Study finds LLMs exhibit substantial authority bias in academic paper recommendations

A new study published on arXiv investigates the phenomenon of "authority bias" in large language models (LLMs) when used for academic paper recommendations. Researchers found that LLMs tend to favor papers based on author prestige, venue, and citation counts over the actual content. This bias was observed to be substantial, varied across different LLMs, and only partially mitigated by prompt-level debiasing techniques. The study also highlighted a "say-do gap," where debiasing instructions were more effective at suppressing mentions of authority than at correcting the underlying bias in recommendations, suggesting that behavioral bias is underestimated by surface-level auditing. AI

IMPACT Highlights potential for LLMs to perpetuate existing academic inequities, necessitating careful auditing and debiasing.

RANK_REASON Academic paper detailing a new finding about LLM behavior. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• alphaXiv
• arXivLabs
• authority bias
• CatalyzeX Code Finder for Papers
• Connected Papers
• CORE Recommender
• DagsHub
• Gotit.pub
• Hugging Face
• Influence Flower
• large-language models
• Litmaps
• ScienceCast
• scite Smart Citations

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →