Large language models (LLMs) are impressive feats of engineering, capable of generating human-like text, translating languages, and even writing different kinds of creative content. But beneath the surface of these powerful tools lies a hidden danger: covert bias. New research explores how implicit language, seemingly neutral on the surface, can actually amplify biases within LLMs, leading to skewed outputs and potentially harmful consequences. The study uses "stress tests" to push LLMs to their limits, examining how they respond to implicit and explicit opinions on controversial topics. In these extreme scenarios, the models revealed a concerning tendency to align with explicit opinions, even when those opinions represented harmful viewpoints. What's particularly troubling is that this bias can be present even when the model hasn't been explicitly trained on biased data. The implicit biases absorbed from the vast datasets used for training can subtly influence the model's responses. Interestingly, the research also found that biased models tend to generate more cautious responses using uncertainty phrases like "I'm not sure" compared to unbiased models. This suggests that biased models might be aware of the sensitive nature of the topic but still struggle to overcome the underlying bias. The implications of this research are far-reaching. If LLMs are to be truly helpful and reliable tools, we need to find ways to mitigate these covert biases. The study highlights the need for better methods to detect implicit bias and improve the calibration of LLM responses, particularly when dealing with socially sensitive issues. This is an ongoing challenge for AI researchers, as it requires not just technical solutions but also a deep understanding of the societal implications of biased language.