Machine learning models for question answering tasks are known to be vulnerable to adversarial examples. The AddSent algorithm generates such examples by appending misleading sentences, but often requires manual corrections due to grammatical errors and unnatural phrasing. In this paper, we propose a novel adversarial example generation method using large language model (LLM), in which all steps of the AddSent algorithm are replaced with LLM-based generation to eliminate the need for human intervention. Experiments on English datasets show a comparable degradation in F1 scores to that caused by AddSent. In contrast, the impact on Japanese datasets is notably smaller, highlighting both the effectiveness of the proposed method and the challenges of applying it to multilingual settings.