Abstract
<p>The widespread use of large language models (LLMs) is reshaping assessment practices, making it increasingly difficult to determine what learners actually know and can do, as learning processes become less visible in AI-mediated work. Rather than treating AI use as a problem of detection, this study explores how assessment can be reconfigured to make learning visible through learners’ interaction with AI. Focusing on a postgraduate teacher education course, the study introduces Meta-Task Awareness (MTA) as an analytic lens for examining how learners regulate task goals, pedagogical considerations, and epistemic responsibility in LLM-supported lesson design. Using an exploratory qualitative design with descriptive code-and-count summaries, interaction trace data from 11 students were analysed at the episode level using a theory-informed coding framework. Findings show that students’ engagement was primarily oriented toward task clarification, pedagogical reasoning, and evaluative judgment, with varying patterns resulting in three distinct orientations toward AI-mediated design. The study shows how assessment can be reconfigured so that learning becomes more visible, more accountable, and increasingly unavoidable in the age of AI.</p>