Conference Program

As generative AI becomes embedded in educational practice, a critical question emerges: does it cultivate or erode the metacognitive capacities that underpin democratic, agentive learning? The dominant model of human-AI interaction — in which students ask and AI answers — mirrors cognitive offloading rather than pedagogical dialogue, externalising problem-framing and bypassing the productive struggle through which deep learning occurs. This contribution presents empirical evidence from a quasi-experimental study examining whether a theoretically grounded alternative can foster metacognitive awareness among pre-service teachers, rather than merely replacing their cognitive effort.

The study deployed MAIEUTIC — a conversational AI explicitly architected around Vygotsky's Zone of Proximal Development and Zimmerman's Self-Regulated Learning framework — in an undergraduate Educational Technology course at an Italian university (N=188). Grounded in Bakhtinian dialogism, MAIEUTIC treats the learner not as a passive recipient of information but as an active voice in a genuine epistemic exchange. Rather than answering student questions, it inverts the typical human-AI dialogue: through Socratic questioning and adaptive scaffolding organised across a tri-phasic structure (planning → producing → revising), the system elicits student elaboration at each stage. Crucially, the system incorporates anti-offloading mechanisms — refusing to generate complete answers and requiring elaboration before offering further support — operationalising the idea that the AI's role is to sustain thinking, not to substitute it. Across 25 analysed sessions, MAIEUTIC averaged 6.5 questions per session against students' 1.6, yielding a 4:1 system-to-student ratio — what we term the Interaction Flip. Metacognitive awareness was measured via the MAI-19 (Metacognitive Awareness Inventory, 19 items, 5-point Likert scale) in a matched pre-post design (N=144: MAIEUTIC n=71, ChatGPT control n=73), with participant-generated anonymous codes ensuring privacy while enabling longitudinal tracking.

Results indicate markedly divergent trajectories. MAIEUTIC participants showed substantial growth in self-reported metacognitive awareness (ΔM = +0.625), while the ChatGPT control group showed a slight decline (ΔM = −0.068). A mixed ANOVA confirmed a large Group × Time interaction (F(1,142) = 146.31, p < .001, η²p = .507), which remained robust after adjusting for baseline non-equivalence (ANCOVA, η²p = .404). Nearly half of MAIEUTIC participants (49.3%) met the Reliable Change Index threshold for clinically meaningful improvement; none in the control group did.

These findings suggest that AI dialogue structure — specifically, who asks the questions — is not a neutral design choice but a deeply political and pedagogical one. In educational spaces where democratic deliberation should flourish, designing AI as an interlocutor that questions rather than answers may be a concrete strategy for preserving learner agency, supporting teacher professional identity, and resisting the automation of judgment. The study contributes empirical grounding to theoretical calls for critical AI literacy, and we conclude with implications for educators, AI developers, and policy frameworks that currently treat all AI tools as equivalent.