The paper provides an adversarial evaluation of AI-generated content (AIGC) detectors’ effectiveness in identifying AI-written student essays. The study constructs the AIG-ASAP dataset, a collection of AI-written student essays, using various text perturbation methods. The results reveal that current detectors can be easily tricked using simple automatic adversarial attacks. The study emphasizes the urgent need for more precise and robust methods to detect AI-generated student essays in the education domain.
Publication date: 2 Feb 2024
Project Page: https://github.com/xinlinpeng/AIG-ASAP
Paper: https://arxiv.org/pdf/2402.00412
