.