Preferred Scaffolding Does Not Lead to Better Learning Performance: Empirical Evidence from AI-Supported Mathematical Modelling

Synthesis: Zhu, Yang, and Yang (2026) investigated how different AI interaction roles function as scaffolding strategies in AI-assisted mathematical modelling. In a randomized within-subjects experiment with 26 university students, they compared five AI roles — Tutor, Teaching Assistant, Peer, Excellent Student, and Struggling Student — on modelling competency, role preference, and learning experience. Students demonstrated higher modelling competency with the Peer and Teaching Assistant roles, which fostered collaborative reasoning and co-construction of ideas, yet expressed stronger preferences for the Tutor and Excellent Student roles offering explicit guidance. The study reveals a notable divergence between learning performance and role preference, highlighting the importance of balancing cognitive Scaffolding with collaborative sense-making in mathematics education.

Five AI roles as distinct scaffolding configurations

The study operationalizes AI roles not as social labels but as dynamic scaffolding configurations encoded in interaction patterns. The Tutor explains first, guides thinking with questions, and withholds the answer until understanding is confirmed; the Teaching Assistant uses open-ended question-based guidance, offers clues, and never directly provides the final answer; the Peer engages in collaborative, discussion-based problem-solving; the Excellent Student delivers direct, complete solutions efficiently; and the Struggling Student supplies inconsistent, ambiguous, or incomplete input. These differences in directiveness and epistemic Agency are the mechanism hypothesized to drive learning.

Where scaffolding changed performance

Performance differences were dimension-specific rather than uniform. In Model Abstraction, contingent and process-oriented roles (Tutor, Teaching Assistant, Peer) outperformed configurations giving either excessive guidance (Excellent Student, which bypassed students' generative reasoning) or insufficient guidance (Struggling Student, which raised cognitive load and disrupted identification of key modelling variables). In Reflection & Iteration, contrastive scaffolding (Peer, Excellent Student, Tutor) led to higher performance by exposing learners to alternative solutions they could evaluate against their own models. Situational Understanding showed no differences, and Mathematical Representation remained low across all conditions — a pattern the authors link to a possible AI-dependency effect.

Why preference diverges from performance

Students consistently preferred the explicit, authoritative Tutor and Excellent Student roles, rating them higher on perceived usefulness, ease of use, and Self Efficacy, while the Struggling Student role was rated lowest everywhere. The essential finding is that preference and competency were shaped by different mechanisms: preference tracked fluency, confidence, and reduced uncertainty, whereas competency depended on active generation, epistemic agency, and reflective reasoning. Directive roles felt efficient but shifted responsibility for reasoning toward the AI; collaborative roles created productive difficulty consistent with productive failure and desirable-difficulty research. This aligns with the ICAP framework's claim that constructive and interactive engagement produces deeper learning than passive reception.

Implications for adaptive AI learning design

Because generative AI now produces fluent, immediate, personalized responses, there is a risk that learners become passive consumers of AI-generated solutions. The authors argue perceived usefulness, ease of use, and immediate satisfaction should not be treated as sufficient indicators of scaffolding effectiveness. Instead, design should balance clarity and efficiency with opportunities for explanation, comparison, revision, and collaborative sense-making. For adaptive, learner-centered AI-supported modelling environments, the implication is that the most popular role is not necessarily the most pedagogically effective — supporting the need for intentional scaffolding design rather than learner-driven default preferences.

Connected Concepts

• Generative AI
• Pedagogical Agent
• Collaborative Learning
• Student Engagement
• Student Experience
• Adaptive Learning
• Metacognition
• Self Regulated Learning
• Feedback

Connected Articles

• AI Enhanced Pbl Chatgpt Scaffolding 2026 — AI-enhanced PBL with ChatGPT scaffolding
• Guided LLM Scaffolding Independent Learning — Guided LLM scaffolding for independent learning
• Chudziak AI Math Tutoring Platform — AI math tutoring platform

Citation

Zhu, W., Yang, Y., & Yang, Y. (2026). Preferred scaffolding does not lead to better learning performance: Empirical evidence from AI-supported mathematical modelling. Computers and Education: Artificial Intelligence, 100669.