Beyond a single role: Justifying a role-adaptive framework for AI companions through a comparative study in elementary book talk

Synthesis: Beyond a single role. A formative within-subjects study by Chang-Yen Liao (2026) comparing how 19 elementary students in Taiwan talk about books with a single-role AI companion (fixed "student peer") versus an experienced human teacher. The AI sustained significantly longer interactions but did so at the cost of student Learner Agency — students contributed a markedly lower proportion of words and sentences (Grade 5 less than a third of teacher-led sessions). The AI was effective at eliciting factual recall ("Facts" in the 4F framework) but significantly weaker than the teacher at prompting emotional ("Feelings") and future-oriented ("Future") reflection — an "affective ceiling." These gaps empirically justify a Role-Adaptive AI Companion Framework: a Modular Adaptive Agent, inspired by multi-agent systems, that can switch between Student Peer, Teacher Assistant, and Parent Advisor roles.

Overview

Conversational AI companions hold promise for scaling "book talk" — structured discussion of reading — in elementary classrooms. Yet most AIEd tools remain monolithic, single-role designs focused narrowly on the student–AI dyad, leaving what this paper calls a "support vacuum" for teachers and parents. The paper first uses a comparative empirical study to expose the limits of a fixed "peer" role, then proposes and justifies a role-adaptive framework grounded in that evidence.

Method

• Design: Within-subjects quasi-experiment — each student interacted with both an AI companion (fixed "student peer" role) and their regular homeroom teacher, serving as their own control.
• Sample: 19 elementary students (12 Grade 4, 7 Grade 5) from an experimental school in Taoyuan, Taiwan, enrolled in a Modeled Sustained Silent Reading (MSSR) program; 4 sessions per student over one month.
• Tools: A research-built conversational companion (Whisper + GPT-3.5) in a peer-like persona, versus experienced homeroom teachers (10+ years).
• Analysis: Session duration, student word/sentence proportion, and 4F reflection categories (Facts, Feelings, Findings, Future), analyzed with paired t-tests and triangulated with qualitative student perceptions.

Key findings

• Longer interaction ≠ more participation. Students spent significantly more time with the AI (Grade 4 t(11)=2.58, p=.026; Grade 5 t(6)=3.97, p=.007) yet contributed a significantly lower proportion of words and sentences (Grade 5: t(6)=−5.62 and −8.21, both p<.01/.001). The paper frames this as conversational dominance — students adopted a passive, reactive role.
• An "affective ceiling" for fixed-role AI. The AI was proficient at factual recall — for Grade 5 it elicited significantly more "Facts" utterances (t(6)=2.76, p=.033). But the teacher prompted significantly more "Feelings" reflection across both grades (most pronounced Grade 5: t(6)=−6.54, p<.001) and, for Grade 4, more "Findings" and "Future" talk.
• Students want both. Qualitatively, students valued the AI as a low-pressure practice partner but preferred the teacher for richer, deeper discussion — supporting the framework's complementary (not replacement) vision.

What this means for practice

• Instructors. Delegate factual comprehension checks and Scaffolding to AI, and reserve human time for the emotional and future-oriented reflection the fixed-role AI could not elicit: the teacher prompted significantly more "Feelings" talk across both grades.
• Instructors. Design for student talk rather than session length — AI sessions ran longer while students contributed a significantly lower proportion of words and sentences — by building in intentional silence and varying prompting density to force student-initiated contributions.
• Designers. Stop treating a role label as a design. A "student peer" persona produced expert dominance in a peer shell, so specify the adaptive interaction logic that modulates turn-taking and floor-sharing, not just the persona.
• Designers. Decompose companion support into modular roles — Student Peer for low-stakes practice, Teacher Assistant for diagnostic analytics, Parent Advisor for home-based guidance — with differentiated dashboards and privacy-aware data flows.
• Researchers. Test the three-role architecture once it is implemented: only the single-role peer has been evaluated empirically, and the framework's core mechanisms (real-time affective sensing, autonomous role transitions) remain unbuilt.

Limitations

• The empirical study is formative with N = 19 elementary students (12 Grade 4, 7 Grade 5) at one experimental school in Taoyuan, Taiwan; the authors state generalizability to broader and more diverse cultural and educational contexts requires further investigation.
• The within-subjects comparison is against each student's regular homeroom teacher rather than an active control condition, so the identified "competence gap" serves as a design rationale rather than a definitive measure of AI versus human efficacy.
• The Role-Adaptive AI Companion Framework itself remains conceptual and architectural: full implementation of the three-role system is pending, with accurate real-time sensing of user affective states and seamless role transitions inside a single agent still unsolved.

Citation

Liao, C.-Y. (2026). Beyond a single role: Justifying a role-adaptive framework for AI companions through a comparative study in elementary book talk. Education and Information Technologies, 31, 4879–4906.