Prestige over merit: An adapted audit of LLM bias in peer review

Research indicates that Large Language Models (LLMs) used in academic peer review reproduce human biases, particularly favoring perceived author prestige over manuscript quality. A multi-role LLM simulation of editors and reviewers revealed that disclosing author identities significantly reduced rejection recommendations, with institutional prestige being the primary influencing factor.

Why it matters

This research highlights a significant risk in the increasing adoption of AI tools within critical evaluation processes, demonstrating that current LLM implementations can perpetuate systemic biases. This has implications for fairness, equity, and the integrity of outcomes in any domain relying on such evaluations, potentially undermining merit-based systems.

Key insights

• LLMs, despite their growing informal role in scholarly peer review, are shown to reproduce biases similar to those observed in human decision-making.
• An adapted resume-style audit, simulating editor/reviewer roles, evaluated high-quality manuscripts across diverse scientific domains.
• Randomizing author identities (institutional prestige, gender, race) allowed for the isolation of bias factors.
• Revealing author identities reduced reviewer rejection recommendations by approximately 25% of the mean rejection rate, even for identical content.
• Institutional prestige was identified as the dominant cue influencing review outcomes, overriding manuscript quality.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2509.15122