Recently, generative artificial intelligence (AI)-powered chatbots such as ChatGPT and Bing Chat have garnered increasing attention on a global scale. Previous studies have focused mostly on the influence of generative AI on writing while few researchers have investigated how generative AI can facilitate students' multimodal writing process. To fill in this gap, we explored the generative AI-assisted composing processes of two groups of English as a foreign language (EFL) writers over two weeks in this qualitative study. One group completed a multimodal PowerPoint (PPT) project, and the other group completed a traditional argumentative essay project. Our data consist of students’ screen recordings with think-aloud protocols, final multimodal texts, and post-project interviews. Our analysis showed different patterns in text production across the two groups. Students in the PPT group tended to construct more bridge texts and examples to corroborate their sub-claims in the hierarchical order. They also inclined to borrow the summarized search results from the Bing Chat to expand texts for their PPT slides. With regard to image generation for PPT slides, descriptions of AI images from ChatGPT were used as effective prompts to generate AI images from Bing Image Creator. Moreover, students were interested in producing and refining AI images following the recommended prompts by Bing Chat. They also evaluated these AI images from different perspectives. We conclude the study with a discussion on the pedagogical implications and suggestions for further study.

最近，ChatGPT 和 Bing Chat 等基于生成人工智能 (AI) 的聊天机器人在全球范围内引起了越来越多的关注。以往的研究主要集中在生成式人工智能对写作的影响，而很少有研究人员研究生成式人工智能如何促进学生的多模式写作过程。为了填补这一空白，我们在这项定性研究中用了两周的时间探索了两组英语作为外语 (EFL) 作家的生成式人工智能辅助写作过程。一组完成多模式 PowerPoint (PPT) 项目，另一组完成传统的议论文项目。我们的数据包括学生的有声思考协议的屏幕录像、最终的多模式文本和项目后访谈。我们的分析显示了两组的文本生成模式不同。PPT组的学生倾向于构造更多的桥梁文本和例子来按层次顺序证实他们的从属主张。他们还倾向于借用 Bing Chat 的汇总搜索结果来扩展 PPT 幻灯片的文本。在PPT幻灯片的图像生成方面，利用ChatGPT中的AI图像描述作为有效提示，从Bing Image Creator生成AI图像。此外，学生们对按照 Bing Chat 的推荐提示制作和完善 AI 图像很感兴趣。他们还从不同的角度评估了这些人工智能图像。我们通过讨论教学意义和进一步研究的建议来结束本研究。