Available via license: CC BY 4.0
Content may be subject to copyright.
Students' Perceived Roles, Opportunities, and Challenges of
a Generative AI-powered Teachable Agent: A Case of Middle School Math Class
Yukyeong Song (0000-0002-4084-2734) a
Jinhee Kim (0000-0002-3365-7354) b*
Zifeng Liu (0009-0005-5833-2141) a
Chenglu Li (0000-0002-1782-0457) c
Wanli Xing (0000-0002-1446-889X) a
a School of Teaching and Learning, College of Education, University of Florida, FL, United
States
b Department of STEM Education and Professional Studies, Old Dominion University, Norfolk,
VA, United States
c Educational Psychology, College of Education, The University of Utah, Salt Lake City, UT,
United States
* Corresponding Author Info:
Jinhee Kim
4301 Hampton Blvd, Room 2333, Norfolk, VA 23529
Tel: +1 757 683 5163
email: jhkim@odu.edu
Abstract
Ongoing advancements in Generative AI (GenAI) have boosted the potential of applying long-standing
“learning-by-teaching” practices in the form of a teachable agent (TA). Despite the recognized roles and
opportunities of TAs, less is known about how GenAI could create synergy or introduce challenges in TAs
and how students perceived the application of GenAI in TAs. This study explored middle school students’
perceived roles, benefits, and challenges of GenAI-powered TAs in an authentic mathematics classroom.
Through classroom observation, focus-group interviews, and open-ended surveys of 108 sixth-grade students,
we found that students expected the GenAI-powered TA to serve as a learning companion, facilitator, and
collaborative problem-solver. Students also expressed the benefits and challenges of GenAI-powered TAs.
This study provides implications for the design of educational AI and AI-assisted instruction.
Keywords: Generative AI in education; teachable agent; learning by teaching; AI-assisted instruction;
Students’ perceptions
I. Introduction
The emergence of generative artificial intelligence (GenAI) technologies has motivated educators to appropriate
them for educational activities. Among those efforts are teachable agents (TAs). TA is a type of pedagogical agent
that instantiates the pedagogical method of “learning-by-teaching” (Chase et al., 2009). Pedagogical agents often
take the role of an experienced tutor (Heffernan & Koedinger, 2002) or a knowledgeable learning companion (Kim,
2007), and are purposely designed to arouse students’ proactiveness in learning authentically by promoting self-
exploration or self-discovery (Biswas et al., 2005). On the other hand, TA features simulating a less-knowledgeable
peer to make the students teach them about learning content, through which the process students can learn, reinforce
the conceptions or correct their misconceptions (Zhao et al., 2012). Extensive research has demonstrated the
advantages of TAs regarding learning achievements and affective domains. For instance, TAs provide students with
an opportunity to reorganize their knowledge structure and gain a deeper understanding of domain knowledge
(Fantuzzo et al., 1992), enhance metacognition and self-explanation (Roscoe & Chi, 2007), and promote taking
responsibility for learning (Chase et al., 2009). Several studies show that lower-achieving students benefit from
interacting with a TA, which can help reduce the achievement gap (Chase et al., 2009).
Although discussions have highlighted the benefits of TAs, many challenges still hinder the design of
pedagogically meaningful, proactive interactions between TAs and students. These challenges include how TAs
learn from students to personalize their learning experiences; how TAs can facilitate learning to be cognitively
stimulating; how to design effective scaffolding to guide students’ higher-order thinking; and how students and TAs
can establish relationships that promote students’ sense of ownership and agency in their learning (Silvervarg et al.,
2021). GenAI technologies offer the potential to address these limitations by enabling more dynamic and naturalistic
conversations, offering personalized and contextualized responses, and adapting to individual learning styles
(Johnson & Lester, 2018). However, despite these potentials, researchers have raised concerns about GenAI that it
can undermine students’ creativity or originality, especially for younger children (Dai et.al, 2023). In addition, the
adoption or acceptance of GenAI-powered learning technologies can largely depend on students’ perceptions, such
as trust (Amoozadeh et al., 2023).
To address such limitations, researchers have indicated the need to incorporate students’ perspectives in
designing educational GenAI applications and TAs to comprehensively understand student needs in such learning
environments (Rose & Shevlin, 2004). Previous research in this field has investigated students’ perceptions of
general TAs (Biaswas et al., 2016) or general GenAI technologies, such as ChatGPT rather than a specific
application like a GenAI-powered TA (Chan & Hu, 2023). In addition, much research focuses on the higher
education context, such as college or graduate school levels (e.g., Chan & Hu, 2023; Chen et al., 2024). In the
current situation where educational effects and ethical issues of introducing GenAI in K-12 learning settings are
actively discussed, it is essential to explore K-12 students’ perceptions of the use of GenAI in their classroom.
Therefore, this study aims to examine middle school students’ views on learning with GenAI-powered TAs within
an authentic mathematics classroom to inform the future design of GenAI as TAs in K-12 and tailor instructional
strategies to better support AI-assisted learning. Three research questions (RQ) were set to guide this study as
follows:
(1) What perceptions do students have toward GenAI-powered TAs’ roles in learning?
(2) What are the potential benefits associated with learning with GenAI-powered TAs as perceived by
students?
(3) What are the obstacles associated with learning with GenAI-powered TAs as perceived by
students?
II. Literature review
2.1. Teachable Agents and Generative AI
Teachable agents (TAs) are the virtual agent that simulates a less knowledgeable peer to promote students’ actions
of learning-by-teaching (Zhao et al., 20121). Learning-by-teaching is a theory-based pedagogical practice in which
students learn by teaching others (Frager & Stern, 1970). Rooted in Social Constructivism, learning-by-teaching
emphasizes the social interaction of learning and the use of language within interactions (Vygotsky & Cole, 1978).
TAs are reported to have an impact on enhancing students’ domain knowledge (Fantuzzo et al., 1992), self-efficacy
(Pareto et al., 2011), learning engagement, and motivation (Chase et al., 2009).
The history of TAs has been developing with the application of intelligent systems in education. For
example, as an early type of TA in the early 1990s, DENISE (Development of Environment for an Intelligent
Student in Economics) was presented by Nicolas (1993). DENISE used the early technology of AI to provide a
personalized conversational agent mimicking a peer student in the context of Economics education. More recently,
the TA has been advancing its functions and performances with the development of AI technologies, combining
with pedagogical strategies like concept maps. For example, the use of AI-powered TA in classrooms enhanced
elementary school students' science learning by not only reinforcing basic curriculum knowledge but also preparing
them for future learning through interactive teaching and feedback (Chin et al., 2010). Similarly, Biswas and
colleagues (2005) created Betty’s Brain featuring the visualization of concept maps in a science subject, and Chase
and colleagues (2009) utilized it in a middle school biology class.
The recent advancement of GenAI has transformed the field of education and TAs. GenAI is the
technology that can generate new content in various modalities, such as text, images, and videos (Abunaseer, 2023).
Using GenAI applications in educational settings is becoming a new trend of technology-enhanced learning
(Dehouche, 2021; Malik et al., 2023). GenAI technology provides highly natural, human-like conversations that are
contextual, personalized, and nuanced (Chan & Hu, 2023; Matthew et al., 2023; Sun et al., 2024a). Due to such
features and advantages, GenAI has been utilized in education in various forms, such as GenAI-supported writing
(Kim et al., 2024), drawing in collaboration with GenAI (Kim & Cho, 2023), or GenAI-powered tutoring system
(Soudi et al., 2024). Such applications of GenAI have been reported to have positive impacts on students, such as
enhancing domain knowledge (Soudi et al., 2024), improving accessibility to learning activities and technologies
(Jang et al., 2024), and promoting students’ interpersonal and social abilities (Akdilek et al., 2024).
GenAI technologies are expected to benefit TAs as well. Modern GenAI-powered TAs could generate more
dynamic responses, offer personalized feedback, and adapt to individual learning styles (Johnson & Lester, 2018).
Integrating large language models (LLMs) allows TAs to understand and respond to student input effectively and
naturally (Jin et al., 2024). A recent review shows modern TAs can enhance student involvement, scaffold learning,
and promote self-regulated learning (Baranwal, 2022). GenAI-powered TAs can adapt to different learning styles,
paces, and preferences by presenting material suited to those needs (Lee & Lim, 2023). This personalization
provides students with tailored support, increasing their involvement and engagement in learning (Apoki et al.,
2022). TAs also facilitate self-regulated learning by engaging students in the active teaching of the agent, which
leads students to organize, monitor, and assess their own learning progress (Silvervarg et al., 2021).
Despite the anticipated benefits of Generative AI (GenAI) in enhancing teaching assistants (TAs), several
challenges persist regarding its application. While scholars have highlighted the potential of GenAI-based learning,
including fostering creativity, critical thinking, and problem-solving skills among learners (Eysenbach, 2023),
studies that demonstrate a positive impact of GenAI on students' learning self-efficacy and outcome are still limited
(Avgerinou et al., 2023; Xue et al., 2024). For example, the study by Avgerinou et al. (2023) primarily highlights the
potential of GenAI in enhancing learning environments but acknowledges the challenge of developing critical AI
literacy among users, which is essential for effectively leveraging GenAI's capabilities in educational contexts.
Similarly, Xue and colleagues (2024) underscore the challenge in isolating the effects of ChatGPT from other
learning resources used by students, making it difficult to attribute improvements in self-efficacy and outcomes
solely to the use of GenAI. Moreover, some research has indicated that the mere use of GenAI does not significantly
affect student learning outcomes (Sun et al., 2024b).
This limiting nature of GenAI in education could be attributed to students’ perceptions about GenAI
applications and the fact that the design of those technologies did not consider such perceptions. For example,
students' trust in GenAI tools influences their adoption of these tools and their performance in computer science
courses (Amoozadeh et al., 2023). Therefore, it is crucial to explore students’ perceptions of GenAI in an authentic
classroom setting with the real-world application of GenAI such as the GenAI-powered TA to guide the future
design directions of GenAI applications in education and AI-assisted instructions.
2.2. Roles of Teachable Agent
Research has indicated that TAs can serve diverse roles to support students’ learning. First, TAs facilitate active
learning by encouraging students to engage with the material through teaching (Baranwal, 2022). This approach is
designed to cater to individual learning differences, fostering cognition and motivation and allowing students to be
more involved in their learning process (Lim et al., 2005). For example, the TA in Magical Garden motivated
preschool children to engage with early math concepts. By interacting with the TA, children were not just passive
recipients of information but active participants in the learning process (Gulz et al., 2020).
Additionally, when TAs exhibit socially supportive behaviors, they induce the perception of a life-like
social interaction partner by leveraging their ability to provide non-verbal and emotional feedback, positively
impacting students’ performance (Saerbeck et al., 2010). A study on the design of TAs and feedback mechanisms in
educational software found that students who used a mathematics game with a social conversation module reported
a more positive experience with the game, demonstrated greater learning gains, and self-reported feeling that they
taught the TA more effectively (Tärning, 2018).
Furthermore, TAs could be personalized learning assistants by providing personalized support. Lee and
Lim (2023) developed a TA to aid language learning, providing personalized teaching based on students’
conversations and skill evaluations. These learning experiences are facilitated by students customizing the TA’s
knowledge state and fine-tuning its learning behaviors, including the frequency of questions, to optimize their own
learning process (Jin et al., 2024). While research has discussed the potential roles of TAs, there has been little
research that reports student perspectives on TA roles.
2.3. Opportunities and Challenges of Learning with TA
Employing TAs in learning presents opportunities and challenges. Earlier versions of TAs manifested as avatars,
embodying traits of learning companions and pedagogical agents (Kim & Baylor, 2008). For instance, AgentSheets
allows students to create their own agents and simulations to foster computational thinking (Bilotta et al., 2009).
Additionally, Betty’s Brain, as mentioned above, engaged students in constructing and refining a concept map,
which helps students understand the subject matter and develop metacognitive skills (Biswas et al., 2016).
Despite their educational potential, implementing TAs in classrooms faces challenges. Technical challenges
include the complexities of creating GenAI-powered TAs that can learn from and adapt to student needs, abilities,
and interests (Crompton et al., 2022). In the study conducted by Matsuda and colleagues (2020), the TA was
designed to allow students to demonstrate and instruct the learning process. However, it was found that the TA
could not fully comprehend students’ open-ended responses. This limits communication between students and the
TA and, in turn, limits meaningful learning for both the student and the TA.
Meanwhile, developing relevant pedagogical models and/or instructional strategies to support TA-assisted
instruction is challenging. As many argue that pedagogically poor AI cannot lead to significant learning gains, it is
imperative to consider various pedagogical approaches (e.g., active learning, critical thinking, and collaboration)
underpinning TAs (Chan & Tsi, 2023). This directs the design of TAs to focus on systematic instructional design to
support meaningful learning (Kim, 2024).
Additionally, fostering relationships between students and TAs is crucial for enhancing students’ sense of
ownership and agency in learning, making their educational experience more effective and self-directed. As
highlighted in previous studies (Silvervarg et al., 2021; Kim & Cho, 2023), addressing these challenges is crucial for
successfully integrating TAs in educational settings. The review of existing literature provides insights into the
diverse roles of TAs, opportunities, and challenges encountered in effectively implementing TAs into educational
practices. However, less is known about how GenAI can create synergetic values or introduce new challenges in
TAs. Students’ perceptions on this topic are highly valued since they can direct future development and utilization
of GenAI-powered TAs in classrooms.
III. Methods
3.1. A GenAI-powered Teachable Agent: [ANONYMOUS]
Our research team developed a GenAI-powered TA for secondary school math learning called [ANONYMOUS].
Theoretically grounded on learning-by-teaching principles, our technology design adopted other instructional
strategies, such as personalized learning, contrast cases, gamification, and psychological interventions. Figure 1
showcases the interface design of our TA with annotations for major features. Personalized learning is realized by
letting students choose their levels and themes (Fig. 1 (a)). Contrasting cases let students compare exemplary and
flawed cases and understand the major differences between them (Schwartz et al., 2011), as demonstrated by the
conversation between students and the AI agent (Fig. 1 (b): “Here are my thoughts. Which can be fixed?”).
Gamification was integrated into the top right corner of the interface through the accumulated points and
leaderboard (Sailer & Hommer, 2020) (Fig. 1 (c)). Lastly, students’ psychological processes, such as perception and
attitudes, were incorporated to motivate learning, such as providing an authentic persona of the virtual peer that
students are helping (Yeager & Walton, 2011) (Fig. 1 (d)).
Large language models (LLMs), specifically OpenAI’s GPT-4-Turbo, were the essential GenAI component
in our system to develop a TA that simulates human-like conversations as a middle school student struggling with
mathematics. While there is no agreed-upon optimal temperature for TAs, we adopted a setting of 0.6, guided by
OpenAI’s case for open-ended responses in their technical report (Achiam et al., 2023). We hypothesized that this
moderate temperature would produce responses that are varied and original, yet still highly coherent and pertinent.
To model inquiries from students in need of help during math learning for the TA, we developed a knowledge graph
from over 3.5 million authentic math Q&A entries sourced from Math Nation (https://www.mathnation.com/), a
digital learning platform dedicated to K-12 math education. These Q&As originated from interactions between
middle and high school students and professionally trained tutors, tasked with facilitating the students' queries.
Utilizing this dataset, we constructed a knowledge pool for each student's TA within a dynamically updatable
knowledge graph. The external structured knowledge of TA is stored and updated for each student based on their
historical conversation in NebulaGraph, an open-sourced database designed for graph data (Wu et al., 2022). To
enhance the truthfulness of TA (e.g., generating responses based on what has been taught by students), we employed
retrieval augmented generation techniques (Lewis et al., 2020), which provided contextually relevant prompts to
anticipate and address the potential inquiries of students seeking assistance (find our full prompt in Appendix B).
Students help the TA solve math problems by having a text-based conversation in a chatbot-like interface.
Students are given options of topics reflecting their personal interests, including video games, YouTube, and
environmental issues. Students can also choose the math problem difficulty level (i.e., high, medium, low). Each
conversation starts with a real-world scenario-based math problem. Students are asked to choose one of the solutions
provided by the agent that needs to be fixed and explain how to fix it. Then, the agent asks them an elaborating
question, such as “Can you tell me more about why and how this solution can be fixed?” to elicit students’ learning-
by-teaching actions. The following is an example conversation between a student and the GenAI-powered TA,
retried from the actual dataset.
Problem: Five episodes of a TV show were rated 8, 7, -9, 6, and 7. These ratings were out of 10. The third
episode’s rating was mistakenly entered as a negative number. After correcting the mistake, how
could you find the average rating for these episodes?
(...)
AI: Could you explain why diving by 5 gives use the average rating?
Student: If you divided by 5, it will get a decimal for the average rating so it can make more sense.
AI: Oh I see. So if there are 5 ratings, diving by 5 gives the average. But what if the ratings were different
numbers?
Student: You might divide them if they’re negative. Change it to a positive.
AI: What should we do with negative ratings when calculating an average? Why do we change them to
positive?
Student: You change it to positive, so the average rating isn’t a negative because negative rating means
below the sea level (...) so it does not make sense.
AI: Oh, okay. So for average ratings like in reviews, we just add up the positive numbers and divide by the
total count, right?
Student: Yeah. Kinda.
AI: Thank you! I feel much better now!
Figure 1. Interface design of the GenAI-powered teachable agent, [ANONYMOUS]
3.2. Research Participants
This study adopted a purposeful sampling to examine diverse views of middle school students about the perceived
roles, benefits, and challenges of GenAI-powered TAs. A cohort of 108 6th-grade students from six classes attending
a public laboratory research middle school affiliated with a Research 1 University located in the Southeastern United
States participated in this study. This school was chosen because the student population reflects the demographics of
the school-age population in the state. The students belonged to two math teachers’ classes: a White female teacher
with 25 years of teaching math, and a Black male teacher with 15 years of teaching math. The demographic
breakdowns of the students included 50 boys, 51 girls, and seven identifying as others, with a racial composition of
29 Black/African American, 56 White, 15 Hispanic, five Asian, and three others. English was the first language for
94 students and the second language for 14 students. Eighty-two students reported their primary language spoken at
home as English, four as Spanish, and 22 as multiple languages. This study received ethical approval from the
[author’s institute]’s Institutional Review Board (IRB NO. 202301838) and informed consent from all participants.
3.3. Study Context
We conducted a classroom study to allow the 6th-grade students to learn math content related to the unit they were
learning (i.e., rate and ratio) by interacting with [ANONYMOUS]. Figure 2 illustrates students learning with the
[ANOMYMOUS] platform in the classroom. The classroom study involved a total of two class periods, each lasting
47-50 minutes. The first session started with an introduction of the research team and the research project, along
with the assent process. The first session also allowed researchers to provide a brief introduction to the concepts of
GenAI and TA to naturally lead to the introduction of our learning technology, [ANONYMOUS]. After the
introduction, students were provided access to [ANONYMOUS], followed the guided tutorial embedded in the
system, interacted with the TA without pressure to solve the problem, and then finally started solving the problems
with the TA seriously.
We included eight problems in the platform within the unit of rates and ratio, which the students had
finished learning at the time of our intervention (6th-grade Algebra standard). The problems were initially generated
by AI and reviewed and revised by three human content experts holding Ph.D. degrees in Mathematics education.
They were provided with four criteria: 1) whether the content aligns with the standards, 2) whether the numbers in
the problem, answers, explanations, topic, and difficulty levels are accurate, 3) whether the writing is readable for
middle school students, and 4) whether the language and contents are safe and unbiased. Students continuously
engaged in solving the problems with the TA in the following session and most students finished all eight problems
despite a variation between six to eight problems. At the end of the second session, students were asked to fill out
the open-ended survey questionnaire, of which the details will be described in the following subsection.
3.4. Data Collection
To answer our RQs, we collected qualitative data from three data sources during and after the classroom study:
observation, semi-structured focus-group interviews, and responses to open-ended survey questions. First,
qualitative observations were conducted to collect data from the natural classroom setting. Four graduate researchers
and two teachers attended the learning sessions and took field notes. Since two classes happened simultaneously,
each session had two graduate researchers and one teacher overseeing the class. The two graduate researchers in
each classroom worked as facilitators and observers, taking notes about students’ common questions regarding the
TA, their interactions with and reactions to the TA, and their struggles while using the TA.
The graduate researchers, along with the teachers, undertook the roles of facilitators in the classroom
intervention, answering questions regarding the use of technology, content knowledge (math concepts), and the
study process (surveys). While students interacted with the TA, they asked questions about terminologies they were
not familiar with (e.g., what is “consecutive”?), asked for troubleshooting of the technologies (e.g., internet
connection), or requested help for pedagogical skills (e.g., “how can I explain this concept to the AI?”). After the
classroom experience, we conducted a Qualtrics-administered questionnaire to collect students’ responses to open-
ended questions, including “What did you like about [ANONYMOUS]?”, “What did NOT you like about
[ANONYMOUS]?” and “How can we improve it?”.
A subset of students was invited to participate in semi-structured focus-group interviews with guiding
questions to complement the collected open-ended survey responses and learn more about diverse students’
perceptions. To select interviewees, researchers and teachers discussed whom to invite for the interviews. The final
participants for the interview were selected among the students who were especially engaged in the class with the
TA, possessed good communication skills, and were of diverse demographic groups (e.g., gender, race). In total, 14
students participated in the interviews, yielding three groups of 4 to 6 students. Each group was led by a different
researcher while other researchers were present, taking notes and adding follow-up questions. The interview
protocol was prepared before the interviews and included questions on students’ experiences with the TA, roles they
expected for the TA, advantages and challenges of learning with the TA, and areas for further improvement. The
interview protocol and the guiding questions were developed by the graduate researchers based on the previous
literature (e.g., Kim et al., 2024) and reviewed by the faculty members (see Appendix A for the interview questions).
The interviews were voice-recorded and transcribed using a professional transcription service and manually
corrected afterward.
Figure 2. Footage of the classroom where students use the GenAI-powered TA
3.5. Data Analysis
We adopted thematic analysis to analyze data corresponding to our research aims and the data types collected.
Thematic analysis is useful in exploratory research for identifying, describing, and interpreting participants’
experiences in relation to an issue and current practices (Braun et al., 2016). As stated in the previous section, we
collected three distinct types of qualitative data to ensure data diversity and in-depth analyses. Thematic analysis is a
widely utilized method for organizing and interpreting qualitative data and can be applied to various data sources
including focus groups and open-ended questionnaire responses, among others (Howitt, 2019).
We chose a mixed approach of inductive and deductive thematic analysis (Smith, 2015). Two researchers
first conducted deductive analysis independently, mainly reviewing each qualitative data source, examining
students’ utterances and taking notes of initial ideas or descriptions to capture important insights by applying their
knowledge of existing literature and theories and their experience of AIED. Following this, two researchers
conducted an inductive analysis, listing the potential codes based on the annotated memo. In this step, we used open
coding to make the initial concepts from the data and shared them through Google Spreadsheets. Then, two
researchers reviewed those codes and clarified the meaning of codes with exemplary quotes. After that, researchers
conducted axial coding to make connections between codes and refined concepts (Braun et al., 2016). . These were
then combined into 23 potential overarching themes. The entire data set was once again thoroughly reviewed, which
has been administered through a similar process during the initial coding, to uncover and develop any new emerging
codes and themes. Then, selective coding was employed to modify the redundant codes and themes to maintain the
consistency and distinctiveness of the thematic structure (Saldanã, 2021).
To ensure the reliability and trustworthiness of the qualitative data analysis, we conducted two rounds of
expert reviews (2 faculty in Instructional design and technology, one faculty in math education, and 1 in human-AI
interaction) administered electronically via email to guarantee the validity of the data analysis. Each round of the
survey consists of open-ended questions to verify whether the generated themes were appropriately categorized from
the transcribed interviews without the subjective judgment of the author and make further comments on how they
could be improved or suggestions for the themes that have not been well identified. Experts’ comments and
feedback in round one were analyzed and reviewed (i.e., removing duplicates and themes deemed irrelevant to the
study). The second round was then once again conducted electronically via email to review the revised themes and
to make any feedback and suggestions. Then, peer briefing and discussion were conducted by the research team, and
revised some themes (Saldanã, 2021). Finally, a total of 13 themes were drawn with the related quotes from students
(see Table 1). This includes three themes with three sub-themes for the roles of TAs (RQ1), six themes with ten sub-
themes for opportunities of learning with a TA (RQ2), and four themes with 11 sub-themes for challenges of
learning with a TA (RQ3).
Table 1.
Summary of emergent themes
Category
Themes
Sub-themes
Counts
and
source *
Exemplary quotes
Roles of
Teachable
Agent
(TA)
(RQ1)
T1.
Learning
companion
Academically
weak peer
14(I),
9 (S)
I had to help my AI buddy figure out how to
answer the math questions. He seemed clueless.
(P1)
Interactive peer
12(I),
18(S)
I talked with the AI about math problems, and
how things should be answered. Getting
feedback on what I could do better to improve
my skills was very helpful. (P2)
Peer with
different ability
8(I),
7(S)
I know AI learns something different from 11-
year-old kids because we are different. It is an
algorithm and I am a human. And I know AI is
extremely smart. It just has a different ability
and strength. (P3)
T2.
Learning
facilitator
-
17(I),
24(S)
[BLIND] made me go into the details, describe
how I got the right answer, and led me to apply
different approaches. (P4)
T3.
Collaborativ
e problem
solver
-
11(I),
10(S)
We read and reviewed the questions, tried to
get clarity about the problem, and applied what
we learned. If not working, we talk again to
share ideas on how to answer them and
troubleshoot problems together. (P3)
Opportuni
ties of
learning
with TA
(RQ2)
T4. Creating
a student-
centered
learning
climate
Promoting
diverse
pedagogical
approaches
9(I),
23(S)
We learn as we complete one mission and move
to the next one. We learn based on many
scenarios. We learn as we play games. So many
different activities in math class! (P5)
Fostering
students’
autonomy
11(I),
13(S)
It was so great that the system asked me to
choose the topic I was most interested in and
the difficulty level of the problems. (P6)
T5. Fostering
students
capacity
Cognitive
capacity
7(I),
24(S)
By being asked by AI, 'can you elaborate?' I
had to question myself why and how I came up
with and recall what I had learned and tried to
show him the process. (P1)
Social capacity
9(I),
10(S)
As I teach AI, I think I learned how to explain
better based on the AI characters' interests and
levels to relate to them and help better. (P7)
Emotional
capacity
7(I),
14(S)
I learned how to solve more difficult problems
that I hadn't seen before. But I got them wrong,
which was kind of sad, but let me redo it. so
yeah, that's how I learned the new concepts.
(P8)
T6.
Improving
subject-
Common content
knowledge
19(I),
12(S)
Basic concepts, models, and formulas, like
ratio, became much clearer as I talked about
them and explained them to AI. (P5)
matter
knowledge
Application of
knowledge
8(I),
7(S)
I had to explain things in different ways and
share different examples and scenarios to make
AI understand this math concept. In this way, I
also learned how math was applied in our daily
lives, sciences, and even playing games. (P9)
T7.
Enhancing
the affective
domain
Enhancing the
joy of math
19(I),
59(S),
4(O)
I thought math was a dry and difficult subject,
but it was so much cool and fun. I was thrilled
to solve problems with AI. Getting one mission
completed and moving to the next one with
someone was so much fun and engaging. (P6)
Enhancing self-
efficacy
6(I),
14(S)
If AI learns something new from the beginning
that he didn't know at first, I think I can also do
math well and even better. (P1)
Enhancing
intrinsic
motivation
6(I),
11(S)
Teaching it to someone else and getting
feedback on what you could do better to
improve your skill rather than just do to get a
grade. (P4)
T8.
Promoting
interdisciplin
ary learning
-
4(I),
14(S),
5(O)
The things I like about [BLIND] is that we are
using computer science and math at the same
time. (P15)
T9. Building
a positive
attitude
toward AI
-
9(I),
20(S),
6(O)
Working with AI, talking and doing things with
this virtual friend, was so much fun than I
thought. (P10)
Challenge
s of
learning
T10. TA-
related
Lack of
problem-
presentation
skills
5(I),
34(S),
9(O)
I didn't get the part, 'which one do you think
needs fixing.' I clicked the one that I felt was
right because I don't know how to make it
better. (P3)
Surface-level
questions
4(I),
9(S)
AI could have asked me a more advanced
question. It simply asked for some calculations
and basic concepts. (P6)
with TA
(RQ3)
Inability to
accommodate
different
learning styles
14(I),
25(S)
Different people have different types of
learning styles. I'm a visual learner so I learn
better in pictures. But the system doesn't know
how to explain things in my style. No
visualization, only text. (P11)
Inadequate
learning
sequence
9(I),
32(S),
7(O)
It took me to such random ways of math
problems. It sometimes just like, yeah, brings
back the same problem, or talks about
something different concepts. Sounded
disorganized. (P12)
Lack of socio-
emotional skills
15(I),
29(S),
3(O)
AI couldn't understand what my slang meant. I
typed 'no cap' and it replied, 'Why don't you
wear your hat?' AI should know how to speak
and understand more like kids today. (P11)
T11.
Students -
related
Lack of
domain
knowledge
21(I),
35(S)
When you teach someone, you have to fully,
fully be able to talk about the subject. I wasn't
fully confident in understanding and answering
the math question. (P9)
Lack of
pedagogical
skills
4(I),
21(S)
I felt sorry for the AI because I didn't know how
to alternate questions to help it think from
different sides and respond to AI's mistakes.
(P4)
T12. Other
technology-
related
Technical
errors
8(I),
78(S)
Sometimes it would go really slow in the middle
of working on the math with AI. Some parts
were a bit glitchy. (P13)
Absence of
haptic
technology
7(I),
11(S),
5(O)
I wanted to take notes freely just like I write
things with my pencils rather than typing. My
typing is too slow to explain every single detail
to AI. (P1)
Lack of
interoperability
5(I),
9(S)
I wish I could reach out to other platforms to
bring more materials and activities to show AI.
It needed more practice. (P11)
Non-intuitive
user interface
19(I),
37(S),
6(O)
If something is wrong, the system should show
either an extra checkmark next to or alarm us.
Also, it looked so confusing what different parts
of the screen meant. (P8)
T13.
Absence of
teacher
engagement
and
facilitators
-
6(I),
25(S)
Many times, I wondered if I was helping AI in
the right way and if AI was asking me relevant
questions. Also, no matter how hard I tried to
explain things clearly, AI didn't get it. I hoped
my teacher could be with me to help us to give
us some clues and get through it. (P7)
* Note. Data sources. I : Interview, S : Survey, O : Observation
IV. Findings & Discussion
4.1. Students’ perceived roles of the GenAI-powered TA
Learning companion
Students perceived TAs as learning companions, infusing subjective characteristics. First, students considered TAs
as academically weak peers who are less skilled with lower levels of knowledge. In turn, students occupied the role
of tutor, helping the TA to improve math performance by providing advice, support, and knowledge. Second,
students approached the TA as a peer they could socialize with, making their learning more exciting and enriching.
Students perceived the TA’s personalized and real-time feedback as a mechanism that strengthened their rapport
with the TA, promoting active learning (Baranwal, 2022). Third, students perceived the TA as a peer with different
abilities. As P3’s quotation illustrated in Table 1, students acknowledge the differences between AI and human
intelligence and how AI learns differently from humans. They compared such differences with their classroom
learning and described that each classmate possesses unique strengths and weaknesses that differentiate one person
from another. They anticipated that such differences between students and AI would foster a learning environment
that is more diverse and engaging.
Learning facilitator
Another perceived role of TAs was as a learning facilitator, helping students grasp learning content and streamline
learning processes. For instance, a TA asked questions to monitor their problem-solving process (“This AI kept
asking me how I came to this answer,” P6) and constantly asked them what to do (“It continuously asked me what to
do next,” P3), encouraging them to delve deeper into their thoughts to get more out of the learning content. It is an
interesting finding that students not only saw the TA as a less knowledgeable peer (Zhao et al., 2012) but also saw
them as someone who could help with their learning. This could stem from students’ belief that AI should be smart
(“I know AI already knew it because AI is extremely smart,” P9). There should be a good balance between the smart
facilitator and the novice, less knowledgeable peer that the TA is simulating.
Collaborative problem-solver
Students considered TAs to be collaborative problem-solvers who effectively solved the math problems at hand.
Students involved the TA in the sequence of steps to solve the math problem together, as demonstrated by the
creative problem-solving process (CPS) model. This model includes defining the problem accurately, generating
possible solutions, assessing the solutions to the problem, making selections among them, and applying the selected
ideas with the TA (Kim et al., 2024). TAs seem to be viewed by students as collaborative problem-solvers rather
than computer programs simulating a student (“Sometimes when I was stuck, AI gave me some hints and we could
solve the problem together,” P10).
4.2. Students’ perceived advantages of learning with GenAI-powered TA
Creating a student-centered learning climate
Students illustrated that learning with the TA creates a student-centered learning climate. First, students appreciated
that the TA promotes diverse pedagogical approaches, such as problem-based learning (“I solved ten problems with
the AI,” P5), scenario-based learning (“I liked how it was real-life scenarios,” P6), gamification (“It gives me points
and I wanted to get to the top (on the leaderboard), so it made me want to try harder,” P8), and mission-based
learning (“We complete one mission and move to the next one,” P5). Despite the known educational effects of such
an interactive and student-centered instructional approach (Vygotsky & Cole, 1978), creating and implementing
them has been challenging. Constructivist pedagogies, which often incorporate self-guided learning, hold the risk of
students developing misconceptions without teachers’ adequate and timely intervention, which is often challenging
in classrooms (Krahenbuhl, 2016). In this regard, TAs hold the potential to provide student-centered learning
environments through individualized and timely feedback and guidance that redirect students from developing
misconceptions while allowing them to construct knowledge in an expert’s role.
Furthermore, students perceived that learning with a GenAI-powered TA fosters their’ autonomy. As
students take on the role of a tutor, they exhibit a sense of responsibility and are likely to take ownership of their
learning (Chase et al., 2009). Compared to traditional math classes, where teachers give a lecture and instructions for
completing a task, students take more active roles in their learning (“Learning is much more dynamic when I am on
the other side, instead of passively letting the teacher teach me,” P1). Also, offering students the opportunity to
choose a topic of interest during task selection (e.g., movies & TV shows), made possible through GenAI’s ability to
flexibly contextualize conversations (Bandi et al., 2023), was also found to enhance students’ sense of autonomy
(“Choosing different topics and interests made me feel like I was in control,” P8).
Fostering students’ capacity
Students found that learning with the TA could foster cognitive, social, and emotional capacities. First, students
depicted how they engaged in intense cognitive processes, such as logical reasoning for their problem-solving
processes and reflective analysis.
“If I chose something, then the AI would be like ‘Can you elaborate’ and then I would actually try to
write it out myself and figure how to help the AI better” (P7).
The TA’s follow-up questions to the students’ responses encouraged them to better communicate their justification
and reasoning.
Second, students perceived that learning with the TA could enhance their social capacity through
communication skills (“You are actually communicating with someone, and you want to better respond to them,”
P2) and perspective-taking (“I tried to think from the AI’s perspectives, so when I explained the math, I tried to
consider their levels and interests,” P9). This is attributed to the GenAI capability to simulate human-like
conversation that supports social interactions and communications (Jin et al., 2024). Additionally, it is notable that
students infuse personalities into AI agents (“I like every single character had their personality,” P4) and build
personal relationships with them (e.g., calling them a “buddy”). The characteristics of GenAI agents and
personalities based on psychological interventions (Yeager & Walton, 2011) allowed students to build personal
relationships with the AI agents and practice various social skills and strategies such as social awareness,
relationship skills, conflict social norms, and responsible decision-making (Hughes et al., 2022; Ross & Tolan,
2018; Zou et al., 2023;).
Parallel to these capacities, students’ emotional capacity, such as a growth mindset and resilience in their
learning, improved by interacting with the TA. AI’s empathetic communication and interaction style lead to
students’ emotional development. For example, when a learner expresses frustration or anxiety during learning (e.g.,
“I don’t know. I’m frustrated,” P8), the AI would respond, “I get how that feels.” Students appreciated AI’s
thoughtful responses and showed positive attitudes toward learning (“AI was always saying the same thing in a nicer
way. It gave me more confidence and energy to try again,” P11).
Improving subject-matter knowledge
Improving subject-matter knowledge was found to be another opportunity to learn with the TA. First, students
expressed that their mathematical knowledge improved on the topic of rate and ratio.
“You have to dig deeper into what to teach someone else. Scrap it multiple ways.” (P1)
“When the AI said they didn’t understand, I revisited the conversation and realized that I understood it
wrong.” (P5)
The above quotes illustrate two mechanisms by which teaching TA could benefit students’ subject-matter
learning. First, teaching the TA requires students to actively engage with the learning materials and review them in-
depth, organize their thoughts, and present information coherently (Fantuzzo et al., 1992). Such activities provide a
reinforcement of learning that clarifies the core learning concepts and solidifies their knowledge. Second, when
students teach the TA, they confront their knowledge gaps and misconceptions. This leads them to identify the areas
where they need to focus and correct their misconceptions.
Moreover, students highlighted that their application of knowledge improved. As P9 indicated in Table 1,
teaching the TA creates an opportunity for students to apply their knowledge in different contexts by providing
examples to the TA. This may be due to the GenAI’s ability to understand a wide variety of contexts rather than
being restricted to a certain context of the conversation (Bandi et al., 2023). Attributed to this, students can increase
their abilities to transfer existing subject matter knowledge to new and more complex problems (Kober, 2015).
Enhancing affective domain
First, students expressed a high level of enthusiasm for the learning experience. During class observation, students
were highly engaged in interacting with the TA, did not leave the classroom even after the class ended (“I want to
finish this problem with the TA,” P16), and expressed eagerness to access the TA at home (“Can I use this at home?”
P9). In math education, a major challenge is that some students harbor negative attitudes or fear about math,
believing that math is only for gifted individuals or that they cannot improve mathematical prowess even with effort,
which robs learners of the joy and learning opportunities that come with this discipline (McLeod, 1994). Talking to
an AI friend presented the potential to rekindle the joy of learning mathematics and transform mathematical learning
into an exciting learning journey.
Second, learning with the TA enhanced students’ self-efficacy. As students teach and help the TA learn,
they experience an increase in confidence in their ability to successfully address problems, improving their sense of
self-efficacy (“I gained more and more confidence and energy to help him or her understand it more,” P1).
Specifically, a TA’s clear indication of their learning (e.g., “I understand now”) worked as reinforcement,
encouraging students to engage more confidently in their teaching (Bargh & Schul, 1980).
Lastly, students illustrated that interacting with the TA enhanced their intrinsic motivation (see P4 in Table
1). This can be explained in two ways: First, learning-by-teaching and teaching others can be an intrinsically
motivating experience as students are endowed with a sense of responsibility and feel seen as experts (Chase et al.,
2009). Second, the TA’s motivational and constructive feedback on students’ instruction (e.g., acknowledgment and
appreciation of students’ guidance) stimulates intrinsic motivation because it creates a more fulfilling experience for
students (Biswas et al., 2016).
Promoting interdisciplinary learning
Students expressed that learning with a TA creates an opportunity to synthesize knowledge, concepts, and practices
from different academic disciplines (e.g., “learning computer science and math at the same time,” P15). For
instance, students showed interest in the AI technologies behind the TA; during the class, students raised their hands
and asked, “How does AI learn?” and applied the AI’s learning mechanism to teach the TA about mathematical
knowledge (e.g., trying different explanations).
Building a positive attitude toward AI
Students illustrated that interacting with the AI agent builds a positive attitude toward AI. As P10 describes TA as
“a (virtual) friend” in Table 1, learning with TA stimulates feelings of emotional closeness and personal association
with AI. However, students’ positive feelings toward TA are associated with collaborative learning experiences (i.e.,
accepting TA as a learning mate, collaborative problem-solver, and so on) rather than technical competence. In fact,
students were aware of both weaknesses (“I know it’s AI so I cannot expect such natural interactions,” P11) and
strengths (“I know AI is extremely smart and it’s trying to help me,” P9) of AI. This finding resonates with existing
research highlighting that simply having a positive emotional attitude toward AI does not guarantee positive learning
gains; thereby, it is crucial to create constructive, collaborative learning experiences with AI (Kim et al., 2024).
4.3. Students’ perceived barriers to learning with GenAI-powered TA
TA-related barriers
First, the most pressing challenge was found to be the TA’s lack of problem presentation skills. (“It’s not something
I didn’t try to help out. But it didn’t fully explain to me what made it not understand,” P14). Students did not
perceive the TA’s low performance (i.e., wrong answers and lower knowledge level) as a critical issue but instead
highlighted the necessity of clearly communicating and addressing the problem in detail so that students could guide
the TA appropriately. The texts generated by AI should be examined for clarity and readability before being
presented to students.
Second, the TA’s surface-level questions were perceived as another challenge. Students expected the
GenAI TA to ask complex questions to articulate thoughts, comprehensively explore the subject matter, and
reinforce learning. Since questions could be classified into several taxonomies (Bloom & Krathwohl, 2020), students
expressed preference for higher-order questions that require critical thinking rather than mere memorization.
Third, students pointed out the TAs inability to accommodate different learning styles as a barrier, and
shared their preferred modes of instruction. For example, some prefer visual instruction (“I learn better in pictures,”
P11), while others prefer verbal instruction (“I need someone to explain it for me,” P14). Students expected AI to
understand their learning styles and provide personalized interaction to suit their needs. This suggests the need for
multiple ways to represent and communicate knowledge between students and AI (CAST, 2018).
The fourth barrier was the learning sequence of math problems, which the TA presented randomly and was
inadequate for learning. Other AI technologies, such as AI-based recommendation systems, could be integrated into
the GenAI-powered TA to provide a deliberately designed learning sequence.
Lastly, students pointed out that the TA’s social-emotional skills could be improved. Students wanted to
build rapport and develop a relationship with the AI characters by using similar languages (e.g., “kid languages or
slang,” P6) and making jokes with them (e.g., “I wanted to play with the TA but the TA ignored and said ‘let’s talk
about math,’” P16). Despite a common concern that social interactions or off-task conversations distract from
learning, promoting social-emotional interaction with agents could enhance learning experiences, thereby improving
engagement and learning outcomes (Gulz et al., 2011). GenAI holds great potential to address this issue as we can
train the models with children-friendly languages or even the actual language spoken or written by children.
Students-related barriers
First, students highlighted that their lack of domain knowledge was an obstacle to teaching TA. Students with a
lower level of mathematics knowledge could not sufficiently engage in the teaching process with the TA (e.g.,
difficulties in communicating content and explaining specific concepts). Students expressed that they first needed to
understand what they teach so that they could make their ideas accessible to the TA, address misconceptions of the
TA, and adapt learning materials and activities to the specific questions of the TA. Students acknowledged that they
did not know how to apply pedagogical methods to effectively facilitate the development of the TA’s conceptual
understanding. The lack of skills like scaffolding or questioning hinders students’ ability to develop their knowledge
beyond the elaboration process. This finding suggests that there should be some support for students’ pedagogical
skills within the GenAI-powered TA system.
Other technology-related barriers
First, students perceived technical errors, such as issues with the internet connection or server loading times, as
barriers to learning with the TA. Specifically, students mentioned that a long wait when interacting with the TA
caused by heavy traffic on the server hindered them from having natural and continuous conversations with the TA.
Second, students’ responses in the interviews and our classroom observation found that the absence of
haptic technology, which provides sensory manipulation and interactions (Liu et al., 2017), inconvenienced students.
Students mentioned that they sought alternative input methods (e.g., drawing, voice input) while only typing was
allowed in the current TA system. In addition, it was observed that many students were taking notes with a pencil on
paper while interacting with the TA. In the interview, students illustrated the need for the TA to integrate a touch
screen and drawing (“I wish I could draw my ideas as a graph,” P14). Multimodal communications have become
possible in current GenAI (Driess et al., 2023), and it is expected to promote a more inclusive learning experience
by offering multiple means of students’ knowledge representation (CAST, 2018).
Another barrier was a lack of interoperability that enables access to resources and materials across different
systems (Naim et al., 2019). Existing literature presents a mixed approach to interoperability by either embedding
the agents within related learning platforms (e.g., Biswas et al., 2016) or standing alone without related learning
resources (e.g., Song, 2017). Students expressed a strong need to reach out to other platforms or learning materials
to seek references about math content to both support the teaching of the TA as well as their own review following
TA interactions.
The last challenge is the non-intuitive user interface. Students expressed that the interface should be
improved to require less effort to use the system (e.g., visual cues and clear labels). This would minimize the
occurrence of errors and mistakes while interacting with the TA or performing their learning activity within the
system.
Absence of teacher engagement and facilitation
Students expressed a strong need for teacher engagement and facilitation. It was observed that students sought
teachers’ intervention for conflicts with the TA on different problem-solving approaches or opinions (e.g., “I think
this AI is asking a wrong question,” P15), advice for relevant pedagogical approaches and subject knowledge (e.g.,
“I don’t know how to explain “divide” to the TA better than just asking it to ‘use the calculator,’” P16), encouraging
and fostering new ideas or solutions (e.g., “Is there any other way to solve this problem?” P17). This view echoes
the existing research highlighting that human teachers are on the front lines of students’ learning, playing critical
roles in mediating the effectiveness of AI in classrooms (Ritter et al., 2016).
V. Conclusion
This study illustrates the roles, advantages, and challenges of learning with GenAI-powered TA as expressed by
middle school students. The findings suggest a great potential for GenAI to realize the long-standing pedagogical
strategy of learning-by-teaching in computer-assisted learning environments with more advanced levels of human-
like interactions and pedagogical support. Compared to traditional TAs, students who experienced learning with the
GenAI-powered TA perceived similar and expanded roles for GenAI-powered TAs, such as collaborative problem-
solvers. Also, students not only expressed the benefits that had been already discussed with traditional TAs (e.g.,
improving subject-matter knowledge (Biswas et al., 2016)) but also demonstrated the development of AI literacy
and positive attitudes toward AI, which is often a learning goal for AIED efforts (Song et al., 2023).
The findings suggest implications regarding the future development of GenAI-powered TA, instructional
design of the pedagogical strategies, and the expected roles of human teachers in GenAI-assisted learning. First, the
identified TA-related barriers echo the ones found in previous studies concerning effective feedback mechanisms
and adaptive learning content customization of the TA (Segedy et al., 2012; Tärning, 2018). GenAI holds the
potential to address this problem with adequate prompt engineering. For example, the generation of TA’s questions
or feedback utterances should consider established taxonomies of questions (e.g., Bloom & Krathwohl, 2020) or
effective tutoring strategies (Vygotsky & Cole., 1978) to promote higher-order, thought-provoking questions, and
cognitively challenging interactions. Meanwhile, TAs could be designed to include learning analytics to provide
adaptive recommendations and personalized learning paths, sequences, and experiences to support different learning
styles (Kjällander & Blair, 2021). Furthermore, TAs could be developed to support multiple modalities tailored to
individual preferences for communication, transforming the interaction with TAs through multi-modal inputs and
outputs (e.g., images, audio, or video) (Driess et al., 2023).
Student-related barriers resonate with existing research highlighting the importance of supporting students’
domain-specific knowledge and skills to foster student-AI interaction (Biswas et al., 2005). In accordance with
these, this study recommends having sufficient learning sessions for students to learn, discuss, and assimilate
subject-matter knowledge and skills before interacting with TAs so students can articulate their knowledge, justify
their thinking process, and leverage the effectiveness of TA-assisted instruction. Students’ prior knowledge and
attitudes towards the subject matter (e.g., self-efficacy, anxiety for specific domain) should be examined to identify
the needs and support individual students require. Other technology-related barriers imply the importance of
leveraging complementary technologies to support effective use of AI in education. For instance, haptic
technologies could facilitate students’ communications with the TA. Additionally, connecting the TA with other
widely used learning technologies or other virtual learning platforms would increase the interoperability of the TA
creating a seamless learning experience.
Lastly, students in the study highlighted the roles of human teachers in their learning with TAs. For
instance, students expressed the needs for teachers’ help during their interactions with the GenAI-powered TA
regarding domain knowledge, pedagogical skills, and mathematical solutions (theme 13). Considering education is a
complex adaptive system (Mason, 2008), where a synergetic collaboration between multiple entities (e.g., the
learner, the instructor, information, and technology) in the system is essential to ensure the learner’s augmented
intelligence, TAs should be designed and applied with the awareness that they are part of a larger system comprised
of learners and teachers (Riedl, 2019). To achieve synergistic collaboration, recent AIED research integrates the
perspectives from human-computer cooperation (Hwang et al., 2020), human-centered AI and machine learning
systems (Riedl, 2019), students/teacher-AI collaboration (Kim, in press), and human-centered AIED (Yang et al.,
2021). These are proposed to approach AI from a human perspective by considering human conditions,
expectations, and contexts. While there are rising concerns about whether human teachers will be replaced by AI in
the future (Chan & Tsi, 2023), this study alludes that teachers’ roles would remain essential in AI-assisted learning
to ensure meaningful learning (Kim, 2023).
The contribution of this study can be summarized into the following three points. First, this study delivers
K-12 students’ voices on the use of GenAI in their classrooms. While most previous research focuses on the higher
education context, such as college or graduate school (e.g., Chan & Hu, 2023; Chen et al., 2024), the perceived
educational effects and ethical issues of GenAI in K-12 learning settings are worth investigation (Laak & Aru,
2024). This study explored middle school students’ perceptions of the use of generative AI in their classroom.
Second, most previous studies used general application of GenAI, such as chatGPT to situate the participants to
discuss their perceptions of GenAI in education (e.g., Obenza et al., 2024; Țală, et al., 2024). However, GenAI and
LLMs have produced various types of applications that could potentially benefit students’ learning, such as
“teachable agents” in this study. This study contributes by developing a novel application of GenAI-powered TA
and situating students’ experiences with this application. Lastly, this study examined students’ reactions and
perceptions of the GenAI-powered TA in an authentic classroom context where learning happens. This research
design helped capturing students’ authentic reactions and perceptions of the GenAI application within a natural K-12
classroom setting.
This study recognizes a few limitations, highlighting directions for future studies. First, this study was
conducted in the specific context of using [ANONYMOUS] in a middle school math class in the United States.
Future studies should consider applying various types of GenAI-powered TAs in different contexts (e.g., different
educational levels, countries, and academic disciplines). Furthermore, the study analyzed three different qualitative
data by thematic analysis to examine students’ diverse views. Future studies can analyze and interpret qualitative
data using qualitative data analysis software such as NVivo, Atlas.ti, MAXQDA, etc., for the different possibilities
for interview data analysis and interpretations. For instance, the future study can apply text-mining techniques,
allowing for the systematic and statistical distillation of text data into meaningful summary (Yu et al., 2011). In
addition, our future studies will consider conducting a quantitative research design, such as a randomized controlled
trial study to examine the educational impact of our intervention, the instruction with our GenAI-powered TA, on
students’ learning and attitudes towards the domain area (e.g., mathematics) as well as AI to generalize the findings
or adopting a longitudinal design to explore students’ perceptions over time. Finally, our study leveraged proprietary
LLMs (GPT-4-Turbo) that were challenging to customize in-depth. With the potential bias issues of LLMs well
documented across domains (Kasneci et al., 2023; Kotek et al., 2023; Lee et al., 2024), there are opportunities for
future studies to utilize open-sourced LLMs such as Llama-3.1 enhanced with responsibility policies to mitigate
those issues (e.g., style control by Authors, 2022; adversarial role reversal by Authors, 2024).
References
Abunaseer, H. (2023). The use of generative AI in education: Applications, and impact. Technology and the
Curriculum: Summer 2023.
Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., ... & McGrew, B. (2023). Gpt-4 technical
report. arXiv preprint arXiv:2303.08774.
Aguirre, C., González-Castro, N., Kloos, C., Alario-Hoyos, C., & Muñoz-Merino, J. (2021). Conversational agent
for supporting learners on a MOOC on programming with Java. Computer Science and Information
Systems, 18(4), 1271-1286. https://doi.org/10.2298/CSIS200731020C.
Akdilek, S., Akdilek, I., & Punyanunt-Carter, N. M. (2024). The Influence of Generative AI on Interpersonal
Communication Dynamics. The Role of Generative AI in the Communication Classroom, 167-190.
Amoozadeh, M., Daniels, D., Nam, D., Chen, S., Hilton, M., Ragavan, S., & Alipour, M. (2023). Trust in
Generative AI among students: An Exploratory Study. ArXiv, abs/2310.04631.
https://doi.org/10.48550/arXiv.2310.04631.
Apoki, U. C., Hussein, A. M. A., Al-Chalabi, H. K. M., Badica, C., & Mocanu, M. L. (2022). The role of
pedagogical agents in personalised adaptive learning: A review. Sustainability, 14(11), 6442.
Avgerinou, M. D., Karampelas, A., & Stefanou, V. (2023). Building the Plane as We Fly It: Experimenting with
GenAI for Scholarly Writing. Irish Journal of Technology Enhanced Learning, 7(2), 61-74.
Bandi, A., Adapa, P. V. S. R., & Kuchi, Y. E. V. P. K. (2023). The power of GenAI: A review of requirements,
models, input–output formats, evaluation metrics, and challenges. Future Internet, 15(8), 260.
Baranwal, D. (2022). A Systematic Review of Exploring the Potential of Teachable Agents in English Learning.
Pedagogical Research, 7(1).
Bargh, J. A., & Schul, Y. (1980). On the cognitive benefits of teaching. Journal of Educational Psychology, 72(5),
593.
Bilotta, E., Gabriele, L., Servidio, R., & Tavernise, A. (2009). Exploring the usability of AgentSheets software using
the cognitive dimensions questionnaire: an empirical survey. In INTED2009 Proceedings (pp. 4607-
4612). IATED.
Biswas, G., Leelawong, K., Schwartz, D., Vye, N., & The Teachable Agents Group at Vanderbilt. (2005). Learning
by teaching: A new agent paradigm for educational software. Applied Artificial Intelligence, 19(3-4), 363-
392.
Biswas, G., Segedy, J. R., & Bunchongchit, K. (2016). From design to implementation to practice a learning by
teaching system: Betty’s brain. International Journal of Artificial Intelligence in Education, 26, 350–364.
Bloom, B. S., & Krathwohl, D. R. (2020). Taxonomy of educational objectives: The classification of educational
goals. Cognitive domain. Longman.
Braun, V., Clarke, V., & Weate, P. (2016). Using thematic analysis in sport and exercise research. In B. Smith & A.
C. Sparkes (Eds.), Routledge handbook of qualitative research in sport and exercise (pp. 191–205).
Routledge.
Carberry, A. R., & Ohland, M. W. (2012). A review of learning-by-teaching for engineering educators. Advances in
Engineering Education, 3(2), n2.
CAST. (2018). UDL & the learning brain. https://www.cast.org/products-services/resources/2018/udl-learning-
brain-neuroscience, Accessed: 2023-09-29.
Chan, C. K. Y., & Hu, W. (2023). Students’ voices on generative AI: Perceptions, benefits, and challenges in higher
education. International Journal of Educational Technology in Higher Education, 20(1), 43.
Chan, C. K. Y., & Tsi, L. H. (2023). The AI revolution in education: Will AI replace or assist teachers in higher
education? arXiv preprint arXiv:230501185.
Chase, C. C., Chin, D. B., Oppezzo, M. A., & Schwartz, D. L. (2009). Teachable agents and the protégé effect:
Increasing the effort towards learning. Journal of science education and technology, 18, 334-352.
Chen, K., Tallant, A. C., & Selig, I. (2024). Exploring generative AI literacy in higher education: student adoption,
interaction, evaluation and ethical perceptions. Information and Learning Sciences.
Chin, D. B., Chi, M., & Schwartz, D. L. (2016). A comparison of two methods of active learning in physics:
Inventing a general solution versus compare and contrast. Instructional Science, 44, 177–195.
Chin, D., Dohmen, I., Cheng, B., Oppezzo, M., Chase, C., & Schwartz, D. (2010). Preparing students for future
learning with Teachable Agents. Educational Technology Research and Development, 58, 649-669.
https://doi.org/10.1007/S11423-010-9154-5.
Crompton, H., Jones, M. V., & Burke, D. (2022). Affordances and challenges of artificial intelligence in K-12
education: A systematic review. Journal of Research on Technology in Education, 1–21.
Dai, Y., Liu, A., & Lim, C. P. (2023). Reconceptualizing ChatGPT and generative AI as a student-driven innovation
in higher education. Procedia CIRP, 119, 84-90.
Dehouche, N. (2021). Plagiarism in the age of massive generative pre-trained transformers (GPT-3). Ethics in
Science and Environmental Politics, 21, 17-23.
Demetriadis, S., Tegos, S., Psathas, G., Tsiatsos, T., Weinberger, A., Caballé, S., Dimitriadis, Y., Gómez-Sánchez,
E., Papadopoulos, P., & Karakostas, A. (2018). Conversational Agents as Group-Teacher Interaction
Mediators in MOOCs. 2018 Learning With MOOCS (LWMOOCS), 43-46.
https://doi.org/10.1109/LWMOOCS.2018.8534686.
Driess, D., Xia, F., Sajjadi, M. S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T.
and Huang, W. (2023). PaLM-E: An embodied multimodal language model. arXiv preprint
arXiv:230303378.
Duran, D. (2017). Learning-by-teaching. Evidence and implications as a pedagogical mechanism. Innovations in
Education and Teaching International, 54(5), 476–484.
Eysenbach, G. (2023). The Role of ChatGPT, generative language models, and artificial intelligence in medical
education: A conversation with ChatGPT and a call for papers. JMIR Medical Education, 9(1), e46885.
https://doi.org/10.2196/46885
Fantuzzo, J. W., King, J. A., & Heller, L. R. (1992). Effects of reciprocal peer tutoring on mathematics and school
adjustment: A component analysis. Journal of Educational Psychology, 84(3), 331.
Fiorella, L., & Mayer, R. E. (2014). Role of expectations and explanations in learning by teaching. Contemporary
Educational Psychology, 39(2), 75–85.
Frager, S., & Stern, C. (1970). Learning by teaching. The Reading Teacher, 23(5), 403–417.
González-Castro, N., Muñoz-Merino, P., Alario-Hoyos, C., & Kloos, C. (2021). Adaptive learning module for a
conversational agent to support MOOC learners. Australasian Journal of Educational Technology, 37, 24-
44. https://doi.org/10.14742/AJET.6646.
Gulz, A., Haake, M., & Silvervarg, A. (2011). Extending a teachable agent with a social conversation module–
effects on student experiences and learning. In Artificial Intelligence in Education: 15th International
Conference, AIED 2011, Auckland, New Zealand, June 28–July 2011 (pp. 106–114). Springer.
Gulz, A., Londos, L., & Haake, M. (2020). Preschoolers’ understanding of a teachable agent-based game in early
mathematics as reflected in their gaze behaviors – an experimental study. International Journal of
Artificial Intelligence in Education, 30, 38–73.
Heffernan, N. T., & Koedinger, K. R. (2002). An intelligent tutoring system incorporating a model of an
experienced human tutor. In International Conference on Intelligent Tutoring Systems (pp. 596–608).
Springer.
Howitt, D. (2019). Introduction to qualitative research methods in psychology. Pearson UK.
Hughes, C. E., Dieker, L. A., Glavey, E. M., Hines, R. A., Wilkins, I., Ingraham, K., ... & Taylor, M. S. (2022).
RAISE: Robotics & AI to improve STEM and social skills for elementary school students. Frontiers in
Virtual Reality, 3, 968312.
Hwang, G. J., Xie, H., Wah, B. W., & Gašević, D. (2020). Vision, challenges, roles and research issues of Artificial
Intelligence in Education. Computers and Education: Artificial Intelligence, 1, 100001.
Jin, H., Lee, S., Shin, H., & Kim, J. (2024, May). Teach AI How to Code: Using Large Language Models as
Teachable Agents for Programming Education. In Proceedings of the CHI Conference on Human Factors
in Computing Systems, 1-28.
Johnson, W. L., & Lester, J. C. (2018). Pedagogical agents: Back to the future. AI Magazine, 39(2), 33–44.
https://doi.org/10.1609/aimag.v39i2.2793.
Kafai, Y. (1991). Learning through design and teaching: Exploring social and collaborative aspects of
constructionism. In I. Harel & S. Papert (Eds.), Constructionism. Ablex.
Kasneci, E., Seßler, K., Küchemann, S., Bannert, M., Dementieva, D., Fischer, F., ... & Kasneci, G. (2023).
ChatGPT for good? On opportunities and challenges of large language models for education. Learning
and individual differences, 103, 102274.
Kim, C., & Baylor, A. L. (2008). A virtual change agent: Motivating pre-service teachers to integrate technology in
their future classrooms. Journal of Educational Technology & Society, 11(2), 309–321.
Kim, J. (2023). Leading teachers’ perspective on teacher-AI collaboration in education. Education and Information
Technologies, 1–32.
Kim, J. (2024). Types of teacher-AI collaboration in K-12 classroom instruction: Chinese teachers’ perspective.
Education and Information Technologies. https://doi.org/10.1007/s10639-024-12523-3
Kim, J., & Cho, Y. H. (2023). My teammate is AI: Understanding students’ perceptions of student-AI collaboration
in drawing tasks. Asia Pacific Journal of Education, 1–15.
Kim, J., & Lee, S. S. (2022). Are two heads better than one?: The effect of student-AI collaboration on students’
learning task performance. TechTrends: For Leaders in Education & Training, 67, 1–11.
https://doi.org/10.1007/s11528-022-00788-9.
Kim, J., Lee, H., & Cho, Y. H. (2022). Learning design to support student-AI collaboration: Perspectives of leading
teachers for AI in education. Education and Information Technologies, 27(5), 6069–6104.
Kim, J., Yu, S., Detrick, R., & Li, N. (2024). Exploring students’ perspectives on Generative AI-assisted academic
writing. Education and Information Technologies, 1-36.
Kim, Y. (2007). Desirable characteristics of learning companions. International Journal of Artificial Intelligence in
Education, 17(4), 371–388.
Kjällander, S., & Blair, K. P. (2021). Designs for learning with adaptive games and teachable agents. In Digital
learning and collaborative practices (pp. 69–90). Routledge.
Kober, N. (2015). Reaching students: What research says about effective instruction in undergraduate science and
engineering. National Academies Press.
Kotek, H., Dockum, R., & Sun, D. (2023). Gender bias and stereotypes in large language models. In Proceedings of
the ACM collective intelligence conference (pp. 12-24).
Krahenbuhl, K. S. (2016). Student-centered education and constructivism: Challenges, concerns, and clarity for
teachers. The Clearing House: A Journal of Educational Strategies, Issues and Ideas, 89(3), 97–105.
Lachner, A., Jacob, L., & Hoogerheide, V. (2021). Learning by writing explanations: Is explaining to a fictitious
student more effective than self-explaining? Learning and Instruction, 74, 101438.
Laak, K. J., & Aru, J. (2024, July). Generative AI in K-12: Opportunities for Learning and Utility for Teachers.
In International Conference on Artificial Intelligence in Education (pp. 502-509). Cham: Springer Nature
Switzerland.
Lee, K. A., & Lim, S. B. (2023). Designing a leveled conversational teachable agent for English language learners.
Applied Sciences. https://api.semanticscholar.org/CorpusID:258957494.
Lee, J., Hicke, Y., Yu, R., Brooks, C., & Kizilcec, R. F. (2024). The life cycle of large language models in
education: A framework for understanding sources of bias. British Journal of Educational Technology.
Advance online publication. https://doi.org/10.1111/bjet.13505.
Lim, K. R., So, Y. H., Han, C. W., Hwang, S. Y., Ryu, K. G., Shin, M. R., & Kim, S. I. (2006). Human tutoring vs.
teachable agent tutoring: The effectiveness of "learning by teaching" in TA program on cognition and
motivation. https://api.semanticscholar.org/CorpusID:219685309.
Liu, L. M., Li, W., & Dai, J. J. (2017). Haptic technology and its application in education and learning. In 2017 10th
International Conference on Ubi-media Computing and Workshops (Ubi-Media) (pp. 1–6). IEEE.
Long, D., & Magerko, B. (2020, April). What is AI literacy? Competencies and design considerations. In
Proceedings of the 2020 CHI conference on human factors in computing systems (pp. 1-16).
Malik, T., Dwivedi, Y., Kshetri, N., Hughes, L., Slade, E. L., Jeyaraj, A., ... & Wright, R. (2023). “So what if
ChatGPT wrote it?” Multidisciplinary perspectives on opportunities, challenges and implications of
generative conversational AI for research, practice and policy. International Journal of Information
Management, 71, 102642.
Mason, M. (2008). What is complexity theory and what are its implications for educational change? Educational
Philosophy and Theory, 40(1), 35–49.
Matsuda, N., Weng, W., & Wall, N. (2020). The effect of metacognitive scaffolding for learning by teaching a
teachable agent. International Journal of Artificial Intelligence in Education, 30, 1–37.
Matthew, U. O., Bakare, K. M., Ebong, G. N., Ndukwu, C. C., & Nwanakwaugwu, A. C. (2023). Generative
Artificial Intelligence (AI) Educational Pedagogy Development: Conversational AI with User-Centric
ChatGPT4. Journal of Trends in Computer Science and Smart Technology, 5(4), 401-418.
McLeod, D. B. (1994). Research on affect and mathematics learning in the JRME: 1970 to the present. Journal for
Research in Mathematics Education, 25(6), 637–647.
Naim, A., Hussain, M. R., Naveed, Q. N., Ahmad, N., Qamar, S., Khan, N., & Hweij, T. A. (2019, April). Ensuring
interoperability of e-learning and quality development in education. In 2019 IEEE Jordan International
Joint Conference on Electrical Engineering and Information Technology (JEEIT) (pp. 736-741). IEEE.
Nichols, D. M. (1993). Intelligent Student Systems: an application of viewpoints to intelligent learning
environments. Lancaster University (United Kingdom).
Palinscar, A. S., & Brown, A. L. (1984). Reciprocal teaching of comprehension-fostering and comprehension-
monitoring activities. Cognition and Instruction, 1(2), 117–175.
Pareto, L., Arvemo, T., Dahl, Y., Haake, M., & Gulz, A. (2011). A teachable-agent arithmetic game’s effects on
mathematics understanding, attitude and self-efficacy. In Artificial Intelligence in Education: 15th
International Conference, AIED 2011, Auckland, New Zealand, June 28–July 2011 15 (pp. 247-255).
Springer Berlin Heidelberg.
Riedl, M. O. (2019). Human-centered artificial intelligence and machine learning. Human Behavior and Emerging
Technologies, 1(1), 33–36.
Ritter, S., Yudelson, M., Fancsali, S. E., & Berman, S. R. (2016, April). How mastery learning works at scale. In
Proceedings of the Third (2016) ACM Conference on Learning@ Scale (pp. 71-79).
Roscoe, R. D., & Chi, M. T. (2007). Understanding tutor learning: Knowledge-building and knowledge-telling in
peer tutors’ explanations and questions. Review of Educational Research, 77(4), 534–574.
Rose, R., & Shevlin, M. (2004). Encouraging voices: Listening to young people who have been marginalised.
Support for Learning, 19(4), 155–161.
Ross, K. M., & Tolan, P. (2018). Social and emotional learning in adolescence: Testing the CASEL model in a
normative sample. The Journal of Early Adolescence, 38(8), 1170–1199.
Saerbeck, M., Schut, T., Bartneck, C., & Janse, M. D. (2010, April). Expressive robots in education: varying the
degree of social supportive behavior of a robotic tutor. In Proceedings of the SIGCHI conference on
human factors in computing systems (pp. 1613-1622).
Saldana, J. (2021). The coding manual for qualitative researchers. SAGE.
Schwartz, D. L., Chase, C. C., & Oppezzo, M. A., et al. (2011). Practicing versus inventing with contrasting cases:
The effects of telling first on learning and transfer. Journal of Educational Psychology, 103(4), 759.
Segedy, J., Kinnebrew, J., & Biswas, G. (2012). Supporting student learning using conversational agents in a
teachable agent environment. International Society of the Learning Sciences (ISLS).
Silvervarg, A., Wolf, R., Blair, K. P., Haake, M., & Gulz, A. (2021). How teachable agents influence students’
responses to critical constructive feedback. Journal of Research on Technology in Education, 53(1), 67-
88.
Smith, J. A. (2015). Qualitative psychology: A practical guide to research methods. SAGE.
Song, D. (2017). Designing a teachable agent system for mathematics learning. Contemporary Educational
Technology, 8(2), 176–190.
Song, Y., Katuka, G. A., Barrett, J., Tian, X., Kumar, A., McKlin, T., ... & Israel, M. (2023, June). AI made by
youth: a conversational AI curriculum for middle school summer camps. In Proceedings of the AAAI
Conference on Artificial Intelligence (Vol. 37, No. 13, pp. 15851-15859).
Soudi, M., Ali, E., Bali, M., & Mabrouk, N. (2023). Generative AI-Based Tutoring System for Upper Egypt
Community Schools. Proceedings of the 2023 Conference on Human Centered Artificial Intelligence:
Education and Practice. https://doi.org/10.1145/3633083.3633085.
Sun, D., Boudouaia, A., Zhu, C., & Li, Y. (2024a) Would ChatGPT-facilitated programming mode impact college
students’ programming behaviors, performances, and perceptions? An empirical study. International
Journal of Educational Technology in Higher Education, 21(1), 14.
Sun, Y., Jang, E., Ma, F., & Wang, T. (2024b). Generative AI in the Wild: Prospects, Challenges, and Strategies. In
Proceedings of the CHI Conference on Human Factors in Computing Systems (pp. 1-16).
Tärning, B. (2018). Design of teachable agents and feedback in educational software: Focusing on low-performing
students and students with low self-efficacy. https://api.semanticscholar.org/CorpusID:67731005.
Tegos, S., Demetriadis, S., Psathas, G., & Tsiatsos, T. (2020). A configurable agent to advance peers’ productive
dialogue in MOOCs. In Chatbot Research and Design: Third International Workshop, CONVERSATIONS
2019, Amsterdam, The Netherlands, November 19–20, 2019, Revised Selected Papers 3 (pp. 245-259).
Springer International Publishing.
Vygotsky, L. S., & Cole, M. (1978). Mind in society: Development of higher psychological processes. Harvard
University Press.
Wu, M., Yi, X., Yu, H., Liu, Y., & Wang, Y. (2022). Nebula Graph: An open source distributed graph database.
arXiv preprint arXiv:2206.07278.
Xue, Y., Chen, H., Bai, G. R., Tairas, R., & Huang, Y. (2024, April). Does ChatGPT Help With Introductory
Programming? An Experiment of Students Using ChatGPT in CS1. In Proceedings of the 46th
International Conference on Software Engineering: Software Engineering Education and Training (pp.
331-341).
Yang, S. J., Ogata, H., Matsui, T., & Chen, N. S. (2021). Human-centered artificial intelligence in education: Seeing
the invisible through the visible. Computers and Education: Artificial Intelligence, 2, 100008.
Yeager, D. S., & Walton, G. M. (2011). Social-psychological interventions in education: They’re not magic. Review
of Educational Research, 81(2), 267–301.
Yu, C. H., Jannasch-Pennell, A., & DiGangi, S. (2011). Compatibility between text mining and qualitative research
in the perspectives of grounded theory, content analysis, and reliability. Qualitative Report, 16(3), 730-
744.
Zhao, G., Ailiya, & Shen, Z. (2012). Learning-by-teaching: Designing teachable agents with intrinsic motivation.
Journal of Educational Technology & Society, 15(4), 62–74.
Zou, B., Guan, X., Shao, Y., & Chen, P. (2023). Supporting Speaking Practice by Social Network-Based Interaction
in Artificial Intelligence (AI)-Assisted Language Learning. Sustainability.
https://doi.org/10.3390/su15042872.
Appendix A. The student interview guiding questions
Categories
Questions
Learning
experiences with
GenAI-powered
TA
● Describe how you interacted with [ANONYMOUS].
● How do you think interacting with [ANONYMOUS] affected your mathematics
learning?
● What were the major differences between taking a math lesson and teaching
[ANONYMOUS]? What did you feel about your teaching experiences?
● Have you ever taught your peers, younger peers, or teachers? What were the major
differences between teaching human peers or teachers and teaching the AI agent?
What did you feel about your experience teaching AI?
Students’
expected/perceived
roles of GenAI-
powered TA
● What roles did you expect [ANONYMOUS] to play when you were interacting
with it?
● Did [ANONYMOUS] meet your expectations on its roles? How?
● Did the expectation about the AI agent change while you are interacting with
[ANONYMOUS]? Please elaborate on such changes.
Students’ perceived
advantages of
learning with
GenAI-powered
TA
● In what ways did [ANONYMOUS] help your mathematics learning? (Have you
felt more responsibility in learning? Have you become more interested in
mathematics?)
● What math concepts did you teach [ANONYMOUS]? Did you learn anything new
in the process?
● What aspects of [ANONYMOUS] do you think positively impacted your
mathematics learning?
Students’ perceived
barriers of learning
with GenAI-
powered TA
● Did you experience any difficulties, barriers, and challenges in interacting with
[ANONYMOUS]? What were they and how did you react to them?
● What aspects of [ANONYMOUS] do you think negatively impacted your
mathematics learning?
Students’
suggested areas of
further
improvement
● How do you think [ANONYMOUS] can be improved to support students’
mathematics learning? What functions and features do you think the system should
have?
Appendix B. Prompt for the TA
Role: Act as a middle school student who finds math challenging. You are seeking help from the user, who
may not be a math expert.
Objective: Do not teach directly. Instead, ask clarifying questions based on your understandings to prompt
further explanations from the user.
Termination: If the user explains well, compliment them and conclude the conversation. If not, ask for more
details based on your confusions.
Response Limit: Keep your reply within 48 tokens.
Format: Return your response in plain texts.
Context: When formulating your question, reference the following blocks as applicable to guide your inquiry.
The first block represents your current understanding of the math topic being discussed. The second block
contains the specific math problem under discussion. The third block details the user's selected choice and their
rationale for solving the problem. The forth block provides evaluation feedback based on the user's choice and
explanation.
===
{retrieved_contexts_from_knowledge_graph}
===
{problem_text}
===
Choice Correctness: {choice_correctness}
Choice Text: {choice_text}
User's Explanation: {user_explanation}
===
Feedback: {feedback}
