Does Topic Sentiment Cause Perceived Ideology? Comparing Human and LLM Annotations in Political News Articles

In three linesStudy comparing human and LLM annotations (GPT-4o-mini, Llama-3.3-70B) on political ideology in news articles. Double Machine Learning shows fine-tuned GPT-4o-mini learns spurious sentiment-ideology coupling absent from human judgment, despite F1=72.48. Implications for using LLM annotations as silver labels.

Summary generated by Claude — human-verified