From Cognitive Delegation to Epistemic Agency: The Role of “I’m Not Sure” in Human–LLM Interaction

Sammanfattning

As large language models (LLMs) increasingly permeate educational settings, they are predominantly designed as authoritative answer-giving systems or instructional mentors. While recent pedagogical interventions aim to scaffold learning, they often implicitly reinforce AI epistemic authority, leaving their deeper impact on learners’ epistemic agency underexplored. In this mixed-methods study, we compare learner interactions with an authoritative expert-like AI and a non-authoritative peer-like AI to examine how AI epistemic stance shapes the enactment of epistemic agency. Our findings reveal a tension between performance and epistemic engagement. The authoritative AI condition supports higher immediate task performance but is associated with cognitive delegation, where learners rely heavily on system-generated answers and show weaker knowledge transformation. In contrast, the peer-like AI introduces epistemic friction through partial and uncertain guidance, which increases evaluative and productive behaviors, though it also presents challenges for independent knowledge consolidation. Quantitative results show a positive correlation between perceived and observed production behaviors (r = .33), suggesting stronger alignment between metacognitive awareness and epistemic activity in the peer-like condition. Qualitative findings further indicate that this engagement is shaped by broader performative assessment cultures that prioritize efficiency and correctness. Overall, the findings suggest that different AI epistemic stances redistribute cognitive labor between system and learner, shaping both epistemic agency and performance outcomes. We argue for a shift in educational AI design toward adaptive “epistemic peers” that dynamically balance authoritative scaffolding with opportunities for learner-driven knowledge consolidation.