The public release of large language model (LLM)-based conversational artificial intelligence such as ChatGPT has sparked multifaceted discussions about AI-based learning environments in the educational domain. However, most current AI-based learning ...
The public release of large language model (LLM)-based conversational artificial intelligence such as ChatGPT has sparked multifaceted discussions about AI-based learning environments in the educational domain. However, most current AI-based learning environments are being developed based on learner models that are biased toward learners' cognitive achievement, with relatively insufficient integrated consideration of learners’ affective states. In learning, affect is a key variable closely connected to academic achievement and an important factor that determines the qualitative level of learning experience along with cognitive engagement. Particularly, low-achieving learners are more vulnerable to negative affective experiences such as sadness, shame, and anger, while high-achieving learners may also experience considerable psychological burden due to excessive pressure to maintain grades, perfectionist tendencies, and anxiety about failure. Since affective problems are issues that all learners face, personalized interventions for learners' affective support are needed in AI-based learning environments. Empathy has been continuously studied in educational and learning contexts as a method of affective support for learners, and empathy helps learners perceive themselves as valuable and supported, thereby preparing them to actively engage cognitively in learning.
However, previous research on empathetic feedback has shown two key limitations. First, previous empathetic feedback systems have been implemented through rule-based approaches, which have limitations in flexibly responding to learners' diverse individual needs and contextual situations. Recent advances in LLMs provide new possibilities for implementing contextual empathy to overcome these limitations. Nevertheless, research on LLM-based contextual empathy feedback in education is still in its early stages. Second, multifaceted validation studies on the effects of empathetic feedback on learners' affect and cognition remain insufficient. According to learning sciences research, cognition and affect are intertwined and activated together rather than appearing independently, and the effects of empathy may manifest in cognitive responses as well as affective ones, even without learners' conscious awareness. Therefore, to accurately measure the subtle effects of empathetic feedback, a holistic approach using multimodal data including physiological measurement tools such as EEG that can provide objective evidence of learners' cognitive engagement levels and affective stability, along with heart rate variability, skin conductance, web logs, surveys and tests, and interviews is necessary.
Therefore, this study implemented adaptive empathetic feedback that considers learning context in an AI-based learning environment using ChatGPT and verified the effects of adaptive empathetic feedback on learners' cognitive engagement and emotion through multimodal learning analytics. The specific research questions were: 1) What effects does adaptive empathetic feedback have on emotion and learning motivation in AI-based learning environments? 2) What effects does adaptive empathetic feedback have on learners' cognitive engagement and learning outcomes in AI-based learning environments? 3) What effects does adaptive empathetic feedback have on learners' brain activity in AI-based learning environments? 4) What are learners' perceptions of adaptive empathetic feedback in AI-based learning environments? The adaptive empathetic feedback varied the empathetic feedback approach according to learners' contextual information (learning records, emotion records, etc.) and supported learners to feel that their learning situations are understood. To empirically verify the effects of adaptive empathetic feedback, an experimental study was conducted applying a control group pretest-posttest design, and interviews were conducted to confirm learners' perceptions of adaptive empathetic feedback. The study was conducted with 62 adult learners (31 in the experimental group, 31 in the control group), with the experimental group provided an AI-based learning environment with adaptive empathetic feedback and the control group provided the same learning environment with neutral guidance messages. Learners studied by solving English grammar problems in the AI-based learning environment. To confirm the effects of adaptive empathetic feedback during this process, multimodal data including surveys and tests, psychophysiological responses (skin conductance, heart rate variability), EEG, web log data, and interviews were collected and analyzed.
The research results showed that the effects of adaptive empathetic feedback were observed more clearly through EEG, psychophysiological data, and web log data collected during the learning process rather than post-hoc data such as surveys and tests. This suggests that the effects of empathy are better revealed through data from the learning process that learners are not consciously aware of rather than post-hoc data based on learners' conscious awareness. The detailed research results are as follows.
First, adaptive empathetic feedback affected 'enjoyment' among post-learning emotions. This was the only effect of empathetic feedback observed in survey results, with learners in the experimental group reporting greater enjoyment of learning after completion compared to the control group. However, it did not influence emotion scores during learning, psychophysiological stress indicators, or post-learning motivation.
Second, adaptive empathetic feedback partially affected the enhancement of learners' cognitive engagement in medium-difficulty problems. Psychophysiological data analysis revealed that the experimental group exhibited higher frequencies of momentary arousal compared to the control group, with statistically significant differences particularly evident in medium-difficulty problems. However, it did not affect cognitive engagement at low or high difficulty levels, nor learning outcomes, indicating no differences in learning results. When examining interactive engagement through web logs as another indicator of cognitive engagement, the control group demonstrated significantly higher rates of additional question behaviors in high to medium-difficulty problems. However, content analysis revealed that many of these were basic 'remember'-level questions, such as word and sentence translations. Furthermore, learners' situational cognitive engagement confirmed through post-surveys showed a positive correlation with post-learning enjoyment emotions, indicating that learners' enjoyable emotions and their perceived level of cognitive immersion in the learning process were interrelated.
Third, adaptive empathetic feedback had positive effects on learners' brain activity. The experimental group showed strong brain activity expressed as gamma waves during problem-solving, correctness feedback, and learning feedback moments, and showed stabilized brain activity expressed as alpha waves during empathy feedback moments. Particularly, brain wave patterns similar to reward effects were observed in correctness feedback for correct answers, brain wave patterns reflecting active learning in overall learning feedback, and brain wave patterns indicating psychological stability in empathy feedback for incorrect answers.
Finally, interview analysis revealed themes of effective empathy characteristics (intimate tone, cognition-based empathy approach, encouragement), affective effects (negative affect alleviation, self-efficacy and achievement enhancement, learning motivation improvement), learning effects (deepened conceptual understanding, sustained concentration), and improvement points (diversification of empathy expression, character use, personalized personality). Learners provided detailed explanations of why empathy was effective while presenting opinions on improvement points for the adaptive empathetic feedback used in the study.
The results of this study show that adaptive empathetic feedback positively transforms the learning experience by effectively enhancing learners' cognitive engagement and providing affective achievement and stability. Additionally, it suggests that the effects of empathy were better revealed through data from the learning process that learners are not consciously aware of rather than post-hoc data based on learners' conscious awareness. These findings demonstrate that empathy can play an important role as a cognitive and affective support method in AI-based learning environments. This study is expected to contribute to adaptive AI-based learning environment design by strengthening the theoretical foundation for cognitive engagement promotion and affective support through empathy in AI-based learning environments and presenting practical implementation approaches.