When preparing research papers, the use of generative AI as a writing aid is increasing. In fact, the growth rate of global paper production has risen significantly alongside the widespread adoption of generative AI. On the other hand, generative AI has also begun to be used for drafting research proposals. Combined with the deepening of individual research fields and an increase in interdisciplinary studies, the number of research grant applications is also on the rise. While many funding agencies prohibit the use of generative AI in reviews, questions remain as to whether there are ways to utilize AI to reduce the burden on reviewers.
The Ministry of Education, Culture, Sports, Science and Technology (MEXT) has decided to allocate 25 million yen in Grants-in-Aid for Scientific Research (Grant-in-Aid for Special Purposes) to a research group led by Professor Akiko Aizawa of the National Institute of Informatics for an investigative research project aimed at developing an AI agent to support reviews. The project will run until the end of next fiscal year.
The number of published papers is increasing globally, and this growth rate has accelerated since the introduction of ChatGPT in 2022. Accompanying this trend is the expansion of hallucinations. For example, among papers accepted at international conferences in the field of natural language processing, errors in references have been increasing year by year, as has the number of errors per paper.
Regarding AI use in paper preparation, many publishers and academic societies have established policies such as "explicitly stating when generative AI is used" and "prohibiting generative AI from being listed as an author." The underlying idea is that while utilizing AI for tasks like text proofreading is acceptable, authors must not have generative AI write the manuscript itself. Major publishers use generative AI for checking research misconduct and fact-checking references.
Whether AI is used during peer review is often left to the authors' discretion, but a survey of 1,600 researchers revealed that more than half have experience using it to assist in writing review results, summarizing, or checking for plagiarism. Rather than being used by reviewers, AI paper review is increasingly utilized by authors to review their work prior to submission, thereby improving the quality of the paper. However, peer review by general-purpose AI has not yielded very good results, and the development and testing of dedicated systems is currently in its initial stages.
The use of AI in research grant reviews is handled even more cautiously than in paper peer review. AI has been used for paper review, but its use for budget applications will take place in the future. This raises concerns regarding data leaks caused by uploading proposals to external generative AI. There is an overwhelming lack of open data to serve as a foundation for AI research and development.
Papers and research grants differ in terms of field diversity, purposes, and evaluation criteria. AI systems developed for paper peer review cannot be applied as-is for research grant reviews. Expected initial uses of AI in research grant reviews include reviewer matching, formatting review text, providing feedback on review comments, and checking or pointing out improvements for proposals prior to submission.
In fact, the Swiss National Science Foundation (SNSF), the La Caixa Foundation in Spain, and the Dutch Research Council (NWO) utilize AI for reviewer matching. The La Caixa Foundation has reportedly introduced an AI solution into its pre-review system as a tool to identify non-compliant proposals. The UK Meta-Science Unit is conducting an LLM project to support research grant reviews.
Conversely, funding agencies (FAs) such as the Japan Society for the Promotion of Science (JSPS) and the U.S. National Science Foundation (NSF) prohibit the direct use of generative AI in reviews due to risks such as data leaks.
In response to the increasing number of applications and to reduce the burden on reviewers, this project will, under the prerequisite of being human-centric and highly reliable, organize insights to contribute to examining how review support using generative AI should be structured and conduct investigative research toward its realization.
Specifically, using the KAKENHI review system as a use case, the project will carry out the conceptual design and partial verification of an AI agent that performs the following tasks: assisting in reviewer selection (taking into account the balance of age, gender, region, and institution type for each review section based on research content, etc.), assisting in integrating review opinions (a function to integrate and summarize multiple review opinions to feed back the reviewers' comments to applicants), and assisting in academic evaluation (premised on the introduction of a generative AI where sensitive information, such as applicants' research ideas, can be entered and searched by building a secure domestic infrastructure).
With its future integration into the KAKENHI system in mind, this project will provide reference materials for system design and operational considerations at MEXT and JSPS. Specifically, within the research period, the group will organize technical feasibility, system compatibility, and operational issues, and feedback insights to contribute to future decision-making.
This article has been translated by JST with permission from The Science News Ltd. (https://sci-news.co.jp/). Unauthorized reproduction of the article and photographs is prohibited.