This paper addresses the issue of detecting AI-generated student essays, a critical concern given the risks associated with plagiarism and the dissemination of fake news. The authors construct the AIG-ASAP dataset, which contains AI-generated student essays using various text perturbation methods, including word and sentence substitution, to evade detection. Through empirical experiments, they assess the performance of current AI-generated content (AIGC) detectors on the AIG-ASAP dataset. The results show that existing detectors can be easily circumvented using straightforward automatic adversarial attacks, particularly through word and sentence substitution methods. These methods effectively evade detection while maintaining the quality of the generated essays, highlighting the need for more accurate and robust detection methods tailored to the challenges posed by AI-generated student essays. The paper also introduces the AIG-ASAP dataset as a benchmark for evaluating AIGC detectors in the education domain and provides detailed experimental setups and results, including human evaluation and analysis of detection difficulty across different essay types. The findings underscore the vulnerabilities of current detection methods and the urgent need for improved detection techniques.This paper addresses the issue of detecting AI-generated student essays, a critical concern given the risks associated with plagiarism and the dissemination of fake news. The authors construct the AIG-ASAP dataset, which contains AI-generated student essays using various text perturbation methods, including word and sentence substitution, to evade detection. Through empirical experiments, they assess the performance of current AI-generated content (AIGC) detectors on the AIG-ASAP dataset. The results show that existing detectors can be easily circumvented using straightforward automatic adversarial attacks, particularly through word and sentence substitution methods. These methods effectively evade detection while maintaining the quality of the generated essays, highlighting the need for more accurate and robust detection methods tailored to the challenges posed by AI-generated student essays. The paper also introduces the AIG-ASAP dataset as a benchmark for evaluating AIGC detectors in the education domain and provides detailed experimental setups and results, including human evaluation and analysis of detection difficulty across different essay types. The findings underscore the vulnerabilities of current detection methods and the urgent need for improved detection techniques.