Reference: 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.

Does Preferred AI Support Produce Better Learning?

When generative AI is introduced into education, researchers often ask which type of AI students prefer. It is tempting to assume that an AI system perceived as friendly, convenient, and helpful will also produce better learning outcomes.

However, the study by Wangda Zhu, Yin Yang, and Yuqin Yang challenges this assumption. It asks a deceptively simple question: Does the AI scaffolding that students prefer actually lead to better learning performance?

Mathematical modelling requires more than calculating a correct answer. Learners must understand a real-world situation, identify relevant conditions, translate the situation into mathematical representations, interpret the results, and revise the model when necessary. These processes involve abstraction, representation, validation, and iterative reasoning.

If AI provides too much information, students may complete the task more easily while losing opportunities to construct and evaluate the model independently.

Five Different Roles for AI

The researchers designed five AI roles representing different instructional approaches.

The “Tutor” delivered expert explanations and systematic solution procedures. The “Teaching Assistant” avoided giving direct answers and instead used questions and step-by-step hints. The “Peer Student” exchanged ideas and participated in collaborative correction. The “Excellent Student” offered correct answers quickly but provided limited explanation. Finally, the “Struggling Student” displayed uncertainty and errors, requiring learners to judge and correct its reasoning.

Participants completed mathematical modelling tasks while interacting with the different AI roles. They then evaluated their preferences, confidence, perceived helpfulness, and ease of interaction. This allowed the researchers to compare not only whether students liked an AI role but also how that role affected task performance and learning.

Preferred Support Was Not Always Effective Support

Students generally responded positively to AI roles that provided clear and competent assistance. The Tutor and Excellent Student roles were relatively well received, whereas roles involving frequent errors or inefficient communication received lower evaluations.

The most important finding, however, was that students’ preferred form of scaffolding did not necessarily produce the best learning performance. Support that felt comfortable and satisfying was not always the support that stimulated the most productive reasoning.

This result demonstrates that preference and satisfaction should not be treated as direct evidence of learning. Even when students report that AI was helpful, their independent problem-solving ability may not improve if the AI performs the essential reasoning for them.

An AI That Solves Tasks and an AI That Promotes Learning

The study highlights the need to distinguish between an AI that helps students finish a task and an AI that helps them learn how to complete the task independently.

An AI that quickly supplies answers and explanations may increase confidence and satisfaction. However, if it presents the solution before the learner has structured the problem or selected a strategy, it may encourage cognitive offloading. Essential reasoning is then transferred from the learner to the AI.

In contrast, a Teaching Assistant that asks questions and waits for the learner to respond may feel less convenient. Nevertheless, it can encourage students to examine the conditions of the problem, explain their assumptions, and select a solution strategy independently.

Effective educational AI is therefore not necessarily the AI that provides the most information. It may be the AI that identifies when support is needed, intervenes only as much as necessary, and gradually withdraws as the learner becomes more capable.

Satisfaction Alone Cannot Demonstrate Educational Effectiveness

The study warns against evaluating AI-supported education primarily through satisfaction or perceived usefulness. Liking the AI, improving immediate task performance, transferring learning to a new problem, and performing without assistance are different outcomes.

Evaluations of AI learning environments should therefore examine how students use questions and hints, how they correct errors, how often they retry a task, and whether they can continue independently after support has been reduced.

The final answer is only one source of evidence. Learners’ modelling processes and their interaction records with AI are equally important for understanding how learning occurs.

Implications for Physical AI-Based Mathematics Learning

The same principle is relevant to Physical AI-based mathematics education. AI does not need to maintain a single role throughout an activity. Its role can change according to the stage of learning.

During task interpretation, AI may ask questions that help learners identify goals and conditions. During mathematical structuring, it may offer hints about variables and relationships. After physical execution, it may provide feedback comparing predicted and measured outcomes. During model revision, it may act as a reflective peer that encourages learners to explain the causes of errors.

As learners become more capable, the support should gradually fade. The final objective is not continued dependence on AI but independent participation in the cycle of prediction, execution, comparison, and revision.

Limitations and Directions for Further Research

The study is based on a limited sample and relatively short mathematical modelling tasks. It therefore cannot fully explain whether the observed effects persist over time or transfer to new learning situations.

The AI roles were also primarily text-based. Their effects may differ in a Physical AI environment where mathematical predictions are followed by visible movements, measurements, and physical feedback.

This limitation creates an opportunity for subsequent research combining AI interaction records with physical execution data, error-correction processes, model revisions, and independent performance after scaffolding has been withdrawn.

The Best AI Knows When to Step Back

Ultimately, the best AI teaching assistant may not be the one that gives the most explanations or receives the highest satisfaction ratings. It may be the AI that intervenes appropriately when support is needed and steps back when the learner is ready to act independently.

This paper shifts the central question of AI education from “Did students like the AI?” to “Which AI role produced which forms of thinking, action, and learning?” Its most valuable insight is that the role, level, timing, and gradual fading of AI support may matter more than the mere presence of AI.