Abstract To support learner agency in LLM‐powered writing environments, this research introduces a new interface that integrates two key features: an Argument Outline and a Similarity Viewer . These features were designed to enhance intentionality in writing and self‐monitoring of reliance on AI‐generated suggestions. Implemented in CoAuthor , a GPT‐based writing tool, the interface was tested in an experiment with 52 university students who completed an argumentative writing task under two conditions: (1) SV group ( n = 26) with Similarity Viewer only, and (2) SV‐AO group ( n = 26) with both Similarity Viewer and Argument Outline. Writing behaviours such as seeking, dismissing, accepting and modifying AI suggestions were analysed using statistical methods and epistemic network analysis (ENA). The average semantic similarity was calculated to measure the similarity between final texts and AI‐generated content. Follow‐up interviews were conducted to explore participants' perceptions of the two features. Results showed that the SV‐AO group had stronger centrality in behaviours linked to independent text composition and produced final texts with lower semantic similarity. Interview data further indicated that participants viewed the Argument Outline as a useful scaffold for supporting independent writing while enabling the selective integration of AI‐generated suggestions aligned with their writing goals. Overall, this study highlights the potential of interface design in optimizing the balance between learner agency and AI assistance in human–AI co‐writing.
Kim et al. (Sat,) studied this question.