Abstract

Agent-driven instructional models are spreading rapidly in language education, yet the question of for whom such models work—and for whom they carry risks—has received little attention. Using the cognitive offloading-human-AI co-creation continuum as an analytic lens, this study compared two contrasting learner populations under the same agent-driven Business English project-based learning (PBL) model: first-year undergraduates (n = 30) and MBA students (n = 30). The two cohorts were taught by the same instructor with the same textbook and identical agent roles and interaction rules, and completed parallel versions of the same end-of-semester questionnaire (with minor wording adaptations for undergraduates), supplemented by open-ended responses. Analyses combined between-group comparisons, a seven-dimension profile comparison, an exploratory typological cluster analysis of offloading and co-creation, and hierarchical regression. Results showed that the two populations were strikingly similar in offloading containment and across all outcome dimensions: no between-group difference survived Holm correction (|d| <= 0.52), and the only medium-sized item-level difference was confidence in detecting errors in agent output (higher among MBA students, d = 0.67). The typological analysis indicated that the sample divided into a “deeply engaged” type (56.7%) and a “prudent self-reliant” type (43.3%) that cut across both cohorts, with high co-creation and elevated dependence signals co-occurring and the type distribution independent of cohort (p = .193). After controlling for gender, prior generative-AI experience, and use frequency, the incremental explanatory power of cohort approached zero (dR-squared <= .021). The findings suggest that under one and the same rule-governed design, “level is not destiny”: learner level should remain an explicit consideration in the design and scaling of agent-based pedagogy, but the unit of differentiation should descend to learner types, and differentiated design guidelines are derived accordingly.

This work is licensed under a Creative Commons Attribution 4.0 License.