This research paper investigates the ability of current AI text detection methods to identify peer reviews generated by large language models (LLMs). The authors highlight the growing concern of LLMs being used unethically to produce peer reviews, potentially jeopardizing the integrity of scientific publications.
Research Objective: The study aims to evaluate the effectiveness of various AI text detection methods in distinguishing between human-written and LLM-generated peer reviews.
Methodology: The researchers collected real peer reviews from the ICLR conference (2019-2024) and generated synthetic reviews using GPT-4o and Llama-3.1 (70b) models for the same papers. They then tested the performance of open-source and proprietary AI text detection models, including Roberta, Longformer, Originality AI API, and LLM-as-a-judge approaches. Additionally, they proposed a novel method using "Anchor Reviews" for comparison.
Key Findings: The study found that existing AI text detection methods have limitations in accurately identifying AI-generated reviews while maintaining a low false-positive rate. While some methods achieved high detection rates, they often misclassified human-written reviews as AI-generated. Notably, the proposed "Anchor Review" method, based on semantic similarity comparison, demonstrated promising results in detecting GPT-4o generated reviews with a lower false-positive rate.
Main Conclusions: The research concludes that current AI text detection methods are insufficient for reliably identifying AI-generated peer reviews. The authors emphasize the urgent need for more robust and accurate detection tools to address this ethical challenge posed by LLMs in scientific publishing. They suggest further research into methods like the "Anchor Review" approach to improve detection accuracy and safeguard the integrity of the peer review process.
Significance: This study highlights a critical issue at the intersection of AI and academic integrity. As LLMs become increasingly sophisticated, their potential misuse in generating undetectable, artificial content poses a significant threat to the reliability and trustworthiness of scientific publications.
Limitations and Future Research: The study primarily focused on peer reviews from a single computer science conference (ICLR). Future research should explore the generalizability of these findings across other disciplines and publication venues. Additionally, investigating the effectiveness of combining multiple detection methods and developing new approaches tailored for peer review contexts are crucial next steps.
To Another Language
from source content
arxiv.org
ข้อมูลเชิงลึกที่สำคัญจาก
by Sungduk Yu, ... ที่ arxiv.org 10-07-2024
https://arxiv.org/pdf/2410.03019.pdfสอบถามเพิ่มเติม