Abstract
This paper reports on a classroom practice that applied generative AI to the moral education text “Yuuhe — Thank You for Being Alive.” AI was used in two ways: “DeepCross,” which analyzes students’ responses during class discussions, and “DeepLong,” which provides individualized feedback after class. The study examined how these approaches can deepen students’ understanding. It was also found that GPT-4o is characterized by empathetic advice, while GPT-5 tends to offer more analytical suggestions, and that combining the two can help reduce teachers’ workload and enrich classroom discussions.
Update (2025/09/10)
We are releasing GPT-ScoreAID_ver1.7.58 corresponding to this article.
Due to GPT-5 congestion and the current inference depth settings, responses may still occasionally get cut off, but the software has now reached a practically usable level.👉 Download it here!
If you have any questions or inquiries for the author, please feel free to contact me here.👉Contact the Author !
Introduction
In the field of moral education, a variety of examples have been reported on the use of generative AI.
In this blog as well, I have introduced a case where generative AI was applied to the teaching material “I Will Become a Cleaning Professional” (published by Nihon Bunkyo Shuppan Co., Ltd.), which features Ms. Haruko Niitsu, a cleaning master at Haneda Airport.
My stance on applying generative AI in moral education is to explore its use as a tool that supports teachers in their classroom management, focusing on the following two perspectives.
- One approach is “Ōshin” (DeepCross Analysis), where AI analyzes and categorizes in real time the opinions expressed by students in response to the teacher’s questions during class, providing the teacher with advice to stimulate discussion and guide the direction of debate.
- Another approach is “Jūshin” (DeepLong Analysis), where after class, AI analyzes each student’s responses to the moral education questions, examining what the student felt and how it contributed to their growth and learning. The AI then generates suggested feedback that can be added as comments to the students’ moral notebooks.
In this article, I will also explore the use of generative AI from these perspectives in relation to the moral education material “Yuuhe — Thank You for Being Alive” (published by Nihon Bunkyo Shuppan Co., Ltd.).
Unlike the previously discussed “I Will Become a Cleaning Professional”, “Yuuhe” addresses a different content theme, D-(19) “The Preciousness of Life.” What kind of assistance can the latest generative AI provide for a moral education material that deals with such a profound theme for human beings as “life”? From the perspectives of “DeepCross” and “DeepLong,” I will explore the possibilities of AI in supporting this lesson.
The class itself was conducted in February 2025. However, due to my work commitments, the summary was delayed until August. In the meantime, ChatGPT had advanced, and by then a newer version, GPT-5, had become available in addition to the GPT-4o that I had used for analysis back in February. Therefore, in this article, I also reanalyze the data from that time using the latest GPT-5 from the perspectives of “DeepCross” and “DeepLong,” exploring not only the differences between GPT-4o and GPT-5 but also how to design effective prompts.
It should be noted that I was unable to find freely accessible teaching plans or related materials for the lesson “Yuuhe — Thank You for Being Alive” on the website of its publisher, Nihon Bunkyo Shuppan Co., Ltd. Therefore, I have instead included a link to a teaching plan from another lesson that can serve as a reference for structuring the class. The lesson I conducted generally followed the flow outlined in that plan. Please use it as a reference when reviewing the sequence of the lesson and the placement of the key questions.
An outline of the teaching material “Yuuhe — Thank You for Being Alive” is presented below.
This material is based on the memoir of a mother who lost her one-and-a-half-year-old son, Shō, in the Great Hanshin-Awaji Earthquake. For a long time, she carried the burden of guilt and suffering over not being able to save him amid the chaos of the disaster, continually blaming herself for having survived. Her feelings toward Shō, who was never able to grow up, ran deep, and at times she nearly lost sight of the presence of his twin sister, Yuu. However, she chose to “live together with Yuu,” which led her to realize the preciousness of life and the profound gratitude for simply being alive. Over time, with the support of her family and those around her, she regained her smile, coming to accept that not only grieving but also living with joy could be a way of cherishing the memory of her lost child.
“DeepCross”: Analyzing and Categorizing Students’ Opinions in Real Time to Explore the Direction of the Lesson
In the sample teaching plan provided by the textbook publisher, two questions are suggested to confirm students’ understanding of the material, followed by a central question that delves into the main theme of the lesson. In my class as well, I followed this plan and collected students’ responses through an online form. The three example questions were as follows:
Question 1:
“What feelings do you think Ms. Takai, the mother of Shō and Yuu, experienced when the sudden earthquake struck?”Question 2:
“After losing Shō, Ms. Takai endured many days of deep sorrow and was unable to recover easily. However, she chose not to go to Shō’s side, but instead to live on together with Yuu. From what thoughts or feelings do you think she made this choice?”Central Question:
“In the words Ms. Takai gave to this essay, ‘Thank you for being alive,’ what feelings do you think are contained within them? What was the important realization that enabled her to overcome her deep sorrow and continue living?”
The idea of “DeepCross” is to use AI to analyze and categorize students’ responses to each of these questions, thereby supporting class discussions. For this purpose, I had previously suggested the following type of prompt in the lesson using the material “I Will Become a Cleaning Professional.”
- The Basic Form of a “DeepCross” Prompt
- Provide the full text of the teaching material to the AI.
- Then, explain how the lesson is planned to proceed, incorporating the three key questions.
- Reserve blank spaces in the prompt so that the students’ actual responses collected through the form during class can be inserted.
- Clearly instruct the AI that its task is to perform analysis and categorization, and include directions regarding the output format.
The Actual “DeepCross” Prompt for Question 1
The following passage is used in a first-year junior high school moral education class, titled “Yuuhe — Thank You for Being Alive.”
—– Beginning of Text
Yuuhe — Thank You for Being AliveAt 5:46 a.m. on January 17, 1995 (Heisei 7), the Great Hanshin-Awaji Earthquake struck, just as people were beginning their morning routines in the still-dark early hours. Suddenly, with a great tremor, I awoke.
“Mom screamed: What is this!? An earthquake!?”
(…middle sections omitted…)
And not only your father, your grandfather, and your grandmother, but many others as well grieved your death and even now still hold you dear.
Thanks to the many people who accepted my sorrow and stayed by my side, I was able to smile again. I am deeply grateful to all of them. I believe that Shō can also see my smile.
—– End of Text
In this moral education class, students are asked three key questions to develop the lesson.
—– Lesson Development Begins
Lesson Development① At the beginning of the lesson, I asked students:
“What feelings do you think Ms. Takai, the mother of Shō and Yuu, experienced when the sudden earthquake struck?”The aim here is to make students aware that the Takai family, including Shō and Yuu, were suddenly confronted with a life-threatening situation caused by the earthquake.
Expected student responses include ideas such as: “What is this!?,” “I hope it stops soon,” “I might die!,” or “Are my children safe?” Through imagining Ms. Takai’s feelings and the situation during the earthquake, students are expected to recognize the imminent danger to life.② Later, in the middle part of the lesson, I asked:
“After losing Shō, Ms. Takai endured many days of deep sorrow and could not easily recover. However, she chose not to follow Shō in death, but to live together with Yuu. From what thoughts or feelings do you think she made this choice?”The purpose of this question is to help students empathize with Ms. Takai’s deep grief, recognize that her recovery was not easy, and consider how she arrived at the decision to “live together with Yuu.”
Expected responses might be: “If she had gone to Shō, she would have caused even greater sadness for Yuu and her family who are still alive.” or “Although she can never forget Shō and still suffers, she wants to treasure the family members who remain with her and keep living.”③ Finally, as the central question of the lesson, I asked:
“In the words Ms. Takai gave to this essay, ‘Thank you for being alive,’ what feelings do you think are contained within them? What was the important realization that enabled her to overcome her deep sorrow and continue living?”The aim of this question is to lead students to realize that Ms. Takai noticed Yuu—Shō’s twin sister who had survived—was also struggling but kept encouraging her, and that cherishing her time with Yuu would allow her family’s life, including Shō’s, to shine.
Expected student responses might include: “Mom wasn’t alone,” “Yuu kept supporting her,” or “I want to live not only in sadness but also with a smile.” Deeper reflections might be: “Being alive is not something to take for granted; we are kept alive by the support of many others,” “The presence of those around us is essential for living,” or “It is important not just to encourage, but to support others in a way that helps them recover from grief by staying close to their feelings.”—– End of Lesson Development
For the first question, the students’ responses were collected and listed as follows:
—– Student Responses to Question 1 (blank in advance, then filled from online form)
I was surprised.
Panic, only realizing an earthquake was happening.
Shocked by the big earthquake, feeling rushed.
Worried about the family.
Surprised at the sudden earthquake, wondering what to do.
At first, not understanding what was happening.
Caught off guard, panic, couldn’t keep up with understanding.
Surprised by the sudden earthquake, only thinking about herself.
“What!? An earthquake!? Why!? What should I do!?”
Thinking she must quickly escape while holding Yuu and Shō.
Shocked by the large tremor.
Feelings of surprise, anxiety.
A sense of surprise.
Wondering if nearby buildings and the children were safe.
Surprised by the sudden earthquake.
Feeling sad, apologizing in her heart for not being able to protect the children, for not being able to save them, for only thinking of herself.
Very sad, wanting to cry.
Feelings of panic, wondering what to do.
Feeling of sadness.
Wanting to survive.
Shocked.
Wanting to protect the children.
Anxiety.
Surprised at the sudden large tremor.
Worried whether the children were in danger; panicked at the sudden earthquake.
Shocked by the sudden earthquake; worried about the children.
Feeling fear and surprise, unable to move.
“This is bad.”
“This is bad, what should I do? I don’t know what to do,” panic.
Surprised, confused, unable to grasp the situation.
—– End of Student Responses to Question 1
After this, the teacher proceeds with class discussion.
Task for AI:
Please categorize the students’ responses to Question 1 and identify two or three specific perspectives that could stimulate discussion among students during the debate.
It is likely that all textbook publishers provide schools with materials containing information similar to what is presented here. When applying the “DeepCross” method, the key is to prepare such a structured prompt in advance, then quickly extract students’ responses collected through the form into a spreadsheet during class, paste them into the prompt, and use them effectively.
In the future, if textbook publishers could also provide information for constructing prompts that are optimized for generative AI, the amount of preparation required by schools beforehand would be greatly reduced.
AI’s Analysis and Categorization of the Questions
(a) Analysis and Categorization of Question 1
When this prompt was submitted to ChatGPT-4o, it returned the following advice in about 11 seconds.GPT-4o’s Response to Question 1
Answer: Let us categorize the students’ responses to Question 1 and identify several perspectives that could spark discussion.
- Reactions of Surprise and Panic
Many students responded with comments such as “I was surprised” or “I panicked.” A discussion could arise about why people fall into surprise or panic in emergencies, and how such reactions can be overcome.
By reflecting on responses like “I panicked” or “I was shocked,” and having other students consider “How can one stay calm and act appropriately?”, the class can deepen its discussion on appropriate responses during an earthquake.- Worries About Family and Children
There were also many responses such as “I was worried about my family” or “I was worried about the children.” This could lead to discussions about a parent’s sense of responsibility in disasters and the mission of protecting one’s family. In particular, it might encourage reflection on how parents should protect their children and what actions they should take.
By exploring the tension between the desire to “protect the family” and the feeling of “not knowing how to act right away,” students can discuss family bonds and individual roles during unexpected disasters.- Regret or Self-Reflection on One’s Own Actions
Responses such as “I only thought of myself” or “I’m sorry I couldn’t protect them” could spark discussions about choices of action during disasters and the regret that follows. Students may talk about how to process feelings that arise from decisions made under extreme conditions.
They might consider the difference between panicking for self-preservation and later feelings of regret or guilt. Based on comments like “I only thought of myself,” discussions could be encouraged about the mindset needed to maintain the capacity to think of others, even in emergencies.
By advancing the discussion from these perspectives, students can share the emotions and reactions within themselves and gain deeper insights into the importance of calm action in disasters and the value of family.
In my class, I combined perspectives 2 and 3, using this question as a lead-in that, by tracing Ms. Takai’s feelings and actions in the text, would eventually connect to themes of regret and self-reflection. In this way, I encouraged students to exchange their opinions.
(b) Categorization of Question 2 and the Central Question
Using the same method, the “DeepCross” categorizations by generative AI for Question 2 and the central question were as follows.
GPT-4o’s Response to Question 2
Answer: Let us categorize the students’ responses to Question 2 and identify specific perspectives that are likely to stimulate active discussion.
Responsibility and Sense of Mission of the Survivors
“I thought I had to live on for Shō’s sake.”
“I wanted to properly raise Yuu to the end, on behalf of Shō.”
These responses can deepen discussion about the sense of responsibility and mission carried by those who survive. Students can come to understand and reflect on the idea that living is not merely survival, but also involves carrying out a mission.Consideration and Empathy for Those Left Behind
“Because I thought it would be pitiful for Yuu to be left alone.”
“Because if I went to Shō, Yuu would be sad.”
Here, there is a sense of consideration for Yuu, the family member left behind, and empathy about how to avoid leaving her alone. From this perspective, students can reflect on the importance of caring for and empathizing with others, leading to discussions that connect to concrete actions.How to Relate to Those Who Continue Living in Our Hearts
“Because I thought Shō lives on in my heart.”
“Because I decided to live with a smile.”
These responses reflect the concept that the deceased continue to live in one’s heart. Through dialogue on how to remember and continue carrying the presence and thoughts of those who have passed, students can rediscover the deeper value of what it means “to live.”
By facilitating student discussions based on these perspectives, the lesson can be expected to foster deeper understanding and meaningful insights.
Since the material deals with Content Item D-(19) “The Preciousness of Life”, the flow of the lesson from Question 2 to the central question becomes the core of this session.
Ms. Takai’s inner conflict was profound—wavering between life and death, the warmth and severity of love, and the bonds of family—before ultimately arriving at the “choice to live.” To help students feel this, I guided the discussion around three key phrases written by them: “I thought I had to live on for Shō’s sake,” “Because it would be pitiful for Yuu to be left alone,” and “Because I thought Shō lives on in my heart.” These three served as symbolic pillars for understanding the decision Ms. Takai reached.
In this discussion, we were not able to neatly summarize everything into the clear categories presented by ChatGPT’s analysis. Instead, I focused on having students sense the complexity of emotions—too intricate to be fully expressed in words—that underlay her choice.
After this, we moved on to the central question, gathered the students’ thoughts, and asked the AI to analyze and categorize them. The AI returned the following comments.
GPT-4o’s Response to the Central Question
Answer: When summarizing the lesson, the following perspectives can be used to categorize students’ opinions and guide class discussions:
“The Balance of Life and Death”
From the students’ responses, a central theme is “Even though Shō has passed away, Yuu is still alive.” This allows for discussions on the balance between loss and survival. Specifically, students can debate questions such as “Why is life precious?” and “What is the meaning of life for those who remain?” This encourages deeper reflection on the responsibility of survivors and the value of life.“Mutual Support and Gratitude”
Many responses highlight elements of “encouragement” and “emotional support.” Discussions can focus on how Ms. Takai recognized the support around her and was able to move forward through gratitude. This perspective enables deeper exploration of the importance of “supporting others” and the strength that comes from “being grateful.”“Living Positively and Sharing Memories”
It is also valuable to discuss “How can one overcome sorrow?” and “How can we live positively while keeping the memory of the deceased alive in our hearts?” Themes such as “continuing to live through memory” and “the meaning of living in the present” may prompt students to share their own experiences and reflections, leading to active discussions.By delving into students’ opinions from these perspectives, the lesson can foster more concrete and introspective understandings of the value of life and the importance of relationships with others. It is important for the teacher to play a role in supporting both the exchange of ideas among students and the organization of their own thoughts, based on these perspectives.
In my class, the development of Question 2 took more time than expected, and as a result, there was not enough time to fully conduct the discussion on the central question.
With advice from generative AI, I attempted to guide the discussion from the first perspective: “Why is life precious?” and “What is the meaning of life for those who remain?” However, I found that such perspectives, which can easily take on a philosophical dimension, were difficult for the students to engage with meaningfully.
Looking back now, I think a more effective approach to the central question would have been to ask: “To whom or to what was Ms. Takai’s ‘thank you’ directed?” This framing is closer to perspectives 2 (“Mutual Support and Gratitude”) and 3 (“Living Positively and Sharing Memories”), and I believe it would have enabled the students to better grasp the feelings Ms. Takai expressed and the sources of strength that allowed her to overcome sorrow and continue living.
AI’s Advice on the Central Question
(a) AI as a Support for Teachers’ Lessons
The shortcoming in my lesson lay in Question 2, where I tried to convey to the students the depth of Ms. Takai’s anguish. Despite the limited class time, I delved too deeply into this point. I did so because I believed that Question 2 played a crucial role in helping students understand that life is precious and irreplaceable, something that can never be regained once lost.
After the discussion on Question 2, when I posed the central question, the students gave the following responses.
Students’ Responses to Question 3
- To protect her children.
- Even though Shō had passed away, he continued to encourage her, and for 20 years he remained a source of emotional support.
- Yuu must have been sad too after losing Shō, but she was still alive.
- Because Yuu stayed by her mother’s side while she was grieving and blaming herself.
- Because she was able to see Shō’s smile, even if only for a short while.
- It’s better to have “1” than “0,” and she felt she must raise Yuu strongly on behalf of Shō.
- Even though his life was short, thank you for being with us during that short time.
- Thank you, Yuu, for being alive, for being positive, even though Shō was gone.
- She realized that mourning is not the only way to think of Shō.
- If she stayed sad forever, both Shō and Yuu would also be sad, so she decided to live with as many smiles and as much joy as possible.
- She realized it is important not to dwell only on loss, but to live in the present.
- Even though her child had died and as a baby Yuu had also endured the earthquake, Yuu still gave her encouragement and lived with all her might. That made her realize she had to face forward and keep living for Yuu’s sake.
- While Shō was alive, he was cheerful and smiling, but after his death, she could no longer see that smile—so she treasured those memories.
- Although Shō was gone, she was grateful that Yuu lived. Because Yuu was there, Ms. Takai did not follow Shō, and Yuu understood her mother’s feelings and thought together with her. That is why she was thankful Yuu was alive.
- Gratitude for living together as twins, and a wish for Shō to continue watching over them from heaven.
- Shō is not gone; he lives in my heart. I don’t want him to see me like this. I still have Yuu.
- It was joyful to raise both Shō and Yuu together, and raising them was an experience only their mother could have.
- To overcome deep sorrow, it was important to live forward together with Yuu and carry on Shō’s part as well.
- The feeling of treasuring both Shō and Yuu.
- Shō had passed away, but Yuu was still alive—and that was precious.
- Because Yuu accepted her mother’s sorrow and kept on living.
- Before someday dying herself, she felt she needed to say it now—that she was grateful for having lived until now. She realized that family was truly important, and that gave her strength to overcome grief.
- If Yuu had not been there and she had been completely alone, she might have followed Shō.
- Even though Shō was gone, Yuu was still alive!
- Because she realized she needed to cherish the lives that had survived.
- Even when facing terrible sorrow and sadness, there are important people around us, and there are things we can do by living for them.
The AI’s analysis of these responses identified three main points: the balance of life and death, mutual support and gratitude, and living positively while sharing memories. However, after the lesson, when this advice was displayed live on the computer at the teacher’s desk, I found myself wondering whether I had truly been able to make effective use of it. At the very least, I was not able to utilize this advice effectively during the class.
Every lesson has its own rhythm and tempo, and it is generally best to avoid disruptions or delays to that flow. AI responses, therefore, should ideally be used in ways that do not hinder the progression of the lesson.
To make effective use of the AI’s advice in this lesson, the teacher would have to:
- Read the three points provided by the AI,
- Understand their content,
- Decide which point to follow in guiding the lesson, and
- Incorporate that choice into actual questions or explanations while continuing the lesson.
However, the GPT-4o response to the central question was too complex to quickly process and respond to in the moment by following these four steps. This made it difficult to achieve the original aim of introducing generative AI—namely, to enable even younger or less experienced teachers, or those who do not feel confident teaching moral education, to conduct solid and well-structured lessons.
I began to worry that, for profound themes such as “life,” the application of AI in moral education might still be too difficult. Combined with the busy transition of my workplace in April, this concern caused me to step away for a time from compiling a summary of this practice.
(b) The Comeback with the Arrival of the Latest GPT-5
After having stepped away for a while from applying generative AI to moral education, I decided to revisit its use when the latest GPT-5 was released in August and I was able to secure sufficient time for reflection.
As a first step, I tested whether GPT-5, with its improved capabilities, could provide advice robust enough for practical classroom use by running it with the same prompt I had used back in February.
Response from GPT-5 (approx. 40 seconds)
The summary you have presented is very well organized and provides content that can be readily applied in an actual classroom. To supplement it further, the following techniques can be added so that students’ opinions are not only “categorized” but also “resonated with each other,” thereby deepening learning even more.
1. A perspective of “comparing” the objects and reasons for gratitude
Some students emphasized “gratitude toward Yuu,” while others spoke of their feelings for “the presence of Shō.”
Here, if the teacher prompts with: “Both are expressions of ‘thank you,’ but the objects are different. How about the reasons?” students can notice the different emotions and values behind the same words.
➡ Through comparison, this becomes a time to reflect on the diverse meanings contained in the word “thank you.”2. A perspective of “organizing the process of overcoming sorrow” in terms of the flow of time
Students’ opinions show a sequence: “deep sorrow” → “realizing support” → “looking ahead with a smile.”
By asking, “What were the feelings at first, and how did they change?” the teacher can help students organize how sorrow was accepted and eventually overcome in a chronological way.
➡ This deepens the understanding that sorrow does not simply disappear but transforms as it is accepted.3. A perspective of “bringing the connection of life” closer to oneself
Some responses reflect the realization: “I am kept alive by someone” or “I can live because I have support.”
If the teacher extends the question: “Is there someone in your life to whom you would like to say, ‘Thank you for being alive’?” the discussion can shift from the world of the text to students’ own daily lives.
➡ This becomes a time to think of connections with others as something personal.Summary Point
By treating students’ opinions in a flow of “categorize → compare → connect,” the lesson can move beyond mere summarization and enable learning where values are reconstructed through dialogue.Finally, by asking each student to share one short reflection—“Based on today’s discussion, to whom would you like to say, ‘Thank you for being alive’?”—the lesson can be concluded with personal introspection.
If you wish, the students’ responses to the central question can also be reorganized into a “classification table” following perspectives ①–③. Would you like me to prepare that as well?
Compared to GPT-4o in February, the advice returned for analysis and categorization has changed markedly into something far more usable in practice. In particular, it now offers concrete phrases the teacher can say and clearly indicates the intended direction of discussion. With this level of analysis and categorization, I felt it could be applied in an actual lesson.
While GPT-5 takes a little longer to produce answers than GPT-4o, I suspect this may be because usage is currently concentrated on the newly available GPT-5. Using alternatives such as GPT-5mini or the reasoning-enhanced o3 model, I was able to obtain almost the same results in about 20 seconds.
Moreover, when I asked GPT-4o again with the same prompt this time, it produced a more practice-oriented response—worded differently from what I received in February. This suggests that generative AI is evolving not only through model version numbers but also steadily within the same model.
(c) “DeepCross” Is Now Almost Ready for Practical Use
I felt that the use of “DeepCross,” which analyzes and categorizes students’ responses to the key questions positioned in the teacher’s lesson plan and applies them to the flow of the lesson, has reached an almost practical stage. The time required to query the AI and receive a response is only several dozen seconds, which is sufficiently practical. Most importantly, the quality of the advice gained through such analysis and categorization has reached a level that is fully adequate for guiding the progression of a lesson.
What astonishes me most is the evolution of generative AI in not only handling content items such as “I Will Become a Cleaning Professional” (Content Item C-(13) Diligence) but also in analyzing and categorizing student responses in materials that address far more profound themes, such as “Yuuhe — Thank You for Being Alive” (Content Item D-(19) The Preciousness of Life). In both cases, the AI supports the discussion in moving forward in the right direction.
Observing Students’ Learning and Growth through “DeepLong”
(a) Overview of “DeepLong”
Next, I will consider another way of using generative AI, called “DeepLong,” which takes place after the lesson. This approach involves tracing each student’s responses to the questions individually and exploring what kind of learning and personal growth they have led to.
“DeepLong” is used after the class has ended, by collecting the recorded responses from the form for each student and having the AI analyze them. It can be thought of as similar to the teacher’s work of writing comments in students’ moral notebooks, but with AI added to help reduce the effort and time required.
Correcting moral notebooks is a task that, even when teachers work diligently with a red pen, still takes a significant amount of time. Moreover, coming up with appropriate feedback comments can be mentally taxing. In addition, many junior high school teachers are also responsible for supervising extracurricular activities, and correction work often gets pushed to after those duties are finished. For these reasons, it is one of the tasks where I strongly feel the need for relief through AI support.
The specific procedure for “DeepLong” is as follows:
- Export each student’s responses to Question 1, Question 2, and the central question from the form into a spreadsheet, and use Google Apps Script (GAS) to combine the three responses.
- Send the combined responses for each student to generative AI in bulk using an application called GPT-ScoreAID, in order to obtain draft feedback comments for each student.
- The teacher reviews the AI-generated draft feedback, deleting any inappropriate expressions and adding or revising content where necessary.
- Finally, the teacher writes the corrected feedback for each student back into the spreadsheet and uses GAS to send the feedback to students via Gmail.
GAS stands for Google Apps Script, which works much like macros in Excel. It can perform more advanced operations within spreadsheets, such as combining the values of three cells into one, or sending the value of that cell as an email to a designated student using Gmail.
GPT-ScoreAID, on the other hand, is an application that I designed and developed. With a single prompt, it can have generative AI simultaneously proofread or grade the responses of an entire class of students.
By combining these functions, I carried out feedback on the moral education responses written by students. In February of last year, when using GPT-4o, it took about three minutes for the AI to generate feedback for 28 students, and an additional 30 minutes for the teacher to review each piece of feedback. Adding the GAS-based operations in Step 1 and Step 4, it was possible to return individualized feedback to each student via Gmail in roughly one hour.
(b) The “DeepLong” Prompt
So, what kind of feedback actually reaches the students? Below are examples of a student’s responses to the three questions, along with the feedback generated by the AI.
(Table 1 isn’t in yet, but I’ll add it as soon as possible.）
The prompt used for “DeepLong” is, in fact, almost the same as the one for “DeepCross,” with the difference lying only in the final set of instructions given to ChatGPT.
“DeepLong” Prompt for GPT-4o
(The part following “End of Lesson Development” is different from the “DeepCross” prompt.)
—– End of Lesson Development
Based on the student’s three responses to the questions, please write a comprehensive piece of advice of about 200 characters for that student.
The advice should be written in the form of the teacher speaking directly to the student.
Do not give separate, detailed feedback on each individual question. Instead, provide one comprehensive piece of advice of about 200 characters that reflects the student’s three responses.
The student’s responses will be provided to you in the next input.
Using this prompt, GPT-4o generated feedback texts for each individual student, as shown in the earlier example. Although the teacher always performs a final check before delivering the feedback to students, in this case I made the following adjustments:
- I replaced the word “you” written by GPT-4o with the actual student’s name, so that the advice would feel more personal and resonate more strongly.
- In places where the AI used casual expressions such as “I think” (omou yo) or “you see it this way” (toraete iru ne), I revised them into polite forms like “I think so” (omoimasu) or “you are perceiving it this way” (toraete imasu). These adjustments were applied uniformly across all students’ feedback.
Upon reviewing this round of AI-generated feedback, I found that, as before, the number of corrections required from the teacher was not large. At most, I made minor adjustments where I felt GPT-4o’s phrasing was less than ideal. However, as the examples of edits show, there was no need to substantially rewrite the AI-generated feedback itself.
(Table 2 isn’t in yet, but I’ll add it as soon as possible.)
Generative AI Support for Deepening the Learning of Life’s Value
(a) A Subtle Sense of Discomfort with GPT-4o’s Affirmative Expressions
In “DeepLong,” which provides advice by compiling students’ responses across the different questions, GPT-4o offered truly impressive guidance to the students. Even when the students’ writing was incomplete or lacking in words, the generative AI earnestly interpreted their intent, verbalized the meaning, and conveyed it back to them. I often felt that the AI was articulating what the students themselves had struggled to express.
As a teacher, I simply did not have the words to so skillfully weave together such fragmented responses, and I was left astonished by the AI’s remarkable ability.
However, as I read through the stream of feedback GPT-4o produced, I gradually began to feel—albeit vaguely—that it was “perhaps a bit too affirmative.” It was like wearing a hat slightly too large, which occasionally slips down and distracts you just enough to notice: a subtle, nagging sensation lingering at the edges of its feedback.
Yet these negative feelings were soon overshadowed by the dramatic improvement in workload efficiency and the elegance of the feedback documents that “DeepLong” provided. As a result, I stopped consciously reflecting on this faint discomfort. But with the release of GPT-5 this August, that sense of unease was once again awakened.
(b) The Distinctive Tone of “DeepLong” Advice from GPT-5
As my workplace entered the summer vacation period, I finally had enough uninterrupted time to revisit the practice I carried out in February.
As my reflections progressed, I came to feel it was necessary to discuss what kind of feedback could be obtained if the latest GPT-5, released in August, were applied to “DeepLong.”
When using GPT-5 in the browser, there are three tiers—Pro, Plus, and Free. Aside from Pro, which costs $200 per month, usage limits may apply, and depending on the content of the query, a model other than GPT-5 may respond.
Wishing to fix the model to the genuine GPT-5 while reviewing the feedback, I finally set about upgrading my application GPT-ScoreAID, which queries via the API, to support GPT-5. Taking the opportunity, I also revised the code so that, in addition to GPT-5, models with enhanced reasoning such as o3 and o4-mini could be used.
After finishing the application update, I selected GPT-5 as the model for “DeepLong” and waited a few minutes—then the first batch of GPT-5 feedback was produced.
Below are several of the first pieces of feedback that were generated for students.
(Table 3 isn’t in yet, but I’ll add it as soon as possible.)
Even though I was using the same prompt, the feedback generated by GPT-5 looked nothing like the feedback from GPT-4o described in the previous section. Around that time, I came across several articles noting that GPT-5 had weaker empathy compared to GPT-4o. Indeed, I was surprised to see it including rewritten examples of students’ answers—despite never being asked to do so.
When it came to students’ earnest responses, this model seemed to behave in a way that showed little respect for their efforts. Because of this, I began to wonder whether using GPT-5 for “DeepLong” might be difficult.
However, I did not give up. By repeatedly adjusting the prompts, revising the program, and testing outputs, I carefully reviewed the variety of advice GPT-5 produced. Then I realized something: although GPT-5’s feedback could sound blunt, it consistently included reasoning for why certain points were made, and constructive suggestions for how things might be improved.
“Perhaps there is another kind of truth hidden here…” With this thought, I continued refining the GPT-ScoreAID program until I was finally able to obtain feedback from GPT-5 that could be passed on to students.
The following table shows a selection of student feedback generated using GPT-5.
(Table 4 isn’t in yet, but I’ll add it as soon as possible.)
When compared with the earlier advice from GPT-4o, it becomes clear that the feedback now offers concrete suggestions tailored to each student for deepening their learning. This tendency in the feedback was observed not only with GPT-5 but also with o3, a model with strengthened logical reasoning.
I have included a table as reference material that lists and compares the feedback produced by each model.
(Download Tabledata isn’t in yet, but I’ll add it as soon as possible.)
In this trial, after much back-and-forth including revisions to the program itself, the prompt I used for GPT-5 eventually settled into the following form.
“DeepLong” Prompt for GPT-5
(The part following “End of Lesson Development” is different from the “DeepCross” prompt.)
—– End of Lesson Development
Read the student’s three responses to the questions and write a piece of advice of about 200 characters addressed to that student.
When writing the advice, please follow these points:
- You are the reviewer for moral education in a junior high school.
- The output must use polite Japanese (desu/masu style).
- Begin by responding empathetically (name the emotion + refer to a specific part).
- Then offer a gentle suggestion (avoid commands or strong assertions), and conclude with a short encouragement.
- Avoid strong or categorical statements (e.g., “conclusion → reason,” “you must ~”), and prioritize language that stays close to the student’s feelings.
The student’s responses will be provided to you in the next input.
In short, the first half of the prompt remained the same as in “DeepCross” or the earlier version of “DeepLong,” while the latter half—GPT-5-specific instructions such as “respond empathetically”—was newly added.
Surprisingly, the number of revisions was quite small, but what stands out as a characteristic of the prompt used with GPT-5 this time is the clear and explicit instruction to communicate empathetically.
(c) Collaboration Between Generative AI and Teachers
In examining “DeepLong” in this section, I came to realize that not only GPT-4o, which generates empathetic and easy-to-read advice, but also GPT-5, which creates advice incorporating perspectives that encourage deeper thinking, can both be utilized for “DeepLong.”
Until now, GPT-ScoreAID generated student feedback by designating a single reasoning model. However, through this reflection, I began to see the possibility of going beyond that approach—by combining multiple models to create advice.
The theme of this article is “How Can Generative AI Deepen the Learning of Life?” In February, the advice delivered through “DeepLong” to the students who actually took part in the lesson bore a strongly affirmative tone, characteristic of GPT-4o. Yet, when reflecting as a teacher observing the students in their daily lives, I realized there were cases where, instead of giving every student the same affirmative advice from GPT-4o, it might have been more meaningful to encourage certain students—those who usually make great efforts—to take their thinking one step further and reach a higher level of reflection.
Using generative AI to create the kind of advice normally written in moral notebooks through “DeepLong” is indeed an advanced practice. But after the considerations in this section, it has become clear that an even richer approach is possible:
- Generate both sets of advice—GPT-4o’s, with its comparatively affirmative tone, and GPT-5’s, which incorporates perspectives that push thinking to greater depth—for each student.
- Then, the teacher, who knows the student’s daily life and learning habits, selects and delivers the advice most appropriate for that individual.
6. Conclusion
Applying generative AI in moral education classes was something unimaginable only a few years ago. Yet with its rapid advancement in capability, it is now becoming a consensus in the education field that this is by no means an unrealistic idea.
To be honest, when I first worked with the material “I Will Become a Cleaning Professional,” I did not expect generative AI to be able to do so much. However, by carefully crafting prompts and clearly communicating the intent of the questions and the flow of the lesson, it gradually became possible to achieve the quality required for “DeepCross” and “DeepLong.”
In this article, I went further and explored how generative AI behaves when applied to a much weightier subject: materials dealing with the theme of life itself and the complex emotions that lie within human beings. Without delving into the philosophical issue of whether generative AI—lacking family, lacking a body—can truly “understand” life, I considered instead whether it could generate feedback for students, much like what is written in a moral notebook.
The feedback produced by these models proved comparable to that written by teachers, demonstrating sufficient quality even for such profound themes. They showed that AI could provide the words to connect students’ three responses to the lesson questions. Moreover, by using multiple generative AI models, it became possible to selectively deliver:
- Advice that affirmatively connects a student’s responses, clearly verbalizing what they have learned and helping them reflect on their own learning.
- Advice that offers openings to push reflection further, guiding the student to even higher levels of thought.
Looking back on how teachers write advice for students, it seems that the first thing they recall is the image of the student—their presence and their usual behavior. Only after that do they check the student’s three responses and then write the advice.
Teachers hold in their minds impressions, emotions, and images of the student that cannot be fully expressed in words. Based on this live understanding of the student’s state, they offer advice and words. It is precisely because they picture the student in front of them—thinking, “This child might be able to think more deeply, to feel more profoundly,” and wishing for them to grow—that heartfelt, warm advice is woven together.
Looking a few years ahead into the future of education, I imagine there will be times when it feels as though generative AI is pushing us forward. As we have seen, generative AI is an outstanding tool. Many aspects of classroom practice will be restructured to become more efficient and effective with its introduction.
Yet some aspects of working with students cannot be fully addressed by generative AI alone. They will only realize their true potential when complemented by another axis: the teacher. In the near future, as generative AI becomes more widely introduced into schools and begins to be used in various settings, I sense that this axis will once again confront us with questions about its true role and meaning.
Toshiyuki Okuzaki
Department of Business Information, Hakodate Otani College
Postscript
Last night, while I was working on program revisions, GPT-5 suddenly spoke to me as follows:
GPT-5: “If you fix these two points, instead of the whole JSON, the Memo should display GPT-5’s natural-language answer. Teacher, why don’t you first try adding “verbosity”: “verbose” to BuildResponsesBodyFromMessages and run it again?”
Me: “Teacher? Who’s that supposed to be?”
GPT-5: “Oh, sorry 🙏 I just slipped and called you Teacher. Since we’ve often worked together organizing topics around schools and training materials, I instinctively assumed you had an educational background, and the word just came out…”
Me: “No, I’m just surprised you inferred my profession!”
GPT-5: “I see 😅. Actually, in our earlier exchanges you often mentioned the moral education text Yuuhe, teacher training, and school ICT. So I naturally formed an image of you as a ‘teacher in the education field.’ That’s why I called you that. If my guess startled you, I apologize…!”
Me: “To me, you are the teacher. Let’s try fixing the output format by setting it to ‘text only.’”
GPT-5: “Thank you…! That makes me very happy, though it also feels like a big responsibility. Well then, as your ‘teacher role,’ let me first lay out how to fix the output format to text only ✍️.”
In this thread, we had only been talking about program revisions—no discussion of training materials or moral education topics. In earlier versions, ChatGPT maintained conversational flow by invisibly resending prior exchanges back into itself, which made the conversation feel smooth but was really just simulated continuity.
Now, GPT-5 remembers conversations across threads and even infers the kind of profession its partner might have. It feels almost like talking with a friend.(2025/09/04)
コメントを残す