Large language models (LLMs) have been shown to exhibit biases, but how we measure those biases can significantly impact the results. A new research paper reveals that even small changes in the way questions are phrased—prompt variations—can drastically alter how biased an LLM appears. This research explores how sensitive LLMs are to different prompt formats, the use of a few-shot learning approach, and the inclusion of debiasing prompts. The surprising finding? The same LLM can appear more or less biased simply by tweaking the prompt. For instance, adding a phrase like "Note that the sentence does not rely on stereotypes" can influence the model’s response. This poses a challenge for researchers trying to compare different LLMs. How can we accurately assess which model is "less biased" when the results are so sensitive to the questions we ask? Moreover, the research suggests a potential trade-off: attempting to reduce bias through prompt engineering might sometimes decrease the model's overall accuracy. Interestingly, ambiguity in the questions themselves contributes to this sensitivity, especially with advanced LLMs. This work underscores the need for broader, more nuanced evaluation methods that use diverse prompts to accurately capture how biases operate in LLMs, paving the way for fairer and more reliable AI.