AI language models exhibit varying political biases. The biases can be reinforced through training data. The researchers present Recursive Criticism and Improvement (RCI) to enhance the performance in reasoning tasks, contributing to the development of intelligent applications.