Authority Bias in Conversational Search Engines for Academic Paper Recommendation

When you ask an AI chatbot to recommend academic papers, does it judge the paper itself or does it just favor famous authors and top journals? This paper tests that directly by giving eight different LLMs the exact same paper content but changing only the author names, venue, and citation counts. Every model showed a strong, systematic bias toward high-prestige signals, and asking the models to ignore prestige only partially worked: they got better at hiding the bias in their explanations while still acting on it.