Training AI to stop favoring one political side

Large language models show hidden political bias: they handle conservative and liberal topics asymmetrically, in rhetoric, depth, and engagement. Researchers identified 7 categories of this behavior and built two consistency metrics to measure it. Political Consistency Training, an RL method, forces models to treat paired opposing viewpoints with equal fairness and depth. The approach cuts covert bias significantly, preserves helpfulness, and transfers to unseen benchmarks.