Hello everyone. This is Ozaki, a second-year master’s student.

In this article, I will introduce a paper on an inquiry-based learning support system that combines “Generative AI,” which has been rapidly spreading in educational settings in recent years, with image recognition technology that captures learner behavior.

Inquiry-Based Learning is a method where students themselves formulate questions and learn autonomously toward solutions. However, situations where “students give up halfway” or “it is difficult to provide appropriate advice to each individual” are often observed. This paper develops and verifies a system that uses “image recognition and ChatGPT” to address these challenges by providing advice to each student at the optimal time.

 

Paper Information

Title: Improving inquiry-based learning through automated ChatGPT-based inquiry prompt and behavioral image recognition

(Improving inquiry-based learning through automated ChatGPT-based inquiry prompt and behavioral image recognition)

Authors: Pin-Hui Li, Ting-Ting Wu, Yueh-Min Huang

Journal: Education and Information Technologies, 30(17), 25543-25575

Publication Year: 2025

 

Introduction and Background

Inquiry-based learning is a highly effective learner-centered approach for fostering deep understanding and problem-solving skills. However, due to its requirement for “autonomy,” it can also be a significant burden for learners. Students who lack sufficient curiosity or self-efficacy are at risk of becoming confused or frustrated during the activity.

Ideally, a teacher would monitor each student’s progress and provide scaffolding at the appropriate time. On the other hand, in large classes, it is physically difficult for a teacher to grasp the situation of every student in real-time and provide individual guidance.

While there have been support systems using technology in the past, they lacked flexibility, such as insufficient real-time monitoring of “what the learner is doing now” and being limited to returning pre-determined template messages.

Therefore, this study developed a system called “AIPS (Automated Inquiry Prompt System)” that monitors student behavior using image recognition technology and generates context-aware advice using ChatGPT.

 

How AIPS Works

This AIPS system has two main functions.

“Behavior Recognition” using YOLObile

First, it uses YOLObile, a lightweight image recognition model, to detect what the student is doing in real-time. Specifically, it recognizes the following elements through a webcam:

・Face, hands

・Keyboard, mouse, etc.

Furthermore, it determines whether the student’s hands are overlapping with objects (operating them) or if they are facing a neighboring student (discussing), and judges every 3 minutes whether they are “actively inquiring” or “stagnating.”

“Dynamic Prompt Generation” using ChatGPT

Based on the results of behavior recognition, ChatGPT generates and presents the optimal prompt for the learner according to their learning progress. In this study, the inquiry process is divided into four stages (POEE model: Prediction, Observation, Explanation, Evaluation), and four types of prompts are used based on Cognitive Load Theory (CLT) and Self-Determination Theory (SDT):

・Constructive prompts: Provide concrete procedures to relieve the anxiety of students who are lost (“Let’s try ~ first”).

・Collaborative prompts: Encourage dialogue with other students (“Let’s try exchanging opinions with the person next to you”).

・Creative prompts: Encourage free thinking and exploration (“What do you think would happen if ~?”).

・Critical prompts: Encourage deep reflection on one’s own thoughts (“Can you explain the reason for that result?”).

For example, if a student’s hands have stopped during the “Prediction” stage of the POEE model, the system can provide a “Constructive prompt” that gently shows the procedure. Conversely, if they are typing away and making progress, it can provide a “Creative prompt” to encourage deeper thinking, allowing for flexible responses.

 

Research Methodology

To verify the effectiveness of this system, an experiment was conducted in a programming lecture for 74 university students in Taiwan. Students were randomly divided into an “experimental group (using AIPS)” and a “control group (not using AIPS)” and engaged in inquiry activities over four weeks.

The following indicators were used for evaluation:

・Curiosity: 5-dimensional curiosity scale (joyous exploration, deprivation sensitivity, stress tolerance, etc.)

・Self-efficacy: Confidence in completing tasks

・Inquiry skills: Skills in task setting, experimentation, analysis, and explanation

・Knowledge construction: Mastery of expertise through post-tests

 

Results and Discussion

The results of the experiment showed that the group supported by AI achieved significantly higher results in all indicators compared to the group that was not.

(1) Improvement in Curiosity and Self-Efficacy

In particular, “joyous exploration” and “deprivation sensitivity (the desire to know)” improved significantly. This is thought to be because the AI provided hints at the appropriate time, allowing students to maintain a positive feeling of “wanting to know more” without being overwhelmed by the stress of “not understanding.” In addition, because the AI broke down and presented tasks even in difficult situations, it is thought that self-efficacy, the belief that “I can do it too,” was enhanced.

(2) Dramatic Improvement in Inquiry Skills

The most significant difference was in “experimental skills.” This is the result of the AI’s feedback on specific procedures being directly effective. On the other hand, the improvement in “task setting skills” seemed relatively small. This suggests that the ability to formulate high-quality “questions” is difficult to cultivate through short-term intervention alone and requires longer-term support.

(3) Synergistic Effect on Knowledge Construction

The analysis results proved that increases in curiosity, self-efficacy, and inquiry skills directly lead to final knowledge construction. In particular, it was found that the motivation to “solve what I don’t understand” and the “explanation skill” to logically verbalize it play a central role in internalizing knowledge.

 

Conclusion and Limitations

This study demonstrated that by combining image recognition and generative AI, it is possible to move away from conventional uniform instruction and provide “adaptive instruction” tailored to the situation of individual students.

However, there are limitations. While this system is suitable for inquiry in science subjects involving clear actions such as “experimentation” and “observation,” it is considered that behavior recognition alone has limitations in humanistic inquiry where one thinks deeply in their mind, or in activities without clear correct answers. The author states that multimodal analysis that also combines language data, such as student utterances and chat history, will be required in the future.

 

Reasons for Selection and Reflections

Reason for selection:

I chose this paper because I am interested in how to support inquiry-based learning and how to measure its effectiveness. In particular, it was very informative that AI-based support for the inquiry-based learning process showed the potential to synergistically enhance not only knowledge but also general abilities such as the learner’s “curiosity” and “inquiry skills.”

Reflections:

What made me think most after reading this paper is the potential of AI for purposes other than knowledge construction. In this study, “acquisition of expertise” was set as the final goal, but in the context of Japanese education, especially in “Integrated Studies” and “Career Education,” the amount of knowledge is not necessarily the only goal. The process of deeply reflecting on “one’s own way of being and living” or making decisions about future careers is considered important. How do AI-driven prompts and learner behavior analysis function in such abstract inquiry processes without correct answers? And how does that lead to the learner’s “actual behavioral change” or “change in consciousness”? There are still many parts of this that have not been clarified, and it is a theme that I myself want to pursue from now on.

Inquiry-based learning is expected to have a positive impact on initiative and curiosity. On the other hand, the learning process is complex and high-load, and this paper also points out that it is difficult to continue learning autonomously without having initiative and curiosity in the first place. I would like to think more deeply in the future about what kind of support can be designed to realize individually optimized learning for learners in large classes who have varying levels of motivation and ability for inquiry-based learning.

Author: Kohei Ozaki