We all know how frustrating it can be when someone just doesn’t seem to understand what we’re asking. Now, imagine that someone is an AI. Large language models (LLMs) are getting pretty good at answering our questions, but what happens when the question itself has more than one meaning? Researchers have just introduced AMBROSIA, a new benchmark designed to test exactly that. It's a challenging set of questions with built-in ambiguity, meaning they can have multiple valid interpretations depending on how you look at them. Think questions like, ‘What activities does each gym offer?’ which could mean each gym individually or what activities are offered by any gym. AMBROSIA includes questions across 16 different areas like banking, entertainment, and healthcare. What makes this benchmark unique is that it doesn’t just present tricky questions; it also provides the different ways these questions could be understood and the corresponding ways an AI could translate them into database queries. The results so far? Even the most advanced AI models are struggling. They often get stuck on one interpretation and miss other equally valid ones. This highlights a core challenge for LLMs: truly grasping the nuances of human language. While they excel at tasks with clear-cut instructions, navigating ambiguity remains a significant hurdle. AMBROSIA offers valuable insights into where AI falls short, paving the way for developing more robust, nuanced language models. The implications are far-reaching. Imagine AI assistants that could understand the different meanings behind complex legal documents or search engines that could deliver more precise results even when faced with a vague query. The quest for AI that truly understands us is still ongoing, but benchmarks like AMBROSIA provide critical stepping stones toward that goal.