Reflective chatbot feedback improved immediate transfer, while encouragement mainly supported retention

Feedback can invite a learner to reconsider understanding, offer encouragement or simply signal a pause. Jiaqi Yin and colleagues compared these functions inside a chatbot lesson rather than treating every supportive response as equivalent. Their study asked whether reflective and affective messages would influence retention, transfer and the accuracy of learners' judgments about what they knew. It also recorded brain activity, providing an additional research perspective without replacing the behavioral evidence that matters for educational decisions.

The researchers recruited 93 university students in Shanghai who were not biology majors. After six exclusions caused by recording problems or missing screen recordings, the analysis included 87 students, with 29 in each randomly assigned feedback condition. All studied fifteen modules about the cardiovascular system. Each module combined learning, a self-assessment and feedback. The reflective condition prompted monitoring or planning, the affective condition offered encouragement or praise, and the neutral condition provided a break prompt. These were structured feedback arrangements, not a comparison of unrestricted language models.

After accounting for initial knowledge, both reflective and affective feedback produced higher retention scores than neutral feedback. The reflective condition also performed better on questions requiring application to new problems, with mean transfer scores of 9.38 compared with 6.72 in the neutral condition. Its confidence judgments tracked answer accuracy more closely. Affective feedback did not show a significant transfer advantage over the neutral condition. These different patterns support choosing feedback according to the intended learning outcome rather than assuming that a pleasant message accomplishes every instructional purpose.

Functional near-infrared spectroscopy measured blood-oxygen changes over selected cortical regions during the lesson. Different feedback conditions showed different activation patterns, and some signals correlated with metacognitive sensitivity or retention. Such findings can inform hypotheses about how learners respond to feedback. They do not allow a teacher to infer that more activation necessarily means deeper learning, or that a particular student's understanding can be read directly from a brain signal. Predictive modeling within this sample also requires independent validation before any practical use.

The design has clear boundaries. Outcomes came from immediate retention and transfer tests after a short lesson, not from delayed classroom performance. The neutral comparison offered little substantive instructional guidance, so superiority to it does not establish superiority to a strong teacher-feedback approach. Participants came from one university, and the measurement system covered only selected brain areas. The intervention also combined different messages with corresponding reflection or emotion ratings. The evidence therefore supports the tested packages, while leaving the contribution of each individual component unresolved.

For AIEDHK, the most usable implication is a simple feedback-design question: what should the learner do next? A reflective prompt could ask for an explanation, identify an uncertainty or encourage a change in strategy; supportive language could help maintain willingness to continue. A classroom pilot should measure the intended outcome separately, including a later unfamiliar task if transfer is the goal. These are proposed design and evaluation steps beyond the experiment. The study provides a reason to distinguish feedback functions and verify their consequences through learner performance, rather than use neuroscience imagery as a shortcut to educational credibility.