Hidding the Ghostwriters: An Adversarial Evaluation of AI-Generated Student Essay Detection
Project Overview
The document examines the impact of generative AI, particularly large language models (LLMs), on education, focusing on their application in student essay writing. While these models can produce high-quality text, they pose significant challenges, including the potential for plagiarism and the deterioration of fundamental writing skills among students. To address these issues, the authors present the AIG-ASAP dataset, designed to assess the efficacy of AI-generated content (AIGC) detection methods against adversarial attacks related to student essays. The findings highlight vulnerabilities in current detection techniques, underscoring a pressing need for enhanced methods to accurately identify AI-generated work. Overall, the document calls for a careful consideration of the integration of generative AI in educational settings, balancing its advantages with the risks it introduces.
Key Applications
AIG-ASAP dataset for AI-generated essay detection
Context: Educational context focused on student essay writing and assessment.
Implementation: The AIG-ASAP dataset was constructed using various text perturbation methods to evaluate the effectiveness of AIGC detectors.
Outcomes: The study found that existing detection methods could be easily circumvented by simple adversarial attacks, indicating a need for improved detection capabilities.
Challenges: Current AIGC detectors struggle with high rates of false negatives when faced with adversarially perturbed text.
Implementation Barriers
Technical Barrier
Existing AIGC detection methods are vulnerable to simple adversarial attacks, allowing AI-generated essays to evade detection.
Proposed Solutions: Develop more robust AIGC detection methods that can withstand adversarial perturbations.
Educational Barrier
Students may rely on LLMs for essay writing, leading to a lack of development in essential writing skills.
Proposed Solutions: Implement educational policies that promote skill development while integrating AI tools in a balanced manner.
Project Team
Xinlin Peng
Researcher
Ying Zhou
Researcher
Ben He
Researcher
Le Sun
Researcher
Yingfei Sun
Researcher
Contact Information
For information about the paper, please contact the authors.
Authors: Xinlin Peng, Ying Zhou, Ben He, Le Sun, Yingfei Sun
Source Publication: View Original PaperLink opens in a new window
Project Contact: Dr. Jianhua Yang
LLM Model Version: gpt-4o-mini-2024-07-18
Analysis Provider: Openai