AI Peer: Large language models and academic peer review outcomes

Context: High profile successes with Large Language Models (LLMs) and ChatGPT, such as performing well at many natural language processing tasks and passing some student assessments, suggest that they may be powerful enough to support many other text reviewing applications, despite recognised problems including hallucination. A key text review task is academic peer review, which is used for journal quality control and national research evaluations, as well as supporting funding, promotion, and tenure decisions. Although aspects of peer review systems (e.g., reviewer finding) are frequently automated or at least semi-automated, and research shows that AI can assist with human reviewers’ writing, it is not clear whether LLMs can help decide peer review outcome decisions and overall recommendations. Topic: This project is about the potential role of LLMs in the academic peer review decision process across different disciplines. It will investigate whether LLMs have the capabilities to support (e.g., by providing non-random judgements) or threaten (by providing plausible but meaningless reports) different forms of academic peer review judgement. Aims and objectives: This project will systematically assess the extent to which LLMs can review academic work and score it for quality, focusing on expert review for the UK Research Excellence Framework (REF) national research evaluation and academic journal peer review. It will assess the ability of ChatGPT to score journal articles for research quality and to write academic reviews to support these judgements. Two additional techniques will also be applied to improve ChatGPT’s prediction accuracy. Although the project uses artificial intelligence (AI) it does not develop it but applies existing AI to academic peer review, to understand and evaluate it. The project will also build community and provide a forum for academic peer review AI advice by creating a Generative AI in Scholarly Review International Advisory Committee (GAISRIAC). This will connect relevant researchers and other stakeholders, provide a common information point (a website that will host project resources, GAISRIAC outputs and links to other relevant resources), and create workshops and panel events at conferences. Applications and benefits: By comparing LLM-constructed academic peer review reports with human peer review reports, this project will gain insights into the complex and often opaque process of constructing peer review, such as the extent to which reports can be written by combining author-reported limitations with general world information and the nature of the additional insights brought by field specialists. Most of all, however, the project’s findings about the value of ChatGPT to support research quality assessments will allow publishers, editors, the REF team, research managers (e.g., interviewing candidates) and funders to make more informed decisions about whether to exploit ChatGPT for efficiency, completely ban it for plausible but misleading reports, or something in between. Finally, the insights will support the ability to detect reviews written by ChatGPT and other LLMs, which will help to safeguard the academic peer review process. The transfer of these insights into practice will be advised on and sped by the high profile influential steering group and the GAISRIAC initiative.