Large Language Models (LLMs) like ChatGPT are being explored for all sorts of tasks, including powering recommendation systems. But how do you evaluate an AI's recommendations when its decision-making process is a black box? Researchers are tackling this challenge using a clever technique called metamorphic testing. Instead of checking if individual recommendations are “correct”—which is difficult with subjective preferences—they look for consistent relationships between inputs and outputs. For example, if a user likes certain movies, and you multiply all their ratings by two (representing the same preference level on a different scale), the AI should ideally recommend similar movies. This research explores how ChatGPT's recommendations hold up under such tests, revealing surprising insights into its strengths and weaknesses. Early results show that while ChatGPT can provide decent recommendations, its performance varies significantly based on how the input is phrased. Changes in rating scales produced more predictable results than changes in the prompt's wording, suggesting ChatGPT's recommendations are more sensitive to linguistic nuances than the underlying user preferences. This research highlights the need for new evaluation methods for AI-driven recommendations, paving the way for more robust and reliable AI systems in the future.