How Learners Interpret Emotion-Aware Feedback in AI-Supported Learning: Evidence from a Classroom Study

Emotion is increasingly incorporated into AI-supported feedback in education, yet less is known about how learners interpret emotion-related messages once they are presented. This paper reports an exploratory classroom-based study comparing three learner-facing strategies for presenting emotion-aware feedback: inference with explanation, inference without explanation, and deliberate non-inference. Using a Wizard-of-Oz procedure embedded in a web-based classroom activity, 78 undergraduate students completed a conceptual quiz, a brief reflection task, and an applied data-analysis task during a 90-min course session. Following the activity, participants evaluated the system on six 7-point Likert outcomes: Perceived Accuracy, Interpretability, Emotional Comfort, Willingness to Reuse, Perceived Usefulness, and Trust. Significant differences were observed across all six outcomes. Across every dimension, the same ordinal pattern emerged: feedback with explanation received the highest ratings, no inference occupied an intermediate position, and inference without explanation was rated lowest. Notably, deliberate non-inference was evaluated more favorably than unexplained inference across all six outcomes. These findings suggest that the learner-facing value of emotion-aware educational AI depends not only on whether emotion is inferred, but on how such inference is presented and contextualized. The study contributes classroom-based evidence that learner interpretation should be treated as an important criterion in evaluating emotion-aware educational AI and that deliberate non-inference can function as a legitimate response strategy when affective claims cannot be presented in an intelligible and contextually grounded way.