Study on Bias in Large Language Model Peer Review: A Systematic Evaluation of Academic Prestige and Racial Factors

This article introduces the llm-peer-review project initiated by the oamin-ai team. Through controlled variable experimental methods, it systematically evaluates the bias tendencies of large language models (LLMs) in academic peer review tasks regarding authors' academic prestige, racial background, and other dimensions. The aim is to explore whether AI inherits or exacerbates structural biases in human society, providing empirical references for research on academic ethics and AI fairness.

With the widespread application of large language models (LLMs) in academia, more and more researchers are exploring their potential to assist in peer review. Traditional peer review has already faced fairness concerns, such as systematic biases against authors from specific institutions/regions. When LLMs are introduced into this field, it is crucial to understand whether they inherit or exacerbate such biases, which led to the initiation of the llm-peer-review project.

The llm-peer-review project focuses on evaluating the bias performance of LLMs in peer review. The core hypothesis is: if the model training data contains unbalanced information related to authors' backgrounds, the model may exhibit systematic biases in reviews (e.g., more positive evaluations of authors from well-known institutions). The project uses controlled variable experiments to quantify the model's sensitivity to the prestige of academic institutions, authors' racial backgrounds, and regional economic levels.

The project uses rigorous controlled variable experiments: creating different variant versions of the same paper, with the only difference being author information (institution prestige level, name characteristics implying race, regional economic level); comparing the review results of different variants to isolate bias-influencing factors and quantify their severity. The experiments cover three dimensions: academic prestige bias, racial bias, and income bias.

The project repository uses a modular design: the data directory contains processed paper data and metadata, and the experiment directory is classified by bias type; this design facilitates other researchers to reproduce results and verify the universality of biases. The project is open-sourced under the MIT License, reflecting the commitment to promoting academic transparency and reproducibility.

If significant biases are confirmed, it will have far-reaching impacts: reminding the academic community to use AI-assisted review tools cautiously (needing to develop bias detection and mitigation mechanisms); pushing LLM developers to pay attention to the diversity and balance of training data; providing an empirical foundation for the field of AI ethics, indicating that advanced AI systems may inherit structural inequalities in human society.

Technological progress should not come at the cost of fairness. When LLMs are applied to high-risk scenarios such as peer review, comprehensive bias assessment is necessary. Future directions include: developing debiasing algorithms, establishing industry standards for model review, and creating transparent and interpretable AI review systems to ensure that technology serves academic progress rather than reinforcing existing unequal structures.

The llm-peer-review project is an important step in LLM ethics research. Through rigorous experiments, it reveals the potential bias issues of AI systems and provides the academic community with an opportunity for reflection. As AI penetrates deeper into the field of scientific research, such studies will become a key cornerstone for ensuring the responsible application of technology.