![[Adobe Stock]](https://sp-ao.shortpixel.ai/client/to_webp,q_glossy,ret_img,w_750,h_375/https://www.psypost.org/wp-content/uploads/2025/01/AI-chatbot-750x375.jpg)
A new study shows that the type of feedback an educational chatbot provides can significantly alter not only how well students learn but also which parts of their brains become active during the process. Research published in npj Science of Learning reveals that feedback designed to make students reflect on their learning process enhances their ability to apply knowledge, while encouraging feedback is better for retaining facts.
Scientists are increasingly exploring how artificial intelligence can support education. Chatbots, which can offer one-on-one guidance and instant feedback, are a popular tool. However, the design of this feedback is important. Past research often focused on simple corrective feedback, like telling a student if an answer is right or wrong.
Jiaqi Yin and fellow researchers at East China Normal University and Zhejiang University wanted to look at different kinds of non-corrective feedback that support the learning process itself. They were interested in how feedback that promotes self-reflection (metacognitive) or provides emotional support (affective) might change learning outcomes and the underlying brain mechanisms. By using brain imaging, they hoped to gain a deeper understanding of why certain feedback strategies might be more effective than others.
To investigate these questions, the research team recruited ninety-three college students and divided them into three groups. Each student learned about the human cardiovascular system by interacting with a specially designed chatbot. The learning material was broken down into fifteen modules. After each module, students would assess their own understanding, and the chatbot would provide a specific type of feedback based on the group the student was assigned to.
One group received metacognitive feedback. This feedback consisted of questions that prompted students to reflect on their learning strategies and understanding. For example, the chatbot might ask, “Are you confident in mastering the above content?” or “What is the most challenging part?” The goal was to encourage students to actively monitor and regulate their own learning.
A second group received affective feedback. This feedback was designed to be encouraging and emotionally supportive. If a student indicated they understood a concept, the chatbot might say, “Great! I believe you will do better in the future learning.” If they struggled, it might offer reassurance, such as, “Don’t be discouraged; it’s normal to have difficulty understanding when you’re just starting to learn.”
The third group served as a control and received neutral feedback. Regardless of their self-assessment, the chatbot would simply say, “Let’s take a break first and then continue with our learning.” This provided a baseline to compare the effects of the other two more active forms of feedback.
Throughout the entire chatbot interaction, the students wore a device that performed functional near-infrared spectroscopy. This non-invasive brain imaging technique uses light to measure changes in blood oxygen levels in the brain’s cortex, the outer layer. An increase in oxygenated blood in a particular area suggests that region is more active. The device’s sensors were placed over the prefrontal cortex, an area involved in high-level thinking, and the right temporoparietal areas, which are related to social and emotional processing. After the learning session, all students completed a retention test to see how much information they remembered and a transfer test, which required them to apply their new knowledge to solve unfamiliar problems.
The behavioral results showed clear differences between the groups. Students who received metacognitive feedback performed significantly better on the transfer test than students in the other two groups. This suggests that prompting students to think about their own learning process helps them develop a deeper understanding that can be applied more flexibly. This group also showed the highest metacognitive sensitivity, meaning they were more accurate at judging whether their answers were correct or incorrect.
When it came to simply remembering facts, both the metacognitive and affective feedback groups outperformed the neutral feedback group. Their scores on the retention test were significantly higher. This indicates that feedback that engages students either strategically or emotionally is more effective for memory consolidation than a simple, non-engaging prompt.
The brain imaging data provided a window into the neural processes behind these behavioral differences. All three groups showed activation in the dorsolateral prefrontal cortex, a brain region associated with working memory and sustained attention. This was expected, as the learning task required students to hold information in mind and focus.
However, each feedback type was also linked to unique patterns of brain activity. The metacognitive feedback group showed greater activation in the frontopolar area and the middle temporal gyrus. The frontopolar area is considered critical for complex reasoning, planning, and self-reflection. The middle temporal gyrus is involved in processing the meaning of words and concepts. Increased activity in these regions supports the idea that metacognitive feedback encouraged deeper, more meaningful engagement with the material.
The group receiving affective feedback showed higher activation in the supramarginal gyrus. This brain region is part of a network involved in understanding the thoughts and feelings of others, as well as processing emotional cues. This suggests that the encouraging words from the chatbot prompted students to engage in more social and emotional processing during their interaction. The neutral feedback group, by contrast, had higher activation in the dorsolateral prefrontal cortex compared to the other groups, possibly indicating they devoted more mental resources to the basic task demands without the additional cognitive load of metacognitive or affective processing.
The researchers also found direct links between brain activity and learning outcomes. In the metacognitive feedback group, the level of activation in the frontopolar area was positively correlated with the students’ metacognitive sensitivity. This provides a strong neurological link, suggesting that the self-reflective prompts directly engaged the brain’s self-monitoring systems, which in turn improved students’ ability to assess their own knowledge.
To explore the link between brain activity and the difficult-to-achieve skill of knowledge transfer, the team used a machine learning model. This model analyzed complex patterns in the brain activity data to predict students’ scores on the transfer test. The model was highly accurate and identified several key brain regions as important predictors. Across all groups, activity in the frontopolar area, supramarginal gyrus, and dorsolateral prefrontal cortex helped predict transfer success. This finding suggests that effective knowledge transfer relies on an interplay of metacognitive, emotional, and core cognitive processes.
The study’s authors acknowledge some limitations. The experiment did not account for individual student characteristics, such as motivation or preferred learning style, which could influence how they respond to feedback. The feedback was also standardized rather than personalized. Additionally, functional near-infrared spectroscopy can only measure activity in the outer regions of the brain.
Despite these limitations, the research provides important insights for designing educational technologies. The findings suggest that different types of feedback can be strategically used to achieve different learning goals. To help students develop a deep understanding and the ability to apply knowledge in new contexts, chatbots should incorporate metacognitive prompts that encourage reflection. To help students with memory and motivation, affective or encouraging feedback may be beneficial. Future research could use these findings to develop adaptive chatbots that monitor a student’s brain activity in real time and deliver the most effective type of feedback for their specific needs.
The study, “Effects of different AI-driven Chatbot feedback on learning outcomes and brain activity,” was authored by Jiaqi Yin, Haoxin Xu, Yafeng Pan, and Yi Hu.
