Scholarlink

<p>This study examines how generative mediation reshapes the temporal structure of L2 digital writing. Using keystroke logging (Inputlog), we analysed pausing behaviour, revision activity, and writing efficiency among 27 undergraduate EFL students composing an English text with access to ChatGPT. Writing efficiency was defined as a composite of active writing time, word production, task duration, and revision count. Findings show that longer pauses, particularly at lexical boundaries, predicted lower efficiency, while pause frequency had limited explanatory value. Interaction time with ChatGPT was associated with slower transcription but higher efficiency, suggesting that tool consultation may interrupt surface fluency while supporting broader task regulation. No relationship emerged between ChatGPT use and revision frequency, indicating a division of labour in which external mediation affects formulation, whereas revision remains under learner control. Bayesian models revealed that total pause count and mean pause duration predicted interaction time, reflecting individual strategies for integrating external suggestions. L2 proficiency did not significantly predict writing efficiency or tool use. These results suggest that behavioural indicators such as pauses must be interpreted within the sequencing of tool-mediated writing episodes. The study contributes to process-oriented research on L2 writing and highlights the need to reconsider fluency and efficiency under conditions of generative support.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

ChatGPT as Cognitive Mediator: Tracing Pausing, Revision, and Efficiency in L2 Digital Writing

  • 期刊论文

  • 发表日期：

    2026年1月27日

查看全文
  • Aitor Garcés-Manzanera,
  • Sophie McBride

摘要

This study examines how generative mediation reshapes the temporal structure of L2 digital writing. Using keystroke logging (Inputlog), we analysed pausing behaviour, revision activity, and writing efficiency among 27 undergraduate EFL students composing an English text with access to ChatGPT. Writing efficiency was defined as a composite of active writing time, word production, task duration, and revision count. Findings show that longer pauses, particularly at lexical boundaries, predicted lower efficiency, while pause frequency had limited explanatory value. Interaction time with ChatGPT was associated with slower transcription but higher efficiency, suggesting that tool consultation may interrupt surface fluency while supporting broader task regulation. No relationship emerged between ChatGPT use and revision frequency, indicating a division of labour in which external mediation affects formulation, whereas revision remains under learner control. Bayesian models revealed that total pause count and mean pause duration predicted interaction time, reflecting individual strategies for integrating external suggestions. L2 proficiency did not significantly predict writing efficiency or tool use. These results suggest that behavioural indicators such as pauses must be interpreted within the sequencing of tool-mediated writing episodes. The study contributes to process-oriented research on L2 writing and highlights the need to reconsider fluency and efficiency under conditions of generative support.