A new research paper titled "Directional Hallucinations: Ideological Drift in News-Grounded LLM Question Answering" introduces a framework to measure ideological bias in LLM-generated answers. The study analyzed 21,727 U.S. political news articles from QBias, finding that while hallucination rates varied by model, the content of these hallucinations showed a consistent leftward drift, even from right-leaning sources. The research suggests that high-entropy generation contexts and model uncertainty may contribute to this ideological skew, with implications for auditing AI in political information dissemination. AI
IMPACT Highlights potential for ideological bias in LLM-generated political information, necessitating safeguards for AI-mediated news consumption.
RANK_REASON Research paper published on arXiv detailing a new measurement framework for LLM ideological drift. [lever_c_demoted from research: ic=1 ai=1.0]
- arXiv
- Directional Hallucinations: Ideological Drift in News-Grounded LLM Question Answering
- Hugging Face
- QBias
- U.S.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
