Optimising the Human-AI Loop: Prompting Strategies and Output Verification Practices of ESL Undergraduates in Academic Writing

Abstract

As generative artificial intelligence (GenAI) becomes deeply embedded in higher education, understanding the dynamics of the human-AI relationship is critical to ensuring that algorithmic support enhances, rather than diminishes, critical student agency. While considerable research examines automated model capabilities, empirical studies exploring how students actively navigate the collaborative human-in-the-loop workflow remain scarce. This study investigates the prompting strategies and output verification practices employed by English as a Second Language (ESL) undergraduates when integrating large language models (LLMs) into their academic writing workflows. Grounded in an interpretive paradigm, qualitative data from semi-structured focus groups with 23 participants were analysed using thematic analysis. The findings reveal that while ESL undergraduates leverage natural language interfaces for cognitive scaffolding and structural refinement, their prompt development relies primarily on iterative trial-and-error rather than structured technical inputs. Furthermore, output verification remains asymmetrical, exposing critical vulnerabilities to automation bias alongside practices of rigorous cross-checking. By mapping these empirical, user-centric practices, this paper bridges the gap between educational informatics and human-system interaction. Ultimately, these insights provide valuable behavioural data essential for the design of intelligent user interfaces (IUIs) and adaptive human-in-the-loop writing assistants that preserve and support human agency.