Prompting Intentionality and Authorship Preservation in AI-Assisted EFL Academic Writing: A Process-Tracing Inquiry

Abstract

Large language models (LLMs) are reshaping how researchers write and negotiate authorship. As these tools grow increasingly capable of producing fluent academic prose, a pressing question emerges: When writers use LLMs to help produce an academic text, how much of their thinking survives the exchange? Research has examined attitudes toward generative artificial intelligence (GenAI) and the quality of AI-assisted outputs; however, the behavioural mechanisms through which writers maintain or relinquish authorial control remain under-explored. This study introduces prompting intentionality and examines its association with authorship preservation. Sixteen EFL doctoral researchers completed a five-stage process-tracing protocol progressing from human-only ideation through naive, guided, and strategic prompting to final re-authoring. Data were collected from six sources, including prompt texts, conversation threads, authorship trace annotations, reflections, and a Likert-scale instrument, and analysed using a convergent mixed-methods design. Prompting intentionality rose across the three stages (mean 0.31 to 3.56; Friedman χ² = 25.20, p < .001) and was associated with authorship preservation (rs = .76, p < .001; mean Human Agency Preservation Ratio 79.4%). Rather than a uniform rise, criteria showed a clear ordering: Writers encoded personal stance and content readily but assigned the tool a bounded role least often. Strategic prompting was identified as the stage of strongest authorial control, though many did not revise AI output, indicating that agency was exercised more at encoding than at output selection. An emergent conceptual pattern, the Agentic Prompting Loop, is proposed to describe this negotiation of human agency. The findings underscore the importance of explicit prompting literacy instruction for responsible GenAI integration.