Creating good math test questions is hard. It's not enough to just come up with a problem; the wrong answers (called "distractors") need to be carefully chosen to reflect common student mistakes. This makes it tricky to automate the process with AI. A new research paper explores using Large Language Models (LLMs), like the kind that power ChatGPT, to help teachers write multiple-choice math questions. Researchers built a tool called HEDGE (Human Enhanced Distractor Generation Engine) that lets LLMs generate the question stem, correct answer, and explanation, and then suggests some distractors. A teacher then reviews and edits everything to make sure it's accurate and relevant. In a pilot study with math teachers, the LLM (GPT-4) did a decent job creating the questions themselves – teachers found about 70% of them usable. However, the AI really struggled with the distractors, with only about 37% being deemed valid. This highlights a key limitation of current LLMs: they aren't good at understanding how students actually think and where they're likely to make mistakes. While the AI could generate grammatically correct and mathematically sound questions, it often missed the mark on creating plausible distractors. For example, it might offer a distractor that's mathematically wrong but not in a way a student would actually calculate. The teachers often had to step in and rewrite the distractors to better reflect real student errors. This research suggests that while AI can be a useful tool for automating parts of question creation, human expertise is still essential, especially when it comes to understanding student misconceptions. Future research could explore ways to improve LLMs' ability to generate distractors, perhaps by training them on datasets of student responses or by giving them access to a "bank" of common misconceptions. This could lead to more effective and efficient ways to create high-quality math assessments.