How Large Language Models Can Support Learning

Baminihannadige Sunekshi Lasuni Bhareti
University Canada West
MBAR 661: Academic Research Project
Research Supervisor: Mohsen Ghodrat
June 25, 2025

2
Table of Contents
Introduction............................................................................................................................... 6
Background of the study ..................................................................................................... 6
Objective of the study .......................................................................................................... 8
Structure of the research report ......................................................................................... 9
Literature Review .................................................................................................................... 10
Use of LLMs in learning .................................................................................................... 10
Benefits of LLMs in learning ............................................................................................ 10
Identification of drawbacks .............................................................................................. 11
Recommendations and future of LLMs ........................................................................... 12
Methodology ............................................................................................................................ 12
Inclusion and Exclusion criteria ....................................................................................... 13
Explanation of the Databases............................................................................................ 14
Selection and Screening Process ....................................................................................... 15
Screening steps and selection process ............................................................................. 15
Results ..................................................................................................................................... 18
Key characteristics of the reviewed literature................................................................. 18
Key Applications of LLMs in different Learning Environments .................................. 18
Benefits of using LLMs in learning .................................................................................. 19
Limitations and ethical concerns ...................................................................................... 20
Theoretical frameworks supporting the study ....................................................................... 21

3
Discussion of the findings ...................................................................................................... 24
Introduction of AI and its broader application in Learning and Education ........................ 24
Relationship between AI and LLMs ................................................................................ 26
Overview of LLMs ................................................................................................................... 26
Origin and Development of LLMs ................................................................................... 26
LLMs and their sources of knowledge ............................................................................. 27
LLMs in Education ................................................................................................................. 28
Role of LLMs in different modes of learning ........................................................................ 29
Application of LLMs in in-person classrooms ................................................................ 30
Application of LLMs in online education ........................................................................ 31
Impact on leaners: ........................................................................................................... 31
Impact on educators: ....................................................................................................... 32
Application of LLMs in self-directed learning ................................................................ 33
Relationship to the theoretical frameworks .................................................................... 34
In-person classrooms ....................................................................................................... 34
Online education .............................................................................................................. 34
Self-directed learning....................................................................................................... 35
Benefits of using LLMs in learning and education ............................................................... 36
Personalized Learning Experience ................................................................................... 36
Accessibility, and inclusion ............................................................................................... 37
Efficiency and scalability................................................................................................... 38
Promoting lifelong learning .............................................................................................. 38

4
Increased motivation, engagement and enjoyment......................................................... 39
Fostering collaborative learning ....................................................................................... 39
.................................................................................................................................................. 40
Limitations of using LLMs in learning .................................................................................. 40
Limitations on complexity and clarity of content ........................................................... 41
Lack of empirical evidence................................................................................................ 41
The issue of over-reliance .................................................................................................. 42
Lack of digital literacy ....................................................................................................... 43
Limitations and biases in training data ........................................................................... 43
Data privacy concerns ....................................................................................................... 44
Issues related to authorship .............................................................................................. 44
.................................................................................................................................................. 45
Recommendations to mitigate the effects of limitations ........................................................ 45
Introducing relevant socio-legal frameworks .................................................................. 45
Fine-tuning LLMs .............................................................................................................. 46
Developing prompt engineering skills .............................................................................. 46
Policies to enhance academic integrity ............................................................................. 47
Enhancing digital literacy ................................................................................................. 48
Increase empirical studies ................................................................................................. 48
The future of LLMs in Learning; Emerging trends and possible trends ............................. 48
Conclusion............................................................................................................................... 50

5
References ............................................................................................................................... 53
Appendix 1 ............................................................................................................................... 60
Appendix 2 ............................................................................................................................... 70

6
Introduction

Background of the study
The educational systems have been transformed with the integration of AI and its
subsets. Over the past decade, AI powered LLMs have been developed and its application has
been expanded across various aspects of education. OpenAI models such as GPT series,
Google Gemini and many more tools have taken over both the teaching and learning
processes proving to be very effective in offering numerous benefits that bridge the gaps of
the traditional education system.
LLMs are a significant improvement of advanced AI systems which are trained on
massive data sets and consist of powerful Natural Language processes and generative skills
(Imaran & Almusharraf, 2024; see also Shahzad et al., 2025). With the introduction of
transformer architectures in 2017, LLMs such as BERT, & GPT series have proven greater
standards of NLP offering text analysis, interpretation, and content generation (Yan et al.,
2024; see also Shahzad et al., 2025; Imran & Almusharraf, 2024). With such improvements
to the LLMs it has paved the way for their vast application in a range of learning and
education activities going beyond content creation to grading, providing feedback and
creating assessments (Yan et al., 2024; see also Imran & Almusharraf, 2024). The improved
integration of LLMs in education has been able to address many limitations and challenges
posed by traditional learning systems such as heavy workloads of educators, inefficiencies in
assessments, lack of customized learning plans, and inequitable educational access (BuceaManea-Tonis et al., 2022).
The existing literature on the impact of LLMs on learning has highlighted the key
advantages they offer to both educators and learners. Accordingly, LLMs have helped the
quality of education to improve through real-time feedback, convenient and automated

7
grading, Academic profiling, recommendation of resources, generation of curriculum
activities and assistance provided in collaborative activities (Yan et al., 2024; see also
Shahzad et al., 2025; Ruiz-Rojas et al., 2024). For example, GPT-3 and its succeeding LLMs
are currently generating assessment questions, grade them, and provide real-time feedback.
Further, models like Google Gemini go beyond just text generation but also create images,
handle audios and videos and therefore offer more enhanced learning experiences, especially
suiting the different needs, capabilities and backgrounds of students (Imran & Almusharraf,
2024).
One cannot number the advantages offered by LLMs in learning. However, one of
the most discussed and highlighted benefits is personalized learning offered through the
integration of LLMs. LLMs are able to curate the learning experience to match the pace,
complexity, and capacity of students and provide scaffolding where necessary (Shahzad et
al., 2025; see also Praveena & Anupama, 2025; Ruiz-Rojas et al., 2024). Further, LLMs are
greatly appreciated by the existing literature for promoting collaborative learning and
improving critical thinking skills. LLMs such as ChatGPT can be leveraged as virtual
moderators fostering different perspectives and enhancing debate in different group work
settings and therefore creating space for peer feedback (Ruiz-Rojas et al., 2024).
Despite the valuable advantages brought in by the LLMs, their application can give
rise to many practical limitations and challenges. Most of the LLMs are still in early
developmental stages and digital illiteracy of both educators and learners can prevent them
from achieving the expected outcomes. Most of these models have not been tested in realworld classroom settings and therefore, lack of transparency in those models can raise doubts
about replicability and accessibility (Yan et al., 2024). Additionally, overreliance and
dependence on LLMs can hinder the critical thinking skills, creativity and individualism in
the work produced by students (Baltaci et al., 2024). Furthermore, deployment of LLMs in

8
the educational setting has given rise to significant ethical concerns as well. The existing
literature has highlighted the biases resulting from the data on which LLMs are trained and
therefore, reinforcing stereotypes and social inequalities, concerns relating to data protection,
authorship of work generated, and intellectual property rights issues. Accordingly, they have
emphasized the importance of human-integration and judgement to balance the drawback of
the LLMs.
Looking at the current state of the existing research, we can see that many authors
have focused on discussing the benefits offered by the LLMs in learning and how they
support different aspects of education. However, most of the literature consists of small pilot
studies or laboratory testing rather than actual deployment of LLMs in broader classroom
settings. It can be seen that most studies have highlighted some general recommendations to
improve the future of LLMs in learning but have not dived into the specifications or strategic
approaches to address the limitations and challenges posed by LLMs. The availability of
limited knowledge and resources can be one of the main reasons for this gap in the research
and focus should be shifted on to discussing a human-centred approach for integrating LLMs
in education. Such approach will include stakeholder involvement from development to
implementation stages in order ensure the right standards, transparency, and policies to
promote ethical use and equitable access to learning powered by LLMs (Yan et al., 2024; see
also Imran & Almusharraf, 2024; Bucea-Manea-Tonis et al., 2022). The ultimate goal should
be to reflect on LLMs as augmenting the experience of educators and learners but not
replacing them, supplementing human thinking, judgement, creativity but not supplanting
those skills.

Objective of the study
This research is focused on addressing the following objectives:

9
● To analyze the impact of LLMs on supporting learning processes, particularly
through interactive feedback and tutoring.
● To explore methods for integrating LLM-based tools in educational settings to
improve learning outcomes.
● To identify potential improvements in LLM feedback systems to foster deeper
learning and critical thinking.
Further, the below questions are intended to be explored by this research:
● How can LLMs be effectively integrated into learning environments to
enhance feedback and student engagement?
● What are the specific features and techniques of LLMs that could support the
goals of education?
● How does the use of LLMs in educational contexts impact students’ critical
thinking and problem-solving skills?

Structure of the research report
The first section of the discussion section draws the attention of the reader to a brief
introduction of AI and its broader application in learning and education. While this part
mainly describes how AI in general have impacted the education sector, it will also briefly
outline the relationship between AI and its subset of LLMs. The next section will narrow the
focus of the research exclusively to LLMs, starting off with an overview of LLMs, their
origin and development, main sources of knowledge and functionalities. Next, the focus will
be shifted to the application of LLMs in education and discuss the existing literature on the
role of LLMs in different types of learning methods. The subsequent section will outline the
benefits and impact of using LLMs in learning, following with a section emphasizing the
limitations and ethical concerns identified by the existing literature in relation to the use of

10
LLMs in educational settings. The final section of the report will be devoted to discussing the
recommendations that must be implemented to mitigate the effects of challenges and
limitations posed by LLMs and the future of LLMs in learning.

Literature Review
Use of LLMs in learning
The integration of AI and most specifically, LLMs has transformed the traditional
approaches to education and fostering wider space for innovative solutions for prolonging
challenges posed by traditional learning systems. The existing literature has highlighted the
development of LLMs from their original state of probabilistic models to more refined
transformer-based architectures such as GPT-4 and Google Gemini. These sophisticated
models are currently being used in varying learning settings for generation of content,
interactive tutoring, generation of real-time feedback, student engagement and many more
advantages. Studies undertaken by Shahzad et al. (2025) and Yan et al. (2024) have
emphasized how LLMs have supported learning and enhanced the experience of both
educators and learners through personalized feedback, personalized assessment processes,
and adaptive delivery of content. Bewersdorff et al. (2025) building on the same aspects have
highlighted how multimodal LLMs such as Gemini go beyond the basic functionalities and
process images, audios, and videos to offer a richer and more engaging learning experience
by drawing examples from science education.

Benefits of LLMs in learning
Speaking of the benefits offered by LLMs in the learning environment, Research by
Ruiz-Rojas et al. (2024) and Richer et al. (2025) have highlighted the ability of LLMs to
deliver complex concepts in simpler manners to suit the learning capacities of different

11
learners, offer real-time assistance and therefore providing scaffolding where necessary. The
LLMs are therefore able to foster inclusive education to support multilingual learners and
widen accessibility to learners with disabilities or coming from distinct backgrounds and
learning proficiencies (Wu et al. 2025). Sarangi et al. (2024) has further focused on the
flexibility offered by LLMs through 24/7 availability. Researches such as Belkina et al.
(2025) and Lang et al. (2024) have highlighted the importance of LLMs in enhancing
efficiency by automating most of the routine and administrative tasks that take up a
considerable amount of time of educators that can be alternatively used for curriculum
innovation and mentoring of students. Shen et al. (2025)’s work on LLMs have focused on
how LLMs foster collaborative learning environments by playing the role of a ‘third
collaborator/moderator’ in group projects. Studying the interactions of nursing students with
LLMs, Shen et al. (2025) has observed how they keep on comparing the insights generated
by LLMs against the perspectives of their peers to improve the nursing care plans.

Identification of drawbacks
The existing literature has however identified the drawbacks in the LLMs and the
functionalities they perform. Yan et al. (2024), carrying out excellent research on this matter
has confirmed that LLMs tend to create false information when they are not able to generate
relevant and accurate information and that such hallucinations pose a risk at academic
integrity. This is just one limitation of LLMs that standout along with the rest of the issues
discussed by existing literature. Researchers have also shed light on the ethical concerns
surrounding the usage of LLMs in learning and Nguyen Van Viet et al. (2024) and Claman &
Sezgin. (2024) have highlighted the risks of giving rise to systemic biases and social
inequalities through the content generated by LLMs due to the broader biases in their training

12
data. Erkılıç & Ibrahim Cifci (2024) and Guizani et al. (2025) have further discussed
authorship and attribution of intellectual property rights of work generated by LLMs.

Recommendations and future of LLMs
The existing literature has also identified the suitable solution for the discussed issues
and Lee & Palmer (2025) have highlighted the importance of building institutional policies
related to the use and standards of LLMs. While Wang et al. (2021) has emphasized the
importance of enhancing prompt engineering skills of the LLM users, Yan et al. (2024) has
looked for solutions that focus on creating empirical validity of the efficiency of LLMs
through application in actual classroom settings.
Finally, the existing researchers have shed light on the potential role of LLMs in the
future paving the way for more efficient advancements in education. While Al-Dokhny et al.
(2024) has thoroughly discussed how multimodal LLMs can improve student engagement
and the quality of learning experience through more interactive learning, Cain (2025) has
envisioned the future role of LLMs as a cognitive partner rather than mere source of
generating content.
In summary, the existing literature has underlined the transformative nature of LLMs
through their numerous benefits offering to the learning environments. However, the
researchers have also carefully navigated through the limitations of LLMs while identifying
the suitable solutions. The common understanding of these studies is that LLMs must be used
in learning as a tool to compliment the skills of both educators and learners but not to replace
them.

Methodology

13
The methodology that was used to develop this report is a systemic literature review
to analyze how LLMs are supporting the learning and drink experience related to both
students and educators. The approach we have adopted has paved the way for identifying,
selecting, evaluating and integrating the relevant academic literature covering the topic to
build a robust and evidence-based understanding of different aspects of the topic.
In order to ensure that the search for the relevant academic journals was broad but
focused, major academic databases were used. Appendix 1 presents the keywords, synonyms
and the relevant search strings that were used in combination to search for the appropriate
and relevant academic studies. Further, the search was limited to peer reviewed articles
published in the past five years to ensure that the most recent discussions and developments
were taken into consideration as the field of AI and LLMs is constantly evolving.

Inclusion and Exclusion criteria
The following inclusion criteria was used to select the academic articles:
§

Type of publication – For enhanced validity, only peer reviewed academic articles
were selected for the research.

§

Publication date – in order to promote contemporary research and to capture the
most recent developments, the academic articles published in last five years were
selected.

§

Relevance – Academic articles covering the application of AI and LLMs in different
methods of learning, benefits of use of LLMs in learning, limitations rising from the
LLMs in the context of education, recommendations to mitigate the effects of
limitations and future of LLMs in the context of learning.

§

Language – Academic articles published in English was used for the study.

14
Below lies the exclusion criteria that was used for the search of the papers:
§

Focus was restricted to peer reviewed academic articles and therefore, book chapters,
conference papers, reports, dissertations and theses and other opinion pieces were
excluded from the search.

§

Articles covering general aspects of LLMs or not solely focused on the application of
LLMs in learning were excluded.

§

Articles that were not peer reviewed or not published within the last five years were
excluded from the search.

Explanation of the Databases
A comprehensive approach was adopted to access a wide range of scholarly articles
through different academic databases. The Publication Finder for University Canada West
was used as the main path of access to a number of subscribed resources to make sure that
access was allowed to peer reviewed, valid, and relevant journal articles. Directory of Open
Access Journals was a main database that was used to have access to peer-reviewed open
access journals. Another data base that was used is the Google Scholar to search for many
introductory and preliminary studies that covers many academic formats. Academic Search
Ultimate was another multidisciplinary database that was heavily used in the search process
as it allows access to almost every field and their studies. This database was proven to cover a
wide range of journals and extremely helpful to ensure that the search was done extensively.
For more general areas such as the introduction, origin and development of AI and LLMs,
JSTOR proved to be helpful in providing access to primary resource articles. Sage Open was
a similar platform but introduced more extensive studies done in different fields and peer
reviewed.

15
Further, few specialized databases were also used to narrow the search and improve
relevancy. ScienceDirect was one such specialized database that was used to gain access,
particularly through Elsevier, to studies done in the science and medical field and the
application of LLMs in those fields. Another key database that was used is PubMed to gain
access to articles related to biomedical education, life sciences and other medical studies. The
agreement to use a multi-database strategy ensured that a more robust, recent, and
comprehensive set of journal articles were accessed to successfully complete the research and
satisfy the research objectives.

Selection and Screening Process
The articles were carefully scrutinized to extract the relevant data and these data were
categorized based on the relevant subtopics such as key themes and methodologies,
application of LLMs in different types of learning, reported benefits, discussed limitations,
suggested solutions. The extracted data from all the articles were then combined to build a
comprehensive review of the research topic and relevant sub-topics. The apparent strengths,
and weaknesses in the existing literature and the methodologies these studies have deployed
were analyzed top identify the gaps in the current research to recommend practical solutions.
Screening steps and selection process
The screening steps utilized in the process of selecting the most relevant journal articles
from the many databases are listed below:
1. Initial searches through keywords
The selection process was started by conducting initial searches across the many
databases identified in the preceding section. The search was based on the keywords,
their synonyms and search strings created from these keywords as displayed in

16
appendix 1. The keywords and search strings were formulated based on the key
themes, concepts, research questions and objectives used for the research study.
2. Screening of titles and abstracts
Upon gaining access to a vast number of journal articles, the titles and abstracts of the
articles were screened with the help of Notion to see the potential relevance. Articles
were taken into the next stage of the study if they discussed about the application of
AI in educational contexts, particularly LLMs, their benefits, limitations, and future
trends. Any articles that discussed about AI or LLMs in general without specific focus
on educational application were excluded at this stage.
3. Full-Text review
The next step was reviewing the full text of the selected articles to make sure that they
were within the inclusion criteria. The Notion tool was of assistance at this stage to
thoroughly study the contents of the selected articles and exclude any article that did
not fit the inclusion criteria.
4. Extraction of Data and Synthesis
Upon reviewing the full text of the articles, around 52 articles were short listed, a
summary of which is included in Appendix 2. The data supporting the research study
was systematically extracted from these articles and this included information
covering the main theses of the study such as recurring themes, emerging trends,
challenges, proposed solutions in the application of AI and LLMs in different
educational contexts.
This whole process has ensured that the methodology of this study is transparent
rigorous and well-aligned with the objectives of the research. Table 1 below includes the
PRISMA flow diagram for ease of understanding of the numbers that were screened in the
selection process.

17
Figure 1
PRISMA Flow diagram of the selection process.

The characteristics of the articles included in the study can be categorized according
to the research methodology that has been used in each of these articles. This will further
provide a clear view as to the type of the studies contributing to this research paper. Figure 2
illustrates this categorization, and it can be seen that close to half of the studies are
conceptual reviews and only 21.15% of the studies carry empirical evidence.

Figure 2
Distribution of research methodology of the reviewed articles

19
visualizations, assist multimodal science learnings, help educators to curate assessments in a
way to enhance student engagement, critical thinking, and real-time learning.
Most of the literature has identified the application of LLMs in online learning
experiences through many activities such as gamified quizzes, and LLM powered interactive
activities. Further, these educators using these online environments are currently using LLM
powered tools to obtain detailed feedback on student performances, to grade assignments and
curate content according to the varying abilities of students.
Finally, self-directed learning is also being identified as a growing method of
learning, especially in higher educational settings, and the existing literature has highlighted
the important role played by LLMs in promoting autonomy, and independence through
different tools used to generate ideas, find resources, structuring and evaluating assignments
and obtain feedback.

Benefits of using LLMs in learning
Some key benefits have been identified by most of the reviewed literature and these
individual benefits fall under few common categories. Nearly all the articles have discussed
how LLMs enhance personalized learnings that best suit the different needs of students.
These studies have also highlighted how the educators are also benefited by such
personalization to identify the many different performance levels of students and cater
accordingly.
A considerable number of articles have also highlighted how LLMs have eliminated
the barriers of traditional learning systems through facilitating wide access for learners with
disabilities and removing language barriers through multimodal and translating tools. The
24/7 availability has also been highlighted frequently as supportive in increasing flexibility.

20
Studies have also repeatedly highlighted how LLMs have increased efficiency
through automation of routine tasks and promoted life-long learning through independent
platforms without any locational constraints.
Almost all the studies have highlighted how student engagement is enhanced through
the integration of LLMs in learning experiences, especially through personalized feedback
which mitigates comparison within student groups and increases the identification of
individual learning capacities.
Finally, a key aspect the has been discussed in several studies is how LLMs improve
collaborative learning and plays the role of a moderator in learning settings, especially in
group projects, by feeding into critical discussions among peers and refining their outputs.

Limitations and ethical concerns
The selected literature has identified recurring limitations, including ethical concerns
surrounding the use of LLMs in education.
Complexity in the content generated has been highlighted by several studies as
resulting in lack of clarity especially in learning settings of younger students. The issue of
creating incorrect information or ‘hallucinations’ is also constantly discussed in majority of
the literature.
Further, over-reliance on LLMs has also been discussed in several articles to show
how it can reduce the critical thinking skills, and creativity among students and reduce
authenticity and individualism in the work produced by them.
Another issue highlighted is the lack of digital literacy of both students and educators
that leads to obstructions in obtaining highest levels of effective utilization of LLMs.
Turning to the ethical concerns, the existing literature has discussed the systematic
biases resulting from the existing biases in the training data and how it leads to reinforcement

21
of stereotypes. Further, studies have shown how potential leakage of private data of students
can lead to significant privacy concerns. Finally, the issue of authorship of content generated
by LLMs and the related issues surrounding intellectual property rights have also been
identified by several studies.

Theoretical frameworks supporting the study
This report analyses the application of LLM in the context of learning and education.
The objective is to identify the benefits of such application, what limitations and challenges
they pose and what are the practical solutions to overcome these challenges. However, before
the resources can be analyzed to discuss the findings, there are few important theoretical
frameworks that we need to lay down. These theoretical frameworks assist the research in
laying a foundation for the main concepts involved in the study and to ensure that the scope
of the study is broadened yet focused and not strayed. They also provide validity to the
concepts discussed in the report and a sense of origin or understructure to the relevant topics.
The focus of this study is narrowed to the application of LLMs in ‘learning’.
Therefore, in order to understand and restrict the scope of ‘learning’ we discuss here, we have
used Diana Laurillard’s conversational framework (LCF) to guide us. The LCF lays down a
structure for including different learning experiences and emphasize the importance of
collaborative learning and engagement. The focal idea of this concept is that the learner is
central in the learning environment and the development occurs as a result of interactions
between the educator and learner and amongst the learners (Yu et al., 2020). This framework
identifies six different learning types that optimize these interactions (Yu et al., 2020):
1. Acquiring – Refers to the knowledge gaining activities such as reading from
resources (books, websites) or listening to lectures or watching educational
videos.

22
2. Inquiring – this step refers to the stage where learners formulate questions,
explore, compare and contrast the concepts acquired to widen their understanding
of the knowledge gained.
3. Discussing – this is the process where learners share their opinions, ideas,
concerns, and engage in conversations with the educators and peers to express
their thoughts.
4. Practicing – in this stage, learners utilize different tools such as feedback, selfreflection, external sources to evaluate their learning and refine their
understanding or output.
5. Collaborating – This refers to act of collaborating with peers to create shared
outcomes, such as reports, papers, designs, diagrams and illustrations. This
involves the practice of negotiating and re-creating work until a solution that is
agreed by everyone is achieved. This process involves challenging ach other’s
work to find the best outcome.
6. Producing – this is the process where learners utilize what they have learnt for the
generation of a new output to that is evaluated by the educator.
Figure 3
Diana Laurillard’s conversational framework (LCF)

23
Further, in order to restrict our understanding on the objectives of learning, we have
Bloom’s taxonomy that describes six levels of learning objectives. Accordingly, the first level
is ‘Remember’ which refers to recalling facts, information an d definitions learnt. Second
level is ‘Understand’ which refers to the skill of comprehending the meaning of concepts and
information acquired (Eber & Parker, 2007). The learner should be able to paraphrase the
information learnt to express the skill of understanding. The next objective is ‘Apply’
(Adams, 2015). This level requires the utilizing of knowledge for new situations to find the
best solutions in a given scenario. Fourth level is ‘Analyze’ and this refers to dissecting
concepts into sub components and understanding how each component relates to the other
and compare and contrast between the concepts (Adams, 2015). The next is ‘evaluate’ that
refers to making judgements related to the ideas learnt and evaluating facts, ideas and
information learnt against given standards (Adams, 2015). The final level is ‘Create’ and this
means the generation of new knowledge or outcomes with the knowledge acquired and
putting different elements together to build a new creation (Eber & Parker, 2007).
While the application of LLMs in learning will be studies in the context of different
types of learning, their benefits and limitations will be discussed in the context of their impact
on the achievement of objectives of learning.
Figure 4
Bloom’s taxonomy

24
Discussion of the findings
Introduction of AI and its broader application in Learning and Education
Modern technology and its advancements have proven the incorporation of
technological tools in many industries. One such important field where technology has
offered numerous benefits and enhanced quality and experience is education and learning.
Artificial Intelligence (AI) has taken the forefront of the progression of technology and it has
proven a remarkable transformative effect in learning by opening unprecedented
opportunities to bridge the gaps in traditional education systems (Reicher et al., 2025). AI is
considered a critical component of the digital transformation process of education that
facilitates agile solutions for educational market needs (Bucea-Manea-Tonis et al., 2022). AI
has provided reliable and effective solutions to challenges posed by traditional learning
systems such as gaps in pedagogical alignment, lack of self-paced learning and inadequacy of
inclusivity of diverse learners (Reicher et al., 2025). It is said that broader Sustainable
Development goals are intended to be achieved through the integration of AI in education and
therefore promoting quality of education and life-long learning (Ruiz-Rojas et al., 2024). One
may believe that AI supports learning only through automation. However, it has redefined
learning systems through different approaches by offering personalized instructions for
educators that suit distinctive needs of students (Ruiz-Rojas et al., 2024). Creating more
adaptive systems with innovative learning and assessment tools, and empowering both
educators and learners have been the central focus of AI advancements in education (Ajani et
al., 2024). AI’s position or purpose in education is best reflected by stating that it does not
replace the educators but augment their performance (Reicher et al., 2025). AI is believed to
have shaped the critical thinking of learners, promoted teamwork and collaboration through
platforms facilitating virtual meeting spaces, fostered inclusivity and enhanced problem
solving (Ruiz-Rojas et al., 2024).

25
AI utilizes its key domains such as natural language processing (NLP), machine
learning, immersive technology, IoT, and conversational agents in education and for better
creativity, socio-emotional intelligence and cultural competence in learning, these tools must
be functioned with human-led instructions (Praveena & Anupama, 2025). Both the learning
experience of students and teaching techniques of educators are advanced through different
forms and uses of AI. Most educators that are technophobic have embraced AI tools for ease
of preparing lesson plans, curating assessments, simplifying the grading process and
streamlined administrative tasks (Ajani et al., 2024). Further AI tools have helped the
educators to analyze student data and identify the needs for intervention and improvement
and thereby making room for personalized and real-time evaluation (Ajani et al., 2024).
Educators are also using AI for creating educational escape rooms, mind maps,
summarization of length texts, all of which assist student engagement and enhance the quality
of educational content transferred to students (Ruiz-Rojas et al., 2024). On the other hand,
students are also massively benefitted from AI platforms that offer diverse perspectives and
enhance critical thinking through real-time feedback, offer access to guidance that is
available all the time and provide content and pace for different learners with individual
capacity (Ruiz-Rojas et al., 2024). Students can use AI platforms to assess, analyze and
summarize course materials and therefore, create efficient and personalized content for
studying and assignments (Reicher et al., 2025). AI has been able to reshape the outcome of
education as it enables the identification of varying capacities of students and curate the
delivery of knowledge accordingly.
This study will focus on a specific domain of AI, namely Large Language Models
(LLMs) and how they support learning. The introduction about the broader application of AI
in learning has laid the foundation to identify the more specific application of LLMs in
learning for the benefit of an enhanced educational experience.

26
Relationship between AI and LLMs
An exclusive study on LLMs and their relevance and usage in education must be
preceded with an understanding of the relationship between AI and LLMs. LLMs are
considered the most advanced natural language processing and generative engines (Xing et
al., 2024; see also Davar et al., 2025; Yan et al., 2023). LLMs are specifically distinct from
AI from their human-like capabilities that focus on adaptation, creativity, and contextual
relevance. Unlike general AI tools, LLMs are used as thought partners offering reasoning,
explanations, and enhancing collaboration (Ruiz-Rojas et al., 2024; see also Shen et al.,
2025). The journey of AI was founded on rule-based expert systems and search engines and it
has currently evolved to LLM enabled Generative AI. While fostering informational access
and creativity, LLMs go beyond the tasks of AI and generate new content and knowledge,
translate and adapt languages (Nguyen Van Viet et al., see also 2024; Belkina et al., 2025).

Overview of LLMs
Origin and Development of LLMs
The origin of language models was created somewhere between 1950-1960s and the
initial models lacked the computational techniques required for effective natural language
understanding and processing (Shahzad et al., 2025). The next important development took
place between 1980-1990s, when the existing language models started utilizing probabilistic
modelling for the prediction of word sequences (Shahzad et al., 2025). Despite this next step
in development of LLMs, the models still struggled with correctly capturing context and
semantics for effective generation of output. The development of LLMs was significantly
shaped and redefined in the mid 2010s with the progressive use of deep learning. The
introduction of RNNs, which are one step ahead than the traditional deep learning networks,
improved effective sequential predictions and generated conclusions based on prior inputs

27
(Shahzad et al., 2025). The most distinguished RNN, Transformer models which are
associated with LLMs was released in 2017 marking a significant milestone in the
development of LLMs. Unlike the traditional RNNs, Transformer models are able to examine
long text sequences and perform multiple computational steps simultaneously rather than in a
serialized manner (Daver et al., 2025). With the incorporation of transformer models, the
Open Ai introduced the first of their series of Generative Pre-Trained models (GPT) in 2018.
From GPT-1 to GPT-4, Open AI has introduced further advancements in each model
focusing on expanding the number of parameters, improving text generation, including
training from human feedback and refining the outputs (Shahzad et al., 2025).

LLMs and their sources of knowledge
LLMs are exposed to massive datasets consisted of books, peer-reviewed journal
articles, diverse professional guidelines and handbooks, websites, case databases, and many
more sources available online (Newton & Xiromeritis, 2024; see also Lucas et al., 2024; Wu,
Dang & Li, 2025; Yan et al., 2023). Once they are exposed, their training is based on deep
learning on these corpora, allowing the LLMs to absorb every detail on the datasets including
factual knowledge, reasoning patterns, and grammatical conventions across different
disciplines. For example, in the health care sector, LLMs are trained on current clinical
literature, existing patient records, and guidelines for plausible and sophisticated responses
(Lucas et al., 2024). The training of LLMs is not limited to be done on English or Westerncentric datasets but a diverse selection of datasets is used including regions or filed specific
Open Educational Resources (Alfirević et al., 2024; see also Nguyen Van Viet et al., 2024).
One may feel that the foundational knowledge of LLMs is limited and determined by the data
in the training datasets at the time of training development. However, it must be noted that
the effective responsiveness of the LLMs to carefully structured user prompts make them

28
flexible and adaptable to different contexts and evolving needs of the users (Newton &
Xiromeritis, 2024; see also Davar et al., 2025).

LLMs in Education
One simple definition cannot be devoted to the meaning of education as it can include
so many aspects related to the facilitation of learning and acquiring knowledge, skills, values
and habits. A person can undertake education at different stages of life and education can also
take different forms suiting the needs of people. For the ease of relevance of this paper, we
are focusing on formal education that is delivered through formal institutes such as primary to
high schools, universities, and other educational institutes.
Further, before we can dive into the application of LLMs in education and analyze
how they support learning, it is important to understand the objectives of learning to evaluate
whether LLMs have proven to be satisfying those objectives. As mentioned before, the
analysis of this paper will be based on Bloom’s taxonomy and the learning objectives
identified through this framework. A core goal of education is considered to be understanding
higher-order thinking and to go beyond basic memorization and achieve high levels of
understanding and application of knowledge. Accordingly, Bloom’s taxonomy has
categorized the learning objectives into 6 hierarchical levels; remembering, understanding,
applying, analyzing, evaluating and creating. While remembering simply refers to recalling
facts that are learnt, understanding means the ability to explain the ideas and concepts that we
have learnt. Next, Applying refers to the use of knowledge or applying the knowledge learned
to different contexts and for problem solving. Analyzing is the objective of differentiating the
facts and concepts learnt and the ability to organize information. Further, evaluating refers to
the act of judging the value and effectiveness of what is learnt. Finally, creating means
producing new ideas or new work depending on what has been learnt. This paper will focus

30
Application of LLMs in in-person classrooms
Traditional learning style of in-person or classroom education has been reshaped with
the use of LLMs. A study done on Japanese junior high schools has explained how these
schools are using LLM powered tools for visualizing and supporting classroom activities.
The traditional active reading classroom activities are now being alternated by individual and
group activities powered by LMS, e-books and analytics dashboards and daily attendance
logs are also captured through the same tools (Horikoshi et al., 2025). Further, learning
stream plots have been generated to process multimodal data across various platforms to
generate visualizations of timelines of classroom activities (Horikoshi et al., 2025). Such in
person classroom activities augmented by LLM platforms have helped the educators to track
the errors and discrepancies in their lesson plans and activities and identify the gaps in
student performances and issues in student engagement (Horikoshi et al., 2025). It can be
also seen that primary school educators, who are traditionally considered to be well based on
physical textbooks and restricted to classroom activities, are now taking a shift towards
deploying LLM powered chatbots such as ChatGPT for both teaching and assessing student
performance (Uğraş et al., 2024). Another study done on the integration of Multimodal LLMs
in science education has shown that the multimodal nature of science education which
involves not just reading and writing but diagram creation and interpretation, data analysis,
visualization of different solutions has been improved and reshaped with LLMs (Bewersdorff
et al., 2025). Accordingly, multimodal LLMs are currently used for generating visual
explanations that accompany modelling, experiments, and simulations in science lessons
(Bewersdorff et al., 2025). These Multimodal LLMs have also enabled real-time feedback
and evaluation of the content created through them (Bewersdorff et al., 2025). A similar
study has observed the use of ChatGPT to model a drainage of water through a tank as a prelab activity which was later used by the students to critically identify the gaps in the model

31
and improve on it through their own innovation (Keith et al., 2025). This study has tried to
confirm the role of LLMs not as a replacement for the original work of the students but only
as supplementary to guide them and enhance their critical thinking and innovation (Keith et
al., 2025).

Application of LLMs in online education
Impact on leaners:
The impact of LLMs is more commonly seen in online education. A cross-cultural
study done on 205 participants has observed the use of online quizzes enhanced with
gamification and their relation with LLMs. The study has proven that LLMs have been used
for the auto generation of the questions, answers, explanations in those gamified quizzes and
most importantly, real-time feedback is made possible through LLMs as a part of these
quizzes (Issabek et al., 2025). Further, this study has shown that LLMs have supported to
curate these quizzes depending on the capacity and performance level of different students
and also their demographic and cultural backgrounds. Another vital role played by LLMs to
support learning is how they envisioned as interactive conversational agents replacing
conversations between students and educators (Blaise Agüera y Arcas, 2022). These LLM
powered interactive conversational agents are able to respond to very theoretical and
philosophical queries, and provide logical reasoning (Blaise Agüera y Arcas, 2022). An
interesting study done on preschool classrooms that are focused on social sciences has found
out that these LLM powered interactive conversational agents are used in early childhood
education as well to develop language skills, early literacy and numeracy of children through
conversations, storytelling, and visuals (Doğan et al., 2025). A similar role is played by AI
chatbots that heavily assist the modern time flipped classrooms by helping the students to

32
have access to different resources before class and be ready for the lessons while practicing
conversational engagement, language development and brainstorming (Davar et al., 2025).
Impact on educators:
On the other hand, educators are largely using LLMs via GPT models in completely
asynchronous and online environments to measure the quality of knowledge and critical
thinking skills of students and their cognitive presence phases (Castellanos-Rayes et al.,
2025). Discussion posts are a common activity included in the higher education curriculum to
improve critical thinking skills, writing skills and student engagement. LLMs which use
structured prompts and adapted codebooks such as Practice Inquiry Model are being used to
assess hundreds and thousands of discussion posts to classify them into cognitive presence
phases (Castellanos-Rayes et al., 2025). Such very specific usage of LLMs focus on critical
inquiry and metacognitive learning of higher education (Castellanos-Rayes et al., 2025). A
study done exclusively in the context of language education and more specifically on English
language learners, who takes English as their second language, has observed the usage of
LLM models such as GPT-4, GPT-3.5, Google’s PaLM 2, and Anthropic’s Claude 2 for
automated essay scoring (Pack et al., 2024). This study has explained that such automated
essay scoring is consequential to explicit instructional prompts by educators that provide
detailed scoring rubrics to the LLMs. Accordingly, the authors have confirmed that LLMs
have provided not just a grade for the essays but also a detailed reasoning for each score
identifying the areas of improvement, gaps in skills and also highlighting the strengths (Pack
et al., 2024). This study has also focused on comparing the scoring LLMs to human graders
to analyze the benefits of using such automated scoring systems for time savings and more
reliable feedback (Pack et al., 2024). Further, the literature has proven that educators are
greatly benefitted by the generation of lesson plans by LLMs, creation of diverse assessments
and exams that prevent repetition in evaluation models, and facilitating a more convenient

33
and comfortable space for tailored feedback that aims to address the unique drawbacks of
each student (Xing et al., 2024; see also Nguyen Van Viet et ak., 2024; Lang et al., 2024).

Application of LLMs in self-directed learning
Self-directed learning is a modern method of learning that offers the learners complete
control to manage their educational journey. Accordingly, the learners are expected to
identify their goals and objectives in learning, identify and utilize the appropriate resources
and strategies for effective learning. This is one of the main areas of education that is heavily
regulated by LLMs as LLMs have proven to be very supportive in assisting the learners to
organize their self-directed learning schedules. LLMs have been able to democratize access
to best suited methods of learning, especially when engaged in self-directed courses.
Campbell & Cox (2024) studying the application of generative AI in higher education has
established that LLMs are used to carry out the most basic tasks in self-directed learning such
as summarizing of lengthy materials, finding out appropriate resources for assignments,
generating ideas, curation of formats and structures of assignments, and for receiving of
feedback that would critique the both AI and self-generated work and receive potential grades
before actual work is submitted. Özçelik & Yangin Ekşi (2024) have also shown the
importance of using LLMs such as ChatGPT as learning assistants to cultivate writing skills.
A group of researchers have studied a set of radiology postgraduate students in India to assess
their awareness and engagement with LLMs in self-directed learning. This area of study is
complex and LLMs have proven to be of great assistance in simplifying subject matters,
structuring reports, curating differential diagnoses and providing access to a great variety of
related textbook materials (Sarangi et al., 2024.). GPT platforms powered by LLMs have
offered consistent supplementary support in online and asynchronous environments related to
such exclusive and complex areas of study. A literature covering the impact of LLMs in

34
medical education has mentioned how medical students easily have access to up-to-date
medical and clinical content and scenarios through LLMs that a considerable amount of time
is saved as they do not have to manually refer to hundreds of books to gather the relevant
information (Lucas et al, 2024).

Relationship to the theoretical frameworks
In-person classrooms
The above-mentioned applications of LLMs in learning have proven how they
directly support the different types of learning introduced by Diana Laurillards’ LCF and
learning objectives as shown in Bloom’s taxonomy.
Generating visualizations in in-person classrooms and using LLMs to create models and
diagrams in science education support two types of learning in the LCF. Accordingly,
acquiring of knowledge and producing levels are achieved. Using the LLMs to create
visualizations help the students to acquire knowledge more conveniently with ease by
reducing complications. Further, using LLLMs to create science models reflects the effect of
LLMs in producing new work by learners. Turning to the Bloom’s taxonomy, these usages of
LLMs in physical classrooms. Impact the remembering and understanding skills of learners
by helping them to efficiently recall and comprehend large sets of data. Further, by using the
guidance of LLMs to produce new work help them to apply the knowledge effectively and
achieve initial stages of ‘creating’ by laying the foundation for learners through initial models
and diagrams and allowing them to build on them.
Online education
The usage of LLMs in online education to generate quizzes, their questions and
responses and curating the content of quizzes to the different needs of students shows how it
supports acquiring and producing. Further, the interactive conversational agents enhance the

35
skills of inquiring and discussing of learners as they build conversations with these robots to
express and exchange ideas and seek explanations. Further by using these conversational
agents in early childhood educations again supports the acquiring of knowledge of young
students through storytelling and visualizations.
The usage of LLMs in online education is also impacting different levels of the
Bloom’s taxonomy. The conversational agents directly help the learners to comprehend
complex information and remember them in simple terms and these conversational
engagements help them to analyze educational material and apply them in their day to day
studying routines and repetitive questioning and seeking explanations, learners are able to
evaluate the responses provided by LLMs and apply their own judgement to come into
conclusions.
Further, the educators’ usage of LLMs in grading and scoring have also helped to
understand and analyze the performance gaps of students. Further, it has helped the educators
to evaluate their feedbacks against the feedbacks created by LLMs and create the best
matching feedback for the performance of students.
Self-directed learning
In self-directed learning, LLMs are mostly assisting the students in guiding their
course journeys and curating easy ways to get through the educational materials.
Accordingly, by summarizing lengthy materials, structuring assignments and providing
access to resources, LLMs are supporting the acquiring and producing of knowledge in selfdirected learning environments. Further, they help the learners to understand course materials
in the absence of physical educators, remembering facts, apply the knowledge acquired in
assignments and creating their own work through researches, thesis, and responses to exams.

36
Benefits of using LLMs in learning and education
Following the discussion of the application of LLMs in different modes of learning,
the next section is devoted to identifying the benefits highlighted by the current literature of
using LLMs in learning and how they support the improvement and enhancement of
effectiveness of learning. Even though the usage of LLMs can be diversified across different
modes of education such as in-person or classroom learning, online learning, and selfdirected learning, the benefits they offer are common to all LLM driven tools. Therefore, this
paper will discuss a few common groups under which the benefits of LLMs can be captured.

Personalized Learning Experience
One of the most important features that has been highlighted by most of the literature
is how LLMs have offered personalized learning experiences for students with varying skills.
This has allowed learners to engage in learning at their own pace which is shaped by their
individual needs and prior knowledge. AI assistants powered by LLMs are able to explain
any complex text or content in a very simple manner that could be understood by different
individuals with varying understanding capacities. Customized learning pathways,
explanations, feedback, and AI driven tutors and conversational agents are addressing
learning gaps of individuals by offering diversified support in real time (Xing et al., 2024; see
also Nguyen Van Viet et al., 2024; Alfirević et al., 2024; Wu et al., 2025; Cecchini et al.,
2025; Ruiz-Rojas et al., 2024).
Furthermore, the personalized experience is not limited to students but educators are
also benefited by this advantage. LLMS are able to assess the quality of teaching
performance, lesson quality, metacognition, and innovation of educators and give feedback so
that they are able to identify the gaps in their performances and work on to improve (Fang et
al.,2025; see also Belkina et al., 2025).

37
This benefit enables all levels of the LCF and learning objectives of Bloom’s
taxonomy. By customizing the educational content to the different needs of the students, the
LLMs affect acquiring, inquiring, practicing and discussing stages of student learning and it
further supports the achieving of all cognitive levels of the Bloom’s taxonomy.

Accessibility, and inclusion
LLMs have been able to overcome many barriers that could not be surpassed by the
traditional learning systems. They facilitate different content such as multimodal,
multilingual, and translated outputs that can be easily embraced by the learners with
disabilities or language barriers (Xing et al., 2024; see also Wu, Dang & Li, 2025; Cecchini et
al., 2025; Ruiz-Rojas et al., 2024). LLMS are being used to create alternative formats and
adaptive resources for any educational content that cannot be received by individuals with
learning disabilities or representing different backgrounds and cultures. By this way, LLMs
have been able to uphold equitable education with broadened participation and democratized
access which is a vital Sustainable Development Goals (Nguyen Van Viet et al., 2024; see
also Ruiz-Rojas et al., 2024b; Nedungadi et al., 2024).
Further, the support of LLMs is available and accessible 24/7 and therefore, providing
much flexibility for the learners. In traditional educational systems, the educators are
available only during the learning hours and any feedback required after learning hours must
wait until the following working day. However, especially for students who are pursuing
higher education along with part time jobs and other commitments, such limited access to
educators is inconvenient. Therefore, readily available AI tutors or LLMs that provide instant
feedback and guidance are greatly benefiting the current trending lifestyles of university
students (Sarangi et al., 2025; see also Reicher et al., 2025; Xing et al., 2024; Davar et al.,
2025).

38
By clearing the barriers in educational accessibility, LLMs are allowing a wide range
of learners to be included in all levels of the LCF and paving the way for engagement in
higher cognitive skills in the Bloom’s taxonomy.

Efficiency and scalability
LLMs have assisted the educators to increase efficiency in their routine tasks with the
automation of lesson planning, generating of assessments including quizzes, exams, and
assignments, grading, and providing feedback. By this way, educators have also been able to
handle larger classrooms and focus more on student mentorship, individual attention to
students, and curriculum innovation (Wu et al., 2025; see also Belkina et al., 2025; Lang,
Triantoro & Sharp, 2024; Davar et al., 2025). Further, through tools such as customizable
Open Educational Resource LLMs, instructors can create and consistently update resources
(Alfirević et al., 2024; see also Lang et al., 2024).
By improving efficiency and scalability of the educators and their routine tasks,
LLMs are supporting the LCF stages related to the educators and thereby impacting acquiring
and practicing knowledge and their skills through generating assessment contents, and lesson
plans.

Promoting lifelong learning
LLM powered tools and platforms create an independent environment for the learners
to pursue their education autonomously. They get the chance to practice exams, quizzes, get
scripts for presentations and rehearse, get real-time feedback before any assessments (Wu,
Dang & Li, 2025; see also Reicher et al., 2025; Nguyen Van Viet et al., 2024). Lucas et al.,
(2024) has brought in examples from medical education to show how students rehearse
clinical reasoning scenarios and cases with AI conversational agents and using their feedback
for improvement. Further, LLMs have helped students to let go of anxiety, frustration or

39
humiliation that is caused by asking ‘basic’ questions in a classroom setting as LLMs provide
immense independence to enquire about the most basic concepts and points until the students
have fully understood (Lucas et al., 2024).

Increased motivation, engagement and enjoyment
Students have proven to be more motivated and engaged in using AI powered
resources such as gamified quizzes and assessments compared to learning in traditional
classrooms. The more tailored and customized format for learning has helped the students to
be less stressed as the feedback given through LLMs are clear, curated to the individual skill
capacity and therefore, limits comparison and competition with the rest of the learners. As a
result, students have experienced increased enjoyment and motivation along with sustained
efforts and hard work to achieve their educational goals (Ruiz-Rojas et al., 2024; see also
Davar et al., 2025).

Fostering collaborative learning
LLMs have been able to foster collaborative learning in higher education settings. An
exclusive study done on the usage of LLMs by nursing students has shown how LLMs are
used by small groups of students for joint construction and refinement of nursing care plans,
specifically through shared prompt engineering (Shen et al., 2025). These student groups
have cooperated on making, using and updating their prompts to make sure the outputs from
the AI fit the care plan better. The teams have collaborated over many steps and discussed the
LLM results, comparing them against the textbook materials and relied on different
perspectives of each team member to improve the care plans produced by the LLMs (Sgen et
al., 2025). These students have stated that they have experienced the value of LLMs not just
as a knowledge partner but also as a platform encouraging teamwork and collaborative
learning. Newton and Xiromeritis (2024) have also noted that LLMs facilitate collaboration

41
LLMs in learning environments. The existing literature has elucidated the importance of
identifying these limitations in order to ensure that these areas are improved so that they do
not compromise the efficiency and integrity of education but complement the same. This
section will explore the existing literature’s take on the limitations of using LLMs in learning
environments and challenges they pose against effective learning outputs. Further, these
studies have also identified the ethical considerations surrounding the heavy use of LLMs for
educational purposes. This section will therefore shed light on this aspect as well.

Limitations on complexity and clarity of content
A study done on elementary science has explored the scientific validity of responses
provided by ChatGPT to measure the quality and sustainability of education. Accordingly, it
has found out that even though 94.2% of responses provided by ChatGPT for elementary
science questions were valid, only 70.6% responses were clear (Choi, 2025). Further, only
12.8% of responses have been relevant to the curriculum (Choi, 2025). The main issue
presented here was that responses were not effectively received by the young learners and the
advanced terminology that was used was not proportionate to the age and skill levels of the
learners. This issue can be understood as presenting accessibility barriers to the learners and
therefore, hindering the effectiveness of learning. LLMs are criticized for producing
inaccurate responses or presenting the risk of misconceptions. The same study on elementary
science education has found out that the responses generated by ChatGPT involved subtle
misconceptions and therefore, demanding the mediation of an educator to prevent the risk of
misunderstanding by students (Choi, 2025).

Lack of empirical evidence
Another criticism raised is that responses and content generated by LLMs lack
empirical evidence for factual learning improvements and they are greatly considered to be

42
conceptual (Nguyen Van Viet et al., 2024). Further, LLMs are comparatively better in
helping learners for ‘paper understanding’ when the correct information is provided rather
than in open-ended ‘literature review’ tasks where LLMs have proven to be weaker (Wang et
al., 2025; see also Yan et al., 2024). As a result, the issue of generating hallucinations or false
references has been common (Maci & Anesa, 2025; see also Claman & Sezgin, 2024).
Dibrell (2024) has mentioned that the ‘Stochastic parrot’ phenomenon related to the
functionality of LLMs has caused them to produce work that looks credible but often includes
many errors and false details. Highlighting the issue of hallucinations, Yan et al (2024) has
pointed out how AI detection tools such as Turnitin produce false-positive rates causing
unjust decisions that can have negative consequences on student grades and self-perception.
Another result of fabrication of data is academic misconduct that happens when students
present work in the assumption that research was carried out when it actually was not (Erkılıç
& Ibrahim Cifci, 2024).

The issue of over-reliance
In the same study that has observed 48 graduate students and their use of LLMs in
higher education, Wang et al (2025) has mentioned that over-reliance or inappropriate
adoption of LLMs is a rising issue. Due to mismanagement of time or lack of time and tight
deadlines, students tend to rely heavily and completely on tools powered by LLMs. Such over
reliance can result in a number of issues. The creativity and critical thinking of students are
substantially hindered as the heavy use of LLMs prevent students from expressing
authenticity, personal engagement, reflective depth and they produce very generic work that
lack individuality (Dibrell, 2024; see also Claman & Sezgin, 2024). Maci & Anesa (2025)
has stated that LLM powered automated tools can assist learners and educators but they
cannot replace the pivotal role played by the human analyst and that as a result, development

43
of essential human interpretive skills is hindered. Further, when students get used to
producing work generated by LLMs, they often lack deep understanding of the learning
concepts and indicate very less confidence of the surface level work (Dibrell, 2024).

Lack of digital literacy
Another issue related to the use of LLMs is the lack of digital literacy amongst both
students and educators. Most LLM based education tools are currently in lab-testing phases
or during early research (Maci and Anesa, 2025). Even though they are already heavily used
by learners, they have not been sufficiently tried in real classroom settings. This poses a
challenge on the practical efficacy of LLMs and digital illiteracy further feeds the inability of
educators and students to effectively recognize the limitations and biases in AI generated
content or fix these errors (Maci & Anesa, 2025).

Limitations and biases in training data
Turning towards the ethical concerns surrounding the use of LLMs in education, one
of the most vital issues is how learning equity can be restricted due to limitations in training
datasets of LLMs. Nguyen Van Viet et al (2024) has mentioned that training datasets of
LLMs are considered to be globally unbalanced and they are trained by and large on English
data and from western, educated, industrialized, rich and democratic countries. This
limitation on training has therefore marginalized learners representing other cultural
backgrounds from having equitable access and deriving the expected and most effective
outputs from the LLMs (Nguyen Van Viet et al., 2024). Further, the deployment and
maintenance of LLMs is expensive and therefore, institutions with less resources find it
difficult to have access to high level LLMs (Guizani et al., 2025). This again creates
educational inequalities. Another result of such bias in training data is the representation of
systemic inequalities and underrepresentation of the minority (Dibrell, 2024). The cost of

44
modifying or updating model-generated content is high and therefore, this drawback hampers
accuracy in the output created by LLMSs (Guizani et al., 2025).

Data privacy concerns
Another common ethical consideration is the privacy related to using student data for
fine-tuning of LLMs (Claman & Sezgin, 2024). According to Yan et al. (2024) if LLMs are
fine-tuned using private data of students without any strong anonymization, details of
personal identities can leak through future responses produced by LLMs.

Issues related to authorship
When researchers or academic writers depend on LLM powered tools for content
generation, few writers can submit the same work for multiple journals, probably minimally
altered (Erkılıç & Ibrahim Cifci, 2024). This happens due to the duplication of work done by
LLMs and slicing or spilling one work or project into several publications. This is considered
a substantial violation of publication ethics (Erkılıç & Ibrahim Cifci, 2024). Another
important ethical concern surrounding the use of LLMs is the ambiguities related to
authorship and attribution of academic work. AI powered LLMs are not able to fulfill the
requirements of scientific authorship and their ability to generate content blurs academic
records (Erkılıç & Ibrahim Cifci, 2024). Accordingly, they end up removing credit from
actual contributors or incorrectly assign authorship (Erkılıç & Ibrahim Cifci, 2024). This is
also tied to the issue of unclear intellectual property rights of content generated by LLMs
through training on proprietary materials (Erkılıç & Ibrahim Cifci, 2024; see also Guizani et
al., 2025).
Figure 7
Limitations of using LLMs

46
Fine-tuning LLMs
Further, for the biases presented in the outputs of LLMs stemming from the biases
and social inequalities embedded in the training data sets, the existing literature has
recommended to fine-tune LLMs in a way to match the local languages, different cultural
context and literacy levels (Choi, 2025). It is also proposed to use adaptive language
processing which will allow the LLMs to tailor their outputs to the varying backgrounds of
the learners so that the social achievement gaps will not be reinforced through the responses
of LLMs (Choi, 2025). Another solution to this same issue is to calibrate the LLMS for local
academic standards to prevent content generation focused on one-size-fits-all style (Nguyen
Van Viet et al., 2024). This will also help educators from different regions and demographics
to get the optimal assistance from LLMs. To address the issue of hallucinations, Jacobsen &
Weber (2023) has recommended critically evaluating the outputs generated by LLMs and
refine the prompts iteratively to mitigate the errors.

Developing prompt engineering skills
Another important solution to mitigate the negative impact of LLMs is to enhance the
prompt engineering skills of both educators and students. The output of the LLMs is greatly
dependent on the prompt we send to them and if the prompt is poorly constructed, it can lead
to inaccurate or irrelevant results (Lee & Palmer, 2025). Wang et al. (2021) has introduced
the CRISPE framework that outlines the basis to craft precise prompts. Accordingly,
Capacity and role refers to defining the role of the LLM in which we want them to operate
and therefore, confirm whether we want them to act as a copywriter, researcher, or translator.
Next is ‘I’ for insight which refers to providing the necessary information and contextual
background to the LLM for the reason of relevancy. Next, ‘S’ stands for Statement and this
means the core of the prompt. This refers to what we ask the LLM to do and it is

47
recommended to be very specific to get a more precise response. Next is Personality and this
reflects on the personality of the LLMs or how we want the LLMs to respond. For example,
do we require the LLM to provide a professional response or casual friendship? This will
enhance personalization in outputs and individuality. The final part is Experiment which
requires the user to ask LLMs to provide multiple responses or examples and therefore,
having the option to choose the best response or combine few different responses.
A similar approach is using the AIPROMT framework for refining prompts to achieve clarity
and specificity (Korzynski et al., 2023). Another step that can be taken to enhance prompt
engineering skills is to include the same in academic curriculum and train the students on
this. Accordingly, prompt engineering skills of learners should be observed and assessed to
make sure that they are well positioned to obtain optimal results from the use of LLMs.
Knoth et al. (2024) has suggested a quantitative prompt quality score to assess the clarity,
context and restrictions in prompts written by learners.

Policies to enhance academic integrity
To address the issues related to authorship and academic integrity, the educational
institutions must implement policies to robust authorship and practice attribution habits.
Accordingly, learners must be encouraged to clearly distinguish between contributions by
human brain and AI tools and disclose any assistance taken from LLMs in their assignments
and publications (Erkılıç & Cifci, 2025; see also Dibrell, 2024). Further, assessments must be
modelled in such a way that they cannot be easily outsourced to LLMs. To do this,
assessments must be created in a way to focus on emphasizing synthesis, analysis, and realworld examples and applications so that learners cannot simply follow responses provided by
LLMs (Wang et al., 2025). Further, introducing clear institutional policies to guide the

48
students in the usage of AI will mitigate any overreliance and force the students to balance
their involvement with the integration of LLMs in their work (Klar, 2025).

Enhancing digital literacy
Another solution to improve the usage of LLMs amongst the learners is to enhance
their information literacy education. Bećirović et al. (2025) has mentioned that by fostering
vetting skills and teaching the learners to analyze the validity of output generated by LLMs
against trusted sources and therefore be informed and trained to identify any hallucinations
produced by LLMs. It should be taught to the students that LLMs must only be used as
debaters or moderators but not as mere answer providers. Learners must be skilled to
properly justify, critique or build upon the content generated through LLMs (Reicher et al.,
2025).

Increase empirical studies
One of the focal issues related to the development of LLMs is the lack of empirical
evidence related to their impact in actual classroom settings. Therefore, it is recommended
for higher education institutes and high schools to carry out open and real-world pilot studies
that will foster the sharing of methods and outcomes that need to develop a solid knowledge
of how LLMs work (Dibrell, 2024; see also Yan et al., 2024). Further, involving the main
stakeholders, students and educators in the process of designing and evaluation will enhance
the user perspective and improve the identification of real-world issues related to usage that
will pave the way for efficient and practical solutions (Filippi & Motyl, 2024).

The future of LLMs in Learning; Emerging trends and possible trends
The educational sector is looking ahead to identify the emerging trends that would
define the role of LLMs in learning in the future. LLMs are already creating a powerful

49
transformative effect in education even though most of these LLMs are still in their early
stages of development. Therefore, if the creators of these models can address the limitations
and drawbacks hindering the optimal performance of these models with effective solutions,
the power of LLMs and what they can do is beyond one’s imagination.
Parakh et al. (2024), speaking of the future role of LLMs in education, has mentioned
that models such as Llama 3.1, CodeLlama, and Mistral can be used for retrieval-augmented
generation (RAG) systems and vector databases to enhance personalized learning experiences
through more targeted and adaptive content delivery. The aim of such developments is to
encourage engagement and knowledge retention of learners by enhancing the role of LLMs
as both content generators and effective real-time feedback agents (Parakh et al., 2024).
Another emerging trend of LLMs is their potential ability to create a rich learning
experience through multimodal LLMs. Al-Dokhny et al. (2024) has discussed the potential
use of multimodal LLMs mediated by acceptance frameworks such as Task-technology Fit,
AIDUA that will consider the hedonic motivation, social influence, expectancy of effort and
willingness to adopt. Such advancements will result in sustainable improvements in higher
education and align with continuing customization. This study has empirically confirmed that
multimodal LLMs are able to refine personalized education through exclusive fit of
technology to the task and user context (Al-Dokhny et al., 2024). It has also been discussed
that it is likely to adopt new systems based on adaptive learning and automated ‘sensing’
tools for common use of research and practice (Whitehead et al., 2025).
Another potential role that is envisioned for the LLMs is the role to function as
cognitive partners rather than mere information sources. Expanding the application of LLMs
to stakeholder and collaborative project-based assignments show considerable potential of
LLMs improving group work and enhancing collaborative education by facilitating socialemotional skills acting as interactive partners (Cain, 2025). Further, it has been identified that

51
of technology and perspective of learning, the methods of learning has extended beyond the
traditional system to many other methods such as online learning, and self-directed learning.
These new learning systems are consisted of a number of new features that give rise to
different activates that are undertaken to achieve the objectives of learning. The development
of technology and more specifically, artificial intelligence has heavily supported these
learning activates by integrating advancements to the mode of delivery and receipt of
education. AI has reshaped the learning systems, especially through LLMs, it's one of the
dominant counterparts. This study has conducted a systematic overview of the existing
literature on how LLMs support education and the analysis has proven a wide-range of
applications of LLMs in educational context. These LLMs have introduced many new
features that give rise to numerous benefits. The existing literature has identified how LLMs
have reshaped learning and education by bringing personalization, engagement,
collaboration, enhancement of efficiency to not only learning experiences but also to the role
of the educators. Collaborative experiences and real-time feedback have been heavily
highlighted by the current studies as valuable outputs of integrating LLMs in education.
However, the existing researchers have been able to identify drawbacks and limitations
associated with the application of LLMs in learning. These limitations can result from many
reasons but mainly due to the fact that these models are still in the early stages of
development and the lack of empirical evidences in practice. Ethical concerns related the
operation of the LLMs are also requires substantial attention. The current studies have not
failed to highlight the appropriate solutions for these drawbacks and emphasized the
importance of increasing the deployment of LLMs in actual scenarios including physical
classrooms and personal experiences to gain real-time evaluation. By this way, empirical
evidence can be collected to measure the success of the role played by LLMs in learning.

52
It is vital to understand by both learners and educators that LLMs are an efficient tool
to supplement and enhance the learning experience. By understanding how these LLMs work
and trained, the learners will realize that they are not supernatural creatures that offer what
you need. Rather, they are modelled and trained to support the learning process and human
mediation plays important role in achieving the optimal results out of the LLMs. Therefore, it
must be strongly understood that LLMs are here to support the learners and educators but not
replace any of them or any part of the learning process. They are here to be developed and to
be used for attaining the highest results of education through supplementing of human
cognitive skills.

53
References

Adams, N. E. (2015). Bloom's taxonomy of cognitive learning objectives. Journal of the Medical
Library Association, 103(3), 152–153. https://doi.org/10.3163/1536-5050.103.3.010
Agüera y Arcas, B. (2022). Do large language models understand us? Dædalus, 151(2), 183–197.
https://doi.org/10.1162/DAED_a_01909
Ajani, O. A., Akintolu, M., & Afolabi, S. O. (2024). The emergence of artificial intelligence in
higher education: Prospects and challenges of AI. International Journal of Research in
Business and Social Science, 13(8), 157–165. https://doi.org/10.20525/ijrbs.v13i8.3507
Al-Dokhny, A., AlHarthi, H. M., Alalwan, N., & Alkhalaf, S. (2024). Can Multimodal Large
Language Models Enhance Performance Benefits Among Higher Education Students?
Sustainability, 16(10780), 1–28.
Alfirević, N., Gojević, A., Blažević, N., & Crnjac Milić, D. (2024). Custom-Trained Large Language
Models as Open Educational Resources. Information, 15(2), 89.
https://doi.org/10.3390/info15020089
Baltaci, K., Herrmann, M., & Turkmen, A. (2024). Integrating Artificial Intelligence into Electrical
Engineering Education: A Paradigm Shift in Teaching and Learning.
Bećirović, S., Polz, E., & Tinkel, I. (2025). Exploring students’ AI literacy and its effects on their AI
output quality, self-efficacy, and academic performance. Smart Learning Environments,
12(29). https://doi.org/10.1186/s40561-025-00384-3
Belkina, M., Daniel, S., Nikolic, S., Haque, R., Lyden, S., Neal, P., Grundy, S., & Hassan, G. M.
(2025). Implementing generative AI (GenAI) in higher education: A systematic review of
case studies. Computers and Education: Artificial Intelligence, 8, 100407.
https://doi.org/10.1016/j.caeai.2025.100407

54
Bewersdorff, E., Sandmann, A., Ulbricht, C., Schreiber, F., Greefrath, G., & Rumann, S. (2025).
Taking the next step with generative artificial intelligence: The transformative role of
multimodal large language models in science education. Education and Information
Technologies, 30(2), 1205–1234. https://doi.org/10.1007/s10639-025-12345-x
Bucea-Manea-Tonis, R., et al. (2022). Artificial Intelligence Potential in Higher Education
Institutions: Enhanced Learning Environment in Romania and Serbia. Sustainability, 14,
5842.
Cain, W. (2025). Cognitive partners in design: Using generative AI for active online learning in a
graduate-level course. Educational Research: Theory and Practice, 36(1), 42–51.
Campbell, L. O., & Cox, T. D. (2024). Utilizing generative AI in higher education teaching and
learning. Journal of the Scholarship of Teaching and Learning, 24(4), 162–173.
https://doi.org/10.14434/josotl.v24i4.36575
Castellanos-Reyes, D., Olesova, L., & Sadaf, A. (2025). Transforming online learning research:
Leveraging GPT large language models for automated content analysis of cognitive presence.
International Review of Research in Open and Distributed Learning, 26(2), 98–115.
https://doi.org/10.19173/irrodl.v26i2.7850
Cecchini, S., Ceccacci, S., Giaconi, C., & Del Bianco, N. (2025). AI chatbots in education:
Challenges and opportunities. Information, 16(3), 235. https://doi.org/10.3390/info16030235
Choi, Y. (2025). Exploring the scientific validity of ChatGPT’s responses in elementary science for
sustainable education. Sustainability, 17, 2962. https://doi.org/10.3390/su17072962
Claman, D., & Sezgin, E. (2024). Artificial intelligence in dental education: Opportunities and
challenges of large language models and multimodal foundation models. JMIR Medical
Education, 10, e52346. https://doi.org/10.2196/52346
Davar, N. F., Dewan, M. A. A., & Zhang, X. (2025). AI chatbots in education: Challenges and
opportunities. Information, 16(3), 1-25. https://doi.org/10.3390/info16030235

55
Davar, P., Tang, K.-Y., & Wang, F. (2025). Evaluating simulated teaching audio performance using
retrieval-augmented generation and large language models. Information, 16(4), 264.
https://doi.org/10.3390/info16040264
Dibrell, D. (2024). Ethical considerations of using AI in writing pedagogy. Sociological Imagination,
60(1), 20–24.
Doğan, D., Yıldız, S., & Özkan, S. (2025). Effectiveness of artificial intelligence practices in the
teaching of social sciences: Pre-school education. Educational Technology & Society, 28(3),
77–92.
Eber, P. A., & Parker, T. S. (2007). Assessing Student Learning: Applying Bloom's Taxonomy.
Human Service Education, 27(1), 45–53.
Erkılıç, E., & Cifci, I. (2025). Ethical issues and violations in using chatbots for academic writing
and publishing: The answers from ChatGPT. Journal of Multidisciplinary Academic Tourism,
10(1), 111–120. https://doi.org/10.31822/jomat.2025-10-1-111
Filippi, S., & Motyl, B. (2024). Large language models (LLMs) in engineering education: A
systematic review and suggestions for practical adoption. Information, 15(6), 1-19..
https://doi.org/10.3390/info15060345
Guizani, I., Mehmood, R., Leclercq, E., Abuadbba, A., Song, H., & Yasar, A. (2025). A systematic
literature review to implement large language model in higher education: Issues and
solutions. Discover Education, 5(1), 12–34. https://doi.org/10.xxxx/discedu.2025.xxxxx
Horikoshi, K., Ikeda, K., Shibata, Y., & Takahashi, T. (2025). Learning stream plot: Classroom
activity visualization with daily learning log data. Journal of Smart Learning Environments,
12(15), 1–23.
Imran, M., & Almusharraf, N. (2024). Google Gemini as a Next Generation AI Educational Tool: A
Review of Emerging Educational Technology. Smart Learning Environments, 11:22.

56
Issabek, A., Oliveira, W., Hamari, J., & Bogdanchikov, A. (2025). The effects of demographic
factors on learners’ flow experience in gamified educational quizzes. Smart Learning
Environments, 12(25), 1–20. https://doi.org/10.1186/s40561-025-00378-1
Keith, J. R., Deng, S., & Carmichael, P. (2025). Harnessing generative AI in chemical engineering
education through critical engagement. Chemical Engineering Education, 59(1), 4–13.
Klar, M. (2025). Using ChatGPT is easy, using it effectively is tough? A mixed methods study on K12 students’ perceptions, interaction patterns, and support for learning with generative AI
chatbots. Smart Learning Environments, 12(32). https://doi.org/10.1186/s40561-025-00385-2
Lang, M., Triantoro, T., & Sharp, J. H. (2024). Large language models as “educational assistant”:
Automatic writing of teaching cases for IS/business/IT education. Journal of Information
Systems Education, 35(2), 133–144.
Lee, D., & Palmer, E. (2025). Prompt engineering in higher education: A systematic review to help
inform curricula. International Journal of Educational Technology in Higher Education,
22(7). https://doi.org/10.1186/s41239-025-00503-7
Lucas, H. C., Upperman, J. S., & Robinson, J. R. (2024). A systematic review of large language
models and their implications in medical education. Medical Education, 58(11), 1276–1285.
https://doi.org/10.1111/medu.15402
Maci, S. M., & Anesa, P. (2025). The impact of AI on discourse analysis: Challenges and
opportunities. International Journal of Language Studies, 19(2), 1–12.
https://doi.org/10.5281/zenodo.15250758
Mezak Matijevic, M., Pisker, B., & Dokic, K. (2024). Constructing a socio-legal framework proposal
for governing large language model usage and application in education. Social Sciences,
13(9), 479. https://doi.org/10.3390/socsci13090479

57
Nedungadi, P., Raman, R., Deng, X., & Bianchi, B. (2024). Generative AI for achieving SDG 4:
Equity and quality in education. Smart Learning Environments, 11(1), 5.
https://doi.org/10.1186/s40561-024-00336-5
Newton, P., & Xiromeritis, M. (2024). ChatGPT performance on multiple choice question
examinations in higher education: A pragmatic scoping review. Assessment & Evaluation in
Higher Education, 49(6), 781–798. https://doi.org/10.1080/02602938.2023.2299059
Nguyen Van, V., Luong Thi Minh, H., Nguyen, V., Mong Quoc, T., Bui Anh, T., Le Anh, T., Phan
Thi, C., & Nguyen Huu, K. (2024). Revolutionizing education: An extensive analysis of large
language models integration. International Research Journal of Science, Technology,
Education, and Management, 4(4), 10–21. https://doi.org/10.5281/zenodo.14744029
Özçelik, N. P., & Yangın Ekşi, G. (2024). Cultivating writing skills: The role of ChatGPT as a
learning assistant—A case study. Smart Learning Environments, 11(10), 1–20.
https://doi.org/10.1186/s40561-024-00296-8
Pack, A., Barrett, A., & Escalante, J. (2024). Large language models and automated essay scoring of
English language learner writing: Insights into validity and reliability. Computers and
Education: Artificial Intelligence, 6, 100234. https://doi.org/10.1016/j.caeai.2024.100234
Parakh, D., Lodha, P., & Subalalitha, C. N. (2024). Adaptive Personalized Learning System with
Generative AI. Frontiers in Health Informatics, 13(8), 2612–2637.
Praveena, T., & Anupama, K. (2025). Machine learning meets language learning: The transformative
potential of artificial intelligence in English language instruction. Human Research in
Rehabilitation, 15(1), 133–144. https://doi.org/10.21554/hrr.042511
Reicher, H., Frenkel, Y., Lavi, M. J., Nasser, R., Ran-milo, Y., Sheinin, R., Shtaif, M., & Milo, T.
(2025). A generative AI-empowered digital tutor for higher education courses. Information,
16(4), 264. https://doi.org/10.3390/info16040264

58
Ruiz-Rojas, L. I., Salvador-Ullauri, L., & Acosta-Vargas, P. (2024). Collaborative working and
critical thinking: Adoption of generative artificial intelligence tools in higher education.
Sustainability, 16(13), 5367. https://doi.org/10.3390/su16135367
Sarangi, S., Choudhury, S. S., Agarwal, A., Thakker, A., & Cherian, J. (2024). Exploring radiology
postgraduate students’ engagement with large language models for educational purposes: A
study of knowledge, attitudes, and practices. Journal of European CME, 13(1), 224–235.
https://doi.org/10.1080/21614083.2024.1234567
Shahzad, T., Mazhar, T., Tariq, M. U., Ahmad, W., Ouahada, K., & Hamam, H. (2025). A
comprehensive review of large language models: Issues and solutions in learning
environments. Discover Sustainability, 6, Article 27. https://doi.org/10.1007/s43621-02500815-8
Shen, M., Shen, Y., Liu, F., & Jin, J. (2025). Prompts, privacy, and personalized learning: Integrating
AI into nursing education—a qualitative study. BMC Nursing, 24, Article 470.
https://doi.org/10.1186/s12912-025-03115-8
Sun, X., Yang, Y., & Du, Q. (2025). Advancing total quality management in higher education
through AI and big data analytics. Total Quality Management & Business Excellence.
Advance online publication. https://doi.org/10.1080/14783363.2025.2495103
Uğraş, M., Soylu, Y., & Zerengök, D. (2024). ChatGPT-supported education in primary schools:
Teachers’ perspectives. Eğitimde Nitel Araştırmalar Dergisi (Journal of Qualitative Research
in Education), 12(52), 317–338. https://doi.org/10.14689/issn.2148-2624.1.12c.52m.13
Wang, J., Huang, C., Yan, S., Xie, W., & He, D. (2025). When young scholars cooperate with LLMs
in academic tasks: The influence of individual differences and task complexities.
International Journal of Human–Computer Interaction, 41(8), 4624–4639.
https://doi.org/10.1080/10447318.2024.2352919

59
Wu, Q., Dang, Y., & Li, F. (2025). Systematic review of generative artificial intelligence (GenAI) in
higher education: Opportunities, challenges, and practical implications. Information, 16(3),
235. https://doi.org/10.3390/info16030235
Xing, W., Lan, A., Crossley, S., Yu, Z., Denny, P., Nixon, N., & Stamper, J. (2024). Revolutionizing
education: An extensive analysis of large language models integration. International
Research Journal of Science, Technology, Education, and Management, 4(4), 10–21.
https://doi.org/10.5281/zenodo.14744029
Yan, L., Sha, L., Zhao, L., Li, Y., Martinez-Maldonado, R., Chen, G., Li, X., Jin, Y., & Gašević, D.
(2023). Practical and ethical challenges of large language models in education: A systematic
literature review. British Journal of Educational Technology, 54(5), 91–108.
Yu, Y. F., Mai, N., & Hin, H. S. (2020). Enhancing Students' Collaborative Learning Experiences
with Laurillard's Conversational Framework. International Journal of Creative Multimedia,
1(Special Issue 1), 152–167. https://doi.org/10.33093/ijcm.2020.Χ1.13

60
Appendix 1

Keywords, their synonyms and search strings used for the search of academic papers

Key words

AI in education

LLM in eductaion

intelligent turoring
systems
based on AI and
LLM

Synonyms

Search Srting

("artificial intelligence"
OR AI OR "AI tools"
OR "machine learning")
AI tools in education,
AND
machine learning
(educat* OR school*
applications for
OR teach* OR learn*
schools
OR classroom
OR "higher education"
OR universit*)
("LLM" OR "large
language model" OR
"large language
models")
LLM in education
AND
and schools
(educat* OR school*
OR teach* OR learn*
OR classroom OR
universit*)
("intelligent tutoring
system*" OR "ITS" OR
"AI tutor*"
OR "LLM tutor*" OR
"AI tutoring platform*"
OR
"LLM tutoring
platform*" OR
"personalized AI
AI and LLM tutoring tutor*" OR
platforms,
"personalized LLM
personalized AI/
tutor*") AND
LLM tutors
("artificial intelligence"
OR "AI" OR "large
language model*" OR
"LLM" OR
"generative AI" OR
"ChatGPT" OR "GPT3" OR "GPT-4") AND
(educat* OR school*
OR teach* OR learn*
OR classroom* OR

# of
articles

# of
relevant
articles

80

6

16,041

50

82,000

10

61
universit* OR "higher
education" OR "K-12")

Natural Language
Processing in
Eductaion

Virtual Learning
Environments
through AI

Natural Language
Processing for
Learning, Language
AI
in Education

("Natural Language
Processing" OR "NLP"
OR "Language AI" OR
"Text AI")
AND
(educat* OR learn* OR
teach* OR school* OR
universit* OR "higher
education" OR "K-12"
OR classroom*)

30,000

10

Digital classroms,
online
learning spaces

("Virtual Learning
Environment*" OR
"VLE" OR "Digital
Classroom*" OR
"Online Learning
Space*" OR "AIpowered Learning
Environment*" OR
"Intelligent Learning
Platform*")
AND
("Artificial
Intelligence" OR "AI"
OR "Machine
Learning" OR
"Adaptive Learning"
OR "Intelligent
Tutoring System*" OR
"Generative AI" OR
"LLM" OR
"ChatGPT")
AND
(educat* OR learn* OR
teach* OR school* OR
universit* OR "K-12"
OR "higher education"

161

8

62
OR "distance learning"
OR "e-learning")

Virtual Learning
Environments
through LLM

LLM based digital
classroms
and online learning
spaces

Student engagement
with AI

AI driven student
motivation,
learner engagement
with technology

(("Virtual Learning
Environment*" OR
"VLE" OR "LLMbased Digital
Classroom*" OR
"LLM-powered Online
Learning Space*")
AND
("Large Language
Model*" OR "LLM"
OR "Generative AI"
OR "ChatGPT" OR
"GPT-3" OR "GPT-4"
OR "Transformer
Model*")
AND
(educat* OR learn* OR
teach* OR school* OR
universit* OR "K-12"
OR "higher education"
OR "distance
learning"))
("student engagement"
OR "learner
engagement" OR
"student motivation")
AND ("artificial
intelligence" OR "AI"
OR "AI-driven" OR
"adaptive learning")
AND (educat* OR
learn* OR school* OR
universit*)

1

1

1,121

12

63

Student engagement
through LLMs

Collaborative
learning with AI

Collaborative
learning
through LLMs

Knowledge
Retention
with AI

Knowledge
Retention
with LLMs

LLM-driven student
motivation, learner
engagement through
LLMs ("student
engagement" OR
LLMs driven student
"learner engagement")
motivation,
AND
learner engagement
("large language
through LLMs
model*" OR "LLM"
OR "ChatGPT" OR
"generative AI") AND
(educat* OR learn* OR
classroom*)
("collaborative
learning" OR "group
learning" OR "peer
AI supported group
collaboration") AND
learning,
("artificial intelligence"
digital peer
OR "AI" OR "AIcollaboration
supported") AND
(educat* OR school*
OR "higher education")
("collaborative
learning" OR "group
Group learning based learning") AND ("large
language model*" OR
on
"LLM" OR "GPT-4")
LLMs, digital peer
collaboration
AND (educat* OR
universit* OR "online
learning")
("knowledge retention"
OR "memory
enhancement" OR
Memory
"memorization") AND
enhancement tools,
("artificial intelligence"
AI assisted
OR "AI" OR "adaptive
memorization
learning") AND
(educat* OR learn* OR
school*)
("knowledge retention"
OR "memory
LLM supported
enhancement") AND
memory
("large language
enhancement tools,
model*"
LLMs assisted
OR "LLM" OR
memorization
"generative AI") AND
(educat* OR learn* OR
"higher education")

0

0

573

5

5,812

8

191

2

3

9

64

Critical Thinking
Development
with AI

Higher-order
cognitive skills,
reasoning skills with
AI

Critical Thinking
Development
with LLMs

Higher-order
cognitive skills
enahnced by LLMs,
reasoning
skills with LLMs

Assessment
Automation
with AI

AI-graded
assignments,
smart testing systems

Assessment
automation
with LLM

Grading assignments
through
LLMs, smart testing
sytems

Personalized
Feedback
through AI

Adaptive feedback,
instant AI feedback

Personalized
Feedback
through LLMs

Adaptive feedback
and
instant feedback
from LLMs

("critical thinking" OR
"higher-order cognitive
skills" OR "reasoning
skills") AND
("artificial intelligence"
OR "AI" OR "machine
learning") AND
(educat* OR school*
OR universit*)
("critical thinking" OR
"reasoning skills")
AND ("large language
model*" OR "LLM"
OR
"ChatGPT") AND
(educat* OR "higher
education")
("assessment
automation" OR
"automated grading"
OR "smart testing")
AND
("artificial intelligence"
OR "AI" OR "machine
learning") AND
(educat* OR school*
OR universit*)
("automated grading"
OR "assessment
automation") AND
("large language
model*"
OR "LLM" OR "GPT4") AND (educat* OR
"higher education")
("personalized
feedback" OR "adaptive
feedback" OR "instant
feedback") AND
("artificial intelligence"
OR "AI" OR "machine
learning") AND
(educat* OR learn* OR
school*)
("personalized
feedback" OR "adaptive
feedback") AND
("large language
model*" OR
"LLM" OR

1,330

11

434

5

71

4

2

0

294

20

15

10

65
"ChatGPT") AND
(educat* OR universit*
OR "online learning")

AI-enhanced
Creativity

Creative idea
generation by
AI, AI support for
innovation

LLM enhanced
creativity

Creative idea
generation
through LLMs, LLM
support
for innovation

AI-powered
Curriculum
Design

AI in syllabus
development,
intelligent course
design

LLM-powered
Curriculum Design

Syllabus
development and
course designing
through LLMs,

Learning Analytics
via AI

Educational data
mining,
AI for student
performance tracking

("creativity" OR
"creative idea
generation" OR
"innovation") AND
("artificial intelligence"
OR "AI" OR
"generative AI") AND
(educat* OR learn* OR
school*)
("creativity" OR "idea
generation") AND
("large language
model*" OR
"LLM" OR "GPT-4")
AND (educat* OR
"higher education")
("curriculum design"
OR "syllabus
development" OR
"course design") AND
("artificial intelligence"
OR "AI" OR "machine
learning") AND
(educat* OR school*
OR universit*)
("curriculum design"
OR "syllabus
development") AND
("large language
model*"
OR "LLM" OR
"generative AI") AND
(educat* OR "higher
education")
("learning analytics"
OR "educational data
mining" OR "student
performance") AND
("artificial intelligence"
OR "AI" OR
"predictive analytics")
AND (educat* OR
school* OR universit*)

33,017

12

5273

6

228

0

18

12

930

20

66

Learning anlytics via
LLMs

Educational data
mining,
student performance
tracking
through LLMs

AI conversation
Educational Chatbots agents,
through AI
virtual learning
assistants

LLM based
educational
chatbots

LLM based
converstaion
agents, virtual
learning
assistants

Competency-based
Education with
AI

Skills mastery
tracking,
personalized skill
development

Competency-based
Education with
LLMs

Skills mastery
tracking,
personalized skill
development

AI for Skill
Development

Soft skills AI
training,
hard skills AI
learning

("learning analytics"
OR "student
performance") AND
("large language
model*"
OR "LLM" OR
"generative AI") AND
(educat* OR "higher
education")
("educational chatbot*"
OR "virtual assistant*"
OR "conversational
agent*")
AND ("artificial
intelligence" OR "AI"
OR "NLP") AND
(educat* OR learn* OR
school*)
("educational chatbot*"
OR "conversational
agent*") AND ("large
language model*"
OR "LLM" OR
"ChatGPT") AND
(educat* OR "online
learning")
("competency-based
education" OR "skills
mastery" OR "skill
development") AND
("artificial intelligence"
OR "AI" OR "adaptive
learning") AND
(educat* OR school*
OR universit*)
("competency-based
education" OR "skills
mastery") AND ("large
language model*"
OR "LLM" OR
"generative AI") AND
(educat* OR "higher
education")
("skill development"
OR "soft skills" OR
"hard skills") AND
("artificial intelligence"
OR
"AI" OR "machine
learning") AND

46

10

915

25

44

12

425

10

3

0

517

18

67
(educat* OR learn* OR
school*)

LLMs for skill
Development

Soft and hard skills
trainiong and
learning
through LLMS

Microlearning with
AI

Bite-sized learning
modules,
AI-driven short
lessons

Microlearning with
LLMs

Bite-sized learning
modules,
LLM-driven short
lessons

Virtual Reality and
AI Learning

Immersive learning
technologies,
VR/AR blended
education

Virtual Relaity and
LLM learning

Immersive learning
technologies,
VR/AR blended
education

("skill development"
OR "soft skills" OR
"hard skills") AND
("large language
model*"
OR "LLM" OR "GPT4") AND (educat* OR
"vocational training")
("microlearning" OR
"bite-sized learning"
OR "short lessons")
AND ("artificial
intelligence"
OR "AI" OR "adaptive
learning") AND
(educat* OR learn* OR
school*)
("microlearning" OR
"bite-sized learning")
AND ("large language
model*"
OR "LLM" OR
"ChatGPT") AND
(educat* OR "mobile
learning")
("virtual reality" OR
"VR" OR "augmented
reality" OR "immersive
learning")
AND ("artificial
intelligence" OR "AI"
OR "machine learning")
AND (educat* OR
school* OR universit*)
("virtual reality" OR
"VR" OR "augmented
reality") AND ("large
language model*"
OR "LLM" OR
"generative AI") AND
(educat* OR "higher
education")

12

2

8

0

0

0

5,121

2

15

6

68

Gamification with
AI

AI-based educational
games
mechanics in
learning

LLM-based
Gamification through educational games
LLMs
and mechanics in
learning

Ethics in AI
Education

Responsible AI use,
ethical
considerations in
educational AI

Ethics in LLM
supported education

Responsible use of
LLMs,
ethical consideration
of
LLMs in education

Bias and Fairness in
Educational
AI

Algorithmic bias in
education,
fairness in AI
assessments

Bias and Fairness in
Educational LLMs

Algorithmic bias in
education,
fairness in LLM
assessments

("gamification" OR
"educational game*"
OR "game-based
learning") AND
("artificial intelligence"
OR "AI" OR "adaptive
learning") AND
(educat* OR school*
OR universit*)
("gamification" OR
"game-based learning")
AND ("large language
model*" OR
"LLM" OR "GPT-4")
AND (educat* OR
"higher education")
("ethics" OR
"responsible AI" OR
"ethical
considerations") AND
("artificial intelligence"
OR "AI" OR "machine
learning")
AND (educat* OR
school* OR universit*)
("ethics" OR
"responsible AI") AND
("large language
model*" OR "LLM"
OR "generative AI")
AND (educat* OR
"higher education")
("bias" OR "fairness"
OR "algorithmic bias")
AND ("artificial
intelligence" OR
"AI" OR "machine
learning") AND
(educat* OR school*
OR universit*)
("bias" OR "fairness"
OR "algorithmic bias")
AND ("large language
model*" OR
"LLM" OR
"ChatGPT") AND
(educat* OR "higher
education")

646

0

1

0

17,330

0

494

10

16,495

10

402

19

69

AI combining text,
Multimodal Learning
Images, and audio
with AI
for education

LLMs combining
Multimodal Learning texts,
with LLMs
images and audio for
education

Equity and Inclusion
in
AI Education

Accessible AI
learning,
inclusive EdTech

Equity and Inclusion
in
LLMs supported
learning

Accessible learning
with
LLMs, inclusive
EdTech

Teacher-AI
Collaboration

Human-AI coteaching
models educators
using
AI tools

Teachers and LLMs
collaboration

Human-LLM coteaching
models, educators
using LLMs

("multimodal learning"
OR "multimedia
learning") AND
("artificial intelligence"
OR
"AI" OR "computer
vision" OR "speech
recognition") AND
(educat* OR school*
OR universit*)
("multimodal learning"
OR "multimedia
learning") AND ("large
language model*" OR
"LLM" OR "GPT-4
Vision") AND (educat*
OR "higher education")
("equity" OR
"inclusion" OR
"accessible learning")
AND ("artificial
intelligence" OR "AI"
OR "machine learning")
AND (educat* OR
school* OR universit*)
("equity" OR
"inclusion" OR
"accessible learning")
AND ("large language
model*"
OR "LLM" OR
"generative AI") AND
(educat* OR "higher
education")
("teacher collaboration"
OR "co-teaching" OR
"educator tools") AND
("artificial intelligence"
OR "AI" OR "machine
learning") AND
(educat* OR school*
OR universit*)
("teacher collaboration"
OR "co-teaching")
AND ("large language
model*"
OR "LLM" OR
"ChatGPT") AND
(educat* OR "higher
education")

1,684

23

36

5

10,150

22

162

7

37

9

4

0

70

Appendix 2
Summary of the selected peer reviewed academic articles for the research
Authors

Year

Method

Adams, N. E.

2015

Conceptual/Review

Contribution
Explains Bloom's taxonomy
of cognitive learning
objectives.

Agüera y

2022

Conceptual/Perspective

Arcas, B.

Explores whether large
language models truly
"understand" human
language.

Ajani, O. A. et

2024

Review/Conceptual

al.

Discusses the prospects and
challenges of AI's
emergence in higher
education.

Al-Dokhny, A.

2024

Empirical/Survey

et al.

Investigates how multimodal
LLMs enhance performance
benefits among higher
education students.

Alfirević, N. et

2024

Conceptual/Case Study

al.

Explores custom-trained
LLMs as Open Educational
Resources (OER).

Baltaci, K. et
al.

2024

Conceptual/Review

Discusses AI integration in
electrical engineering
education as a paradigm
shift.

71

Bećirović, S. et

2025

Empirical

al.

Examines students' AI
literacy and its effects on AI
output quality, self-efficacy,
and academic performance.

Belkina, M. et

2025

al.

Systematic Review (Case

Provides a systematic review

Studies)

of case studies on
implementing generative AI
in higher education.

Bewersdorff,

2025

Conceptual/Review

E. et al.

Discusses the transformative
role of multimodal LLMs in
science education.

Bucea-Manea-

2022

Empirical/Case Study

Tonis, R. et al.

Explores AI potential for
enhancing learning
environments in higher
education institutions in
Romania and Serbia.

Cain, W.

2025

Case Study/Empirical

Examines using generative
AI as cognitive partners for
active online learning in a
graduate course.

Campbell, L.

2024

Review/Conceptual

Discusses the utilization of

O., & Cox, T.

generative AI in higher

D.

education teaching and
learning.

CastellanosReyes, D. et al.

2025

Empirical/Methodological

Explores leveraging GPT
LLMs for automated content
analysis of cognitive
presence in online learning
discussions.

72

Cecchini, S. et

2025

Review/Conceptual

al.

Outlines challenges and
opportunities of AI chatbots
in education.

Choi, Y.

2025

Empirical/Content Analysis

Investigates the scientific
validity of ChatGPT's
responses in elementary
science.

Claman, D., &

2024

Review/Conceptual

Sezgin, E.

Discusses opportunities and
challenges of LLMs and
multimodal models in dental
education.

Davar, N. F. et

2025

Review/Conceptual

al.

Examines challenges and
opportunities of AI chatbots
in education (similar to
Cecchini, distinct reference).

Davar, P. et al.

2025

Empirical/Experimental

Evaluates simulated teaching
audio performance using
retrieval-augmented
generation (RAG) and
LLMs.

Dibrell, D.

2024

Conceptual/Ethical

Discusses ethical

Analysis

considerations of AI use in
writing pedagogy, including
issues of authenticity.

Doğan, D. et
al.

2025

Empirical

Studies the effectiveness of
AI practices in pre-school
social sciences education.

73

Eber, P. A., &

2007

Conceptual/Practical Guide

Parker, T. S.

Provides guidance on
assessing student learning by
applying Bloom's
Taxonomy.

Erkılıç, E., &

2025

Cifci, I.

Conceptual/Ethical

Examines ethical issues and

Analysis

violations of using chatbots
for academic writing and
publishing.

Filippi, S., &

2024

Systematic Review

Motyl, B.

Provides a systematic review
of LLMs in engineering
education with practical
adoption suggestions.

Guizani, I. et

2025

al.

Systematic Literature

Reviews issues and solutions

Review

for implementing LLMs in
higher education.

Horikoshi, K.

2025

Empirical/Developmental

et al.

Presents "Learning stream
plot" for classroom activity
visualization using daily
learning log data.

Imran, M., &

2024

Review/Conceptual

Reviews Google Gemini as a

Almusharraf,

next-generation AI

N.

educational tool.

Issabek, A. et
al.

2025

Empirical

Studies the effects of
demographic factors on
learner flow experience in
gamified educational
quizzes.

74

Keith, J. R. et

2025

Conceptual/Pedagogical

al.

Discusses harnessing
generative AI in chemical
engineering education
through critical engagement.

Klar, M.

2025

Mixed Methods Study

Explores K-12 students'
perceptions, interaction
patterns, and learning
support with generative AI
chatbots.

Lang, M. et al.

2024

Conceptual/Applied

Examines LLMs as

Research

"educational assistants" for
automatic writing of
teaching cases.

Lee, D., &

2025

Systematic Review

Palmer, E.

Systematic review on
prompt engineering in higher
education to inform
curricula.

Lucas, H. C. et

2024

Systematic Review

al.

Systematic review of LLMs
and their implications in
medical education.

Maci, S. M., &

2025

Conceptual/Review

Anesa, P.

Discusses the impact of AI
on discourse analysis,
highlighting challenges and
opportunities.

Mezak

2024

Conceptual/Legal Analysis

Proposes a socio-legal

Matijevic, M.

framework for governing

et al.

LLM usage and application
in education.

75

Nedungadi, P.

2024

Conceptual/Review

et al.

Examines generative AI's
role in achieving SDG 4
(quality education, equity).

Newton, P., &

2024

Scoping Review

Xiromeritis, M.

Reviews ChatGPT's
performance on multiplechoice examinations in
higher education.

Nguyen Van,

2024

Review/Conceptual

V. et al.

Provides an extensive
analysis of LLM integration
revolutionizing education.

Özçelik, N. P.,

2024

Case Study

Case study on ChatGPT's

& Yangın Ekşi,

role as a learning assistant

G.

for cultivating writing skills.

Pack, A. et al.

2024

Empirical/Quantitative

Examines LLMs for
automated essay scoring of
English language learner
writing, focusing on validity
and reliability.

Parakh, D. et

2024

Conceptual/System Design

al.

Proposes an adaptive
personalized learning system
with generative AI.

Praveena, T., &

2025

Conceptual/Review

Anupama, K.

Discusses the transformative
potential of AI in English
language instruction.

Reicher, H. et
al.

2025

Conceptual/Developmental

Proposes a generative AIempowered digital tutor for
higher education courses.

76

Ruiz-Rojas, L.

2024

Empirical/Survey

I. et al.

Examines the adoption of
generative AI tools for
collaborative working and
critical thinking in higher
education.

Sarangi, S. et

2024

Empirical/Survey

al.

Studies radiology
postgraduate students'
engagement with LLMs for
educational purposes.

Shahzad, T. et

2025

Comprehensive Review

al.

Provides a comprehensive
review of LLM issues and
solutions in learning
environments.

Shen, M. et al.

2025

Qualitative Study

Qualitative study on
integrating AI into nursing
education, focusing on
prompts, privacy, and
personalized learning.

Sun, X. et al.

2025

Conceptual/Review

Discusses advancing total
quality management in
higher education through AI
and big data analytics.

Uğraş, M. et al.

2024

Empirical/Qualitative

Explores primary school
teachers' perspectives on
ChatGPT-supported
education.

Wang, J. et al.

2025

Empirical/Qualitative

Examines how young
scholars cooperate with
LLMs in academic tasks,

77

focusing on individual
differences and task
complexities.
Wu, Q. et al.

2025

Systematic Review

Systematic review of
generative AI in higher
education, covering
opportunities, challenges,
and implications.

Xing, W. et al.

2024

Extensive Analysis/Review

Provides an extensive
analysis of LLM integration
revolutionizing education.

Yan, L. et al.

2023

Systematic Literature

Focuses on practical and

Review

ethical challenges of LLMs
in education.

Yu, Y. F. et al.

2020

Conceptual/Framework

Discusses enhancing

Application

collaborative learning
experiences using
Laurillard's Conversational
Framework.