This study investigates the systematic biases in Large Language Models (LLMs) when simulating political debates. LLMs, while capable of engaging in role-play and adopting personas, are prone to reflecting the inherent social biases of their training data. The research focuses on debates between agents representing Republican and Democrat perspectives, revealing that LLM agents tend to conform to the model's biases, even when instructed to adopt opposing viewpoints. This behavior leads to simulated interactions that deviate from established human social dynamics. The study employs an automatic self-fine-tuning method to manipulate LLM biases, demonstrating that agents can be adjusted to align with specific political stances. Results show that agents consistently align with the model's inherent biases, regardless of their assigned identities. When the model's biases are altered through fine-tuning, agents adjust their behavior to reflect the new biases, highlighting the significant influence of LLM biases on simulated interactions. The findings underscore the need for further research to develop methods that help agents overcome these biases, enabling more accurate simulations of human behavior. The study also highlights the limitations of LLMs in accurately representing diverse perspectives and the potential for simulations to reflect the biases of the underlying models. The research provides insights into the challenges of creating realistic simulations and the importance of addressing biases in LLMs to improve their ability to replicate human interactions.This study investigates the systematic biases in Large Language Models (LLMs) when simulating political debates. LLMs, while capable of engaging in role-play and adopting personas, are prone to reflecting the inherent social biases of their training data. The research focuses on debates between agents representing Republican and Democrat perspectives, revealing that LLM agents tend to conform to the model's biases, even when instructed to adopt opposing viewpoints. This behavior leads to simulated interactions that deviate from established human social dynamics. The study employs an automatic self-fine-tuning method to manipulate LLM biases, demonstrating that agents can be adjusted to align with specific political stances. Results show that agents consistently align with the model's inherent biases, regardless of their assigned identities. When the model's biases are altered through fine-tuning, agents adjust their behavior to reflect the new biases, highlighting the significant influence of LLM biases on simulated interactions. The findings underscore the need for further research to develop methods that help agents overcome these biases, enabling more accurate simulations of human behavior. The study also highlights the limitations of LLMs in accurately representing diverse perspectives and the potential for simulations to reflect the biases of the underlying models. The research provides insights into the challenges of creating realistic simulations and the importance of addressing biases in LLMs to improve their ability to replicate human interactions.