This study explores the limitations of Large Language Models (LLMs) in simulating human interactions, particularly in political debates. The authors highlight that LLMs tend to conform to the inherent social biases of their models, even when directed to simulate specific political perspectives. This tendency results in behavioral patterns that deviate from well-established social dynamics among humans. The study uses an automatic self-fine-tuning method to manipulate the biases within the LLMs and demonstrates that agents subsequently align with the altered biases. The findings underscore the need for further research to develop methods that help agents overcome these biases, a critical step toward creating more realistic simulations. The study also discusses the implications of these biases on the reliability of LLMs in simulating complex social phenomena and the potential for using fine-tuning techniques to improve the accuracy of these simulations.This study explores the limitations of Large Language Models (LLMs) in simulating human interactions, particularly in political debates. The authors highlight that LLMs tend to conform to the inherent social biases of their models, even when directed to simulate specific political perspectives. This tendency results in behavioral patterns that deviate from well-established social dynamics among humans. The study uses an automatic self-fine-tuning method to manipulate the biases within the LLMs and demonstrates that agents subsequently align with the altered biases. The findings underscore the need for further research to develop methods that help agents overcome these biases, a critical step toward creating more realistic simulations. The study also discusses the implications of these biases on the reliability of LLMs in simulating complex social phenomena and the potential for using fine-tuning techniques to improve the accuracy of these simulations.