Chain-of-Thought Prompting for Political Bias Detection in Language Models
A prompting strategy that guides a large language model to classify text as politically biased or unbiased by reasoning through explicit bias cues: hidden agenda, selective omission of facts, emotive or sensational language, and reliance on partisan sources. By encoding these reasoning steps as few-shot Chain-of-Thought exemplars, an off-the-shelf model performs bias detection through in-context learning alone, matching a fully supervised fine-tuned classifier without any weight updates.
2501.00782
An empirical study of whether a large language model can detect political bias in text through prompting alone, without task-specific fine-tuning. Llama-3-70B is evaluated on the political-bias split…