Can an LLM Be Used for Peer Review? Educational Material for Journal Publication Committee Member...

English summary

The video examines the feasibility of using large language models in academic peer review, patterns of AI-assisted reviewing, reported effects on review outcomes, and journal policies for generative AI. In the studies presented, overlap between GPT-4 and human reviews was approximately 30 percent for Nature family journals and 40 percent for ICLR, falling to approximately 1 percent when reviews were mismatched across papers. This suggests that GPT-4 critiques reflected paper-specific content to some extent. AI-assisted reviews were also reported to be associated with higher review scores and a higher acceptance probability for borderline submissions at one conference. The presentation concludes that policy should prioritize the accuracy of the final content, human oversight, and confidentiality rather than attempting to prohibit or police AI use in general.

Key points

• In the analysis by Liang and colleagues, overlap between GPT-4 and human reviews was reported to be similar to human-to-human overlap and was highest for papers with consensual rejection decisions. Human-to-human and GPT-4-to-human agreement was lower for highly scored or accepted papers.
• GPT-4 more often identified criticisms raised by multiple human reviewers and comments ranked as highly important in human reviews. It less consistently addressed missing citations and literature context, machine learning ablation experiments, and the assessment of true scientific novelty.
• Among 308 authors in the prospective survey, 57.4 percent rated GPT-4 feedback as helpful or extremely helpful, and 58.5 percent said they would use the system again. These findings were based on the author survey described in the presentation.

Audience

Journal editors, peer reviewers, researchers and authors, editorial boards, and organizations developing generative AI policies

Scope and limitations

This summary is based only on the supplied evidence notes from the video presentation. Its main evidence comes from studies reported on arXiv and from a policy example at one journal; the video is not itself an independent systematic review. The presentation notes that AI detection tools have both false positive and false negative errors, with potentially higher false positive rates for non-native English writers, so sanctions should not rely solely on detector scores. In addition, associations between AI-assisted reviews and review scores or acceptance outcomes do not establish causation.

Research context

The topic has a neutral thematic connection to Sangzin Ahn's portfolio through medical AI and data science, particularly generative AI evaluation, human oversight, reproducibility, data confidentiality, and research governance.