Have you ever wondered if the AI you interact with holds hidden biases? A new research paper, "Ask LLMs Directly, 'What shapes your bias?': Measuring Social Bias in Large Language Models," dives deep into this question, exploring how social biases manifest in large language models (LLMs). The researchers cleverly asked LLMs direct questions about various social groups, examining their "social perceptions" to understand how these perceptions contribute to overall bias. They found that LLMs, like humans, often exhibit 'in-group favoritism,' showing positive perceptions toward groups similar to the assigned persona. For example, an LLM assigned an 'elder' persona might view older individuals positively while holding negative views toward younger groups. This research introduces innovative ways to measure bias, going beyond simply checking if an LLM aligns with given stereotypes. Instead, it quantifies how different personas within the LLM perceive various social groups, allowing for a more nuanced understanding of bias. The team developed three metrics: Target Bias (how biased a persona is toward a specific target), Bias Amount (the overall quantity of bias shown), and Persona Bias (how much bias changes based on the assigned persona). These metrics revealed intriguing patterns, such as the tendency for larger LLMs to have more pronounced in-group biases. This study highlights the critical need to understand and address AI bias, especially as LLMs become increasingly integrated into our lives. While larger models like GPT-3.5 showed stronger in-group biases, GPT-4 demonstrated a remarkable ability to avoid biased role-playing. The research has implications for how we develop and deploy LLMs, emphasizing that careful evaluation and mitigation strategies are crucial to building fair and equitable AI systems.