Agentic AI as a Dual-Role Lecturer and Teaching Assistant: Effects on Learner Autonomy, Self-Regulated Learning, and Intrinsic Motivation in University-Level EFL Education

Abstract

Agentic AI is entering university English-language education, but whether it supports or supplants learner autonomy remains empirically open. In a 16-week quasi-experiment with 120 B1-level university EFL students, we deployed a dual-role Agentic AI on the Uedu platform — an AI Lecturer that delivers structured input and Socratic questioning, and an AI Teaching Assistant that scaffolds through progressive hints (reflective prompts → conceptual cues → worked examples) with algorithmic fading. The experimental group showed large, simultaneous gains in learner autonomy, self-regulated learning, and intrinsic L2 motivation, alongside a visible shift in questioning from answer-seeking toward elaboration and metacognition.

Problem & Motivation

LLM-based Agentic AI is rapidly entering university English education, yet there is little evidence on whether it genuinely supports — rather than replaces — learner autonomy. Three gaps motivate this study: most AI-assisted language-learning research examines single-function tools rather than a dual-role design that is simultaneously instructor and scaffolding partner; Self-Determination Theory (SDT) and Vygotsky's Zone of Proximal Development (ZPD) are frequently invoked but rarely tested for whether “scaffold fading” can actually be realized in an LLM environment; and it remains unexamined whether Agentic AI can foster autonomous learning behavior within a Confucian-heritage cultural context where performance orientation and autonomy orientation coexist.

Method

A 16-week quasi-experiment (pretest–posttest non-equivalent control-group design) was conducted with 120 B1-level undergraduates at a comprehensive university in northern Taiwan (60 experimental, 60 control), using intact-class assignment, the same instructor, and identical syllabus and assessment criteria. The experimental group used a dual-role Agentic AI deployed on Uedu: an AI Lecturer providing structured English input and initiating Socratic questioning, and an AI Teaching Assistant responding through progressive hint scaffolding (reflective questions → conceptual cues → worked examples only as a last resort), with prompt-architecture-level “Socratic guardrails” preventing direct answer-giving; scaffolding faded algorithmically as weekly performance improved. The control group used the same platform with the Agentic AI features disabled. Learner autonomy (LAS), self-regulated learning strategies (MSLQ), and intrinsic motivation (L2MSS) were analyzed via ANCOVA with pretest scores as covariates; at the learner level, 6,235 AI-Teaching-Assistant dialogue turns were analyzed for shifts in questioning type. The study was approved by the National Taiwan University Research Ethics Committee (NTU-REC 202507EM058).

Findings

• Three outcomes rose together: the experimental group's gains in learner autonomy, SRL strategy use, and intrinsic L2 motivation significantly exceeded the control group's (partial η² = .247, .215, and .183 respectively — all large effects).
• Motivation dissociation: extrinsic goal orientation showed no between-group difference. This “intrinsic up, extrinsic unchanged” pattern aligns with SDT's distinction between autonomous and controlled motivation and is difficult to explain as a mere novelty effect (which would uniformly inflate all motivation indices).
• Visible shift in questioning: 90.0% (54/60) of experimental learners monotonically shifted from answer-seeking toward elaboration/metacognitive questioning across four time periods; the overall corpus moved from 71.3% answer-seeking in weeks 1–4 to 28.9% in weeks 13–16.
• Three qualitative themes emerged from 16 interviews: self-paced agency, adaptive feedback as ZPD scaffolding, and — unanticipated by the original SDT–ZPD frame — affective safety and relational warmth.

Implications

The study's value lies as much in its interpretive discipline as in its effect sizes — a stance central to the Uedu research program. The between-group contrast confounds AI access with task format (Socratic dialogue vs. open-ended writing), so the observed effects reflect an AI-supported learning situation as a whole rather than the isolated action of the AI system. The 90% questioning shift is a descriptive behavioral trajectory, not evidence of internalization, and could arise from interface familiarity or task-difficulty differences across periods. Most importantly, behavior while using the technology is not the same as behavior after its removal: the authors name this the “autonomy–accessibility paradox” — the very features that make AI support effective (immediate availability, adaptive calibration, frictionlessness) may simultaneously erode the beneficial cognitive struggle needed to internalize self-regulation. Models of AI-supported autonomy should therefore incorporate a scaffold-removal temporal dimension, and future work should add delayed, scaffold-free post-tests and treat a “progressive AI-removal protocol” as a first-class pedagogical question alongside in-class scaffold fading.

Citation

@article{li2026agentic_ai_efl,
  author  = {Kuei-Hao Li and Chia-Kai Chang},
  title   = {Agentic {AI} as a Dual-Role Lecturer and Teaching Assistant: Effects on Learner Autonomy, Self-Regulated Learning, and Intrinsic Motivation in University-Level {EFL} Education},
  journal = {Frontiers in Psychology},
  volume  = {17},
  pages   = {1847607},
  year    = {2026},
  doi     = {10.3389/fpsyg.2026.1847607},
}