Context
Jinadu and colleagues tested eight LLMs in a controlled, single-turn academic paper-recommendation setting. Keeping paper titles and abstracts constant while varying authority metadata, they found model-dependent, directional changes in top-one paper choices.
What changed
The peer-reviewed study compares original, flipped and boosted author prestige, venue and citation metadata across five open-weight and three closed-weight models. It reports that prompt-level debiasing reduces explicit authority references more readily than it removes authority-driven recommendation flips.
Why it matters
AI source selection can depend on metadata beyond semantic relevance. Controlled evidence of authority bias is relevant to evaluating whether recommendations reflect document content or perceived prestige.
Visibility OS interpretation
Visibility OS interpretation: distinguish content relevance from authority signals in source-selection experiments, and test actual recommendation choices rather than relying only on generated explanations. The findings do not establish how deployed AI search engines rank web pages.
What remains uncertain
How much of this authority bias persists in retrieval-grounded multi-turn search? Does it affect website citations or product recommendations outside academic papers? Which platform and prompt settings moderate the effect?