Volume 9, Issue 8, August – 2024 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24AUG1297

Ethical Considerations for Companies Implementing

LLMs in Education Software
Mekam Kontche Steve
IUT (university institute of technology) of Douala

Technology will never Replace Great Teachers, but Technology in the Hands of Great Teachers is Transformational
— George Couros

Abstract:- Large Language Models (LLMs) have domains, particularly in education with functionality like
revolutionized natural language processing, offering personalized learning, automates administrative tasks, virtual
significant advancements in educational software through tutors,and various AI-driven tools, However, alongside these
promising innovations come significant ethical challenges
applications like personalized learning and virtual
that companies must carefully address to ensure responsible
tutoring. This position paper investigates the ethical deployment an Ai integration in their offerings for
considerations for companies integrating LLMs into educational settings.it therefore becomes increasingly
educational tools. Key issues include data privacy, with a important to critically examine the ethical considerations of
focus on safeguarding sensitive student information using LLMs in education software. The purpose of this paper
against breaches while ensuring transparency and is to discuss the potential impacts of this new generation of
consent. The paper highlights the risk of misinformation, software and to explore the ethical challenges that arise from
integrating AI into the education sector.
as LLMs might generate incorrect or misleading content
that could affect students’ learning. It also addresses  In light of these Challenges, we Address the Questions:
concerns about algorithmic bias, which can lead to unfair
treatment of students from diverse backgrounds, and the  How can companies ethically implement LLMs in
potential over-reliance on AI, which may undermine education software while mitigating risks and ensuring
critical thinking and human oversight. Additionally, the equitable access?
paper explores the challenge of equitable access to LLM-  What safeguards can companies implement to prevent
based technologies, particularly in underserved over-reliance on AI in education, ensuring human
oversight and critical thinking?
communities. The analysis concludes with practical
 What professional development opportunities should be
recommendations for companies, including robust data provided to educators to help them effectively incorporate
protection measures, balanced AI integration with human LLMs software into their teaching practices?
oversight, and strategies to enhance access for all  How should educational leaders assess the effectiveness
students. By emphasizing these ethical challenges, the of LLMs software in the classroom, including their impact
paper aims to guide responsible AI implementation in on student achievement and engagement?
education, ensuring that technological advancements  To what extent should LLMs be integrated in school
activities and curriculum?
benefit all learners fairly and effectively.
By critically analyzing the ethical implications of LLMs
Keywords:- Large Language Models — Artificial Intelligence in education, this study aims to contribute to the ongoing
— Education — Educational Technologies — Responsible AI. conversation on responsible AI deployment.
I. INTRODUCTION II. OVERVIEW
LLMs an acronym for large language models are models  Brief History and Development of LL and LLMs
that have been trained on a vast corpora of text data, enabling The research of LM has received extensive attention in
them to predict the next word in a sentence and generate the literature, which can be divided into four major
coherent, contextually relevant text across a wide variety of development stages:
tasks [1].The rapid advancement of Large Language Models
(LLMs) in AI has ushered in transformative changes across
 Statistical Language Models (SLMs) in the 1990s.
various sectors, The GPT (Generative Pre-trained
These models were based on statistical learning methods
Transformer) [1] model developed by OpenAI was the first
and utilized the Markov assumption to predict the next
large language model that was publicly released in 2018 ,
word in a sentence using recent context. Known as n-gram
since then new startups have begun sprouting like
models, such as bigrams and trigrams, SLMs were widely
mushrooms, introducing innovative ideas across various
used in tasks like information retrieval and natural

IJISRT24AUG1297 www.ijisrt.com 1856

Volume 9, Issue 8, August – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24AUG1297

language processing. However, they suffered from the  Pre-trained Language Models (PLMs) emerged,
curse of dimensionality due to the vast number of beginning with models like ELMo, which pre-trained
transition probabilities required for higher-order models, bidirectional LSTM networks to capture context-aware
leading to data sparsity issues. To address this, techniques word representations. Later, BERT introduced the use of
like back-off estimation and Good-Turing estimation the Transformer architecture with self-attention
were introduced to smooth data and improve accuracy. mechanisms, setting a new standard for NLP tasks. This
"pre-training and fine-tuning" paradigm became
 Neural Language Models (NLMs), which leveraged dominant, with models like GPT-2 and BART refining
neural networks such as multilayer perceptrons (MLPs) the process through different architectures and pre-
and recurrent neural networks (RNNs) to model word training strategies.
probabilities. A key advancement was the introduction of
distributed word representations, which represented  Large Language Models (LLMs) took pre-trained
words as vectors based on aggregated contextual features. models to an unprecedented scale. By significantly
This innovation led to powerful new techniques like increasing model sizes—such as GPT-3's 175 billion
word2vec, which significantly improved performance parameters—LLMs demonstrated emergent abilities,
across NLP tasks by embedding words into a shared surpassing smaller models in handling complex tasks.
vector space. This shift marked a broader application of LLMs like GPT-3 showcased capabilities like few-shot
language models in representation learning. learning through in-context adaptation. The release of
ChatGPT, based on the GPT series, further demonstrated
the conversational abilities of LLMs, spurring a sharp rise
in research interest and applications in diverse fields.

Fig 1 Chronological Display of Language Model Evolution with Examples

IJISRT24AUG1297 www.ijisrt.com 1857

Volume 9, Issue 8, August – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24AUG1297

Fig 2 Chronological display of LLM releases: blue cards represent ‘pre-trained’ models, while orange cards correspond to
‘instruction-tuned’ models. Models on the upper half signify open-source availability, whereas those on the bottom half are
closed-source. The chart illustrates the increasing trend towards instruction- tuned models and open-source models, highlighting
the evolving landscape and trends in natural language processing research.[2]

 Current Applications in Education  For university students, AI-powered software

Large language models (LLMs) based software have applications have been integrated into the academic
already show many successful application in real life : experience to assist in more advanced reading, writing,
and research tasks. Tools such as Grammarly and
 For elementary school students, several LLM-based ProWritingAid use AI to enhance writing by providing
software applications have already shown many grammar and style suggestions, while Zotero and
successful implementations, assisting in the development Mendeley aid in organizing research materials and
of reading and writing skills. For example, allowing generating citations. Additionally, Turnitin utilizes AI to
children to read aloud sentences or write with a pen detect plagiarism, ensuring academic integrity. AI-driven
device, the software will then recognize the characters and platforms like QuillBot and ChatGPT further assist
sentences written or spoken by the children, and evaluate students in summarizing texts, generating ideas, and
letter formation, correctness, and pronunciation while improving the clarity of their writing. These tools not only
also providing feedback on pronunciation we can take for streamline the writing process but also offer personalized
example mobile software like Speech Blub by Blub Blub feedback that can help students refine their academic
inc or Writing Wizard:Handwriting by L'escapadou. writing and research skills.Stefan A. D. Popenici1* and
Recent studies of Meridith Lovell and Linda Phillips [3] Sharon Kerr [4] in their recent study highlights the
have even shown successful results after evaluation for potential for AI to fundamentally change governance and
several authorized commercial software for teaching the internal architecture of higher education institution
reading and writing in the primary grades.
 For Educators, AI-powered software applications have
 For middle and high school students, several LLM-based been designed to enhance instructional practices and
software applications have been implemented in various support classroom management. Tools like Kahoot! and
ways to support the learning process. for example Socrative use AI to create interactive quizzes and
generating creation content like quizzes,flashcards,course assessments, providing instant feedback and tracking
resume and even multilingual translation for example student progress. Edpuzzle allows teachers to integrate
ImaginAi by ImaginAi. in a recent work of Dijkstra et al. questions and interactive elements into video lessons,
[4], researchers have used GPT-3 to generate multiple- fostering engagement and understanding. Classcraft uses
choice questions and answers for a reading AI to gamify classroom management, helping to motivate
comprehension task and argue that automated generation and monitor student behavior. Additionally, TeacherKit
of quizzes not only reduces the burden of manual quiz offers AI-driven analytics to manage student records,
design for educators but, above all, provides a helpful tool track performance, and organize lesson plans. These
for students to train and test their knowledge any time applications streamline administrative tasks, enhance
while learning from textbooks and other resources. teaching strategies, and provide valuable insights to
support student learning

IJISRT24AUG1297 www.ijisrt.com 1858

Volume 9, Issue 8, August – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24AUG1297

 Key Challenges and Risks Related to the Application of students rely on accurate information to build their
Large Language Models in Education knowledge. A model might provide inaccurate
The integration of Large Language Models (LLMs) in explanations or present information in a way that is
education introduces numerous opportunities, but also incomplete or biased, leading to confusion and
significant challenges and risks that companies must address misinformation.
to ensure ethical implementation.
 A further risk lies in Equitable Access to LLM-based
 From a company’s perspective, Data Privacy is a educational tools. While these technologies hold great
paramount concern. To deliver personalized learning promise, not all students have the same level of access to
experiences, LLMs based Software require access to the necessary resources, such as reliable internet
extensive datasets that often include sensitive student connections or modern devices. This digital divide is
information, such as academic performance, behavioral particularly pronounced in underserved communities,
data, and personal identifiers.Moreover, storing such where students may struggle to engage with AI-driven
large volumes of data makes companies a prime target for learning tools due to financial constraints or inadequate
cyberattacks, heightening the risk of data breaches. Even infrastructure. As LLM-based software becomes more
with advanced encryption and security protocols, the integrated into education, there is a real concern that these
sheer volume of sensitive information processed demands disparities will widen the achievement gap, leaving
constant vigilance. The challenge lies in balancing the behind students who lack access to the technology needed
need for personalization with the obligation to minimize to fully participate in AI-enhanced learning environments.
data collection and ensure its secure handling.
III. PRACTICAL RECOMMENDATIONS
 Another critical issue is the risk of Over-Reliance on AI.
As companies design increasingly sophisticated AI tools The following recommendations provide a
for education, there is a tendency for both students and comprehensive guide for companies seeking to navigate these
educators to depend too heavily on AI-generated complexities and implement LLM-based educational
solutions. This over-reliance can diminish the role of software responsibly.
critical thinking and human oversight in the educational
process, which is problematic from an ethical standpoint.  Development and implementation of robust data privacy
For instance, if teachers overly depend on AI for grading, and security policies that clearly outline the collection,
feedback, or even instructional content, they may neglect storage, and use of student data in compliance with
their role as facilitators of deeper understanding and regulation (GDPR, HIPAA, FERPA) and ethical
critical analysis. Similarly, students might turn to AI- standards. Companies should also be upfront and
generated answers instead of developing independent transparent towards students and their families about the
problem-solving skills. Companies face the challenge of data collection, storage, and use practices, with obligatory
designing software that enhances learning without consent before data collection and use. Educating the
replacing the human element, ensuring that AI serves as a school staff and students about the data privacy and
supportive tool rather than a substitute for traditional security policies, regulations, ethical concerns and best
teaching methods. practices to handle and report related risks can also
improve Ai-Software adoption.
 Bias and Fairness : LLMs are trained on large datasets
that may contain inherent biases related to race, gender,  Incorporate safeguards to prevent over-reliance on AI by
socioeconomic status, and cultural norms. These biases integrating human oversight features like manual
can manifest in the software's outputs, potentially leading overrides and teacher controls. Companies should offer
to unfair treatment or misrepresentation of certain student training to educators on AI’s limitations and encourage
groups. For instance, students from non-dominant critical thinking, ensuring that AI supports rather than
linguistic or cultural backgrounds may receive responses dominates the learning process.
that do not reflect their experiences or realities,
contributing to a learning environment that is  Develop solutions that cater to a wide range of
unintentionally exclusionary. The persistence of such technological environments, including offline
biases in educational software poses serious ethical capabilities, tiered pricing models, and multilingual
concerns, as it can perpetuate inequality and hinder the support. Companies should aim to ensure that AI-driven
goal of providing an inclusive and equitable education for tools are accessible to underserved communities and that
all students. the digital divide is narrowed by making education tools
inclusive and cost-effective.For example, they could
 Misinformation and Content Accuracy:The ability of reduce the price of their offerings and provide their
large language models to generate human-like text can software in underrepresented languages.
make it difficult for students to distinguish between real
knowledge and unverified information. This can lead to We therefore conclude with UNESCO’s call to ensure
students accepting false or misleading information as true, that AI does not widen the technological and educational
without questioning its validity[6 ,7].This risk is divides within and between countries, and recommended
particularly dangerous in educational settings where important strategies for the use of AI in a responsible and fair

IJISRT24AUG1297 www.ijisrt.com 1859

Volume 9, Issue 8, August – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24AUG1297

way to reduce this existing gap instead. According to the IV. CONCLUSION
UNESCO education 2030 Agenda [8]: “UNESCO’s mandate
calls inherently for a human-centered approach to AI. It aims The integration of Large Language Models (LLMs) into
to shift the conversation to include AI’s role in addressing educational software holds transformative potential for
current inequalities regarding access to knowledge, research enhancing teaching and learning experiences. However, this
and the diversity of cultural expressions and to ensure AI potential comes with significant ethical considerations that
does not widen the technological divides within and between companies must navigate to ensure responsible and equitable
countries.The promise of “AI for all” must be that everyone implementation. This paper has explored key challenges
can take advantage of the technological revolution under way including data privacy, misinformation, algorithmic bias,
and access its fruits, notably in terms of innovation and over-reliance on AI, and equitable access.
knowledge.”
To address these challenges, it is essential for
 Provide educators with foundational AI literacy training, companies to implement robust data privacy policies,
focusing on understanding how LLMs work, their ensuring transparency and securing consent while
limitations, and their potential applications in education. safeguarding sensitive student information. Additionally,
This training should include practical sessions on mitigating over-reliance on AI involves incorporating human
integrating AI tools into lesson planning, assessment, and oversight and promoting critical thinking, ensuring that AI
personalized instruction.Create ongoing professional tools complement rather than replace traditional educational
development programs that encourage peer collaboration, methods. Professional development opportunities for
where educators can share experiences, challenges, and educators should focus on AI literacy, ethical use, and
best practices related to using LLMs Software in their collaborative learning networks to support the effective
teaching. These networks can foster continuous learning integration of LLMs into teaching practices.
and adaptation of AI tools in diverse educational contexts.
Educational leaders must assess the effectiveness of
 Companies' software should be used as supplementary LLMs through standardized assessments, engagement
tools to enhance personalized learning. They can provide analytics, and qualitative feedback from teachers to
tailored feedback, offer additional practice materials, and understand the impact on student achievement and classroom
support differentiated instruction based on individual dynamics. Finally, the extent of LLM integration in curricula
student needs. However, they should complement rather should be carefully balanced to enhance learning without
than replace traditional teaching methods and human compromising the role of human interaction and traditional
interaction.for example administrative tasks such as educational approaches.
grading and generating lesson plans, allowing educators
to focus more on direct student engagement. Their By addressing these ethical concerns with thoughtful
integration in these areas should aim to reduce the and responsible practices, companies can harness the benefits
administrative burden while maintaining the quality of of LLMs while contributing to an equitable and effective
education.It’s important that students are guided to educational environment. This balanced approach will ensure
critically evaluate AI-generated content and not solely that advancements in AI serve to enhance learning
rely on it for their learning. opportunities for all students while upholding the principles
of fairness and integrity in education.
 Implement standardized assessments and benchmark tests
to evaluate the impact of LLM software on student REFERENCES
learning outcomes. Compare pre- and post-intervention
performance to gauge improvements in areas such as [1]. Language Models are Few-Shot Learners by Ilya
literacy, critical thinking, and problem-solving skills. Sutskever et al.
Analyze data to determine if there are significant gains in [2]. A Comprehensive Overview of Large Language
student achievement attributable to the use of Ai driven Models. Humza Naveeda, Asad Ullah Khana,∗, Shi
tools.Additionally, monitor changes in student Qiub,∗ et al
engagement and participation using surveys, interviews, [3]. Commercial Software Programs Approved for
and classroom observations, utilize engagement analytics Teaching Reading and Writing in Primary Grades:
from the software, if available, to track student Another Sobering Reality. Meridith Lovell and Linda
interactions and time spent on tasks.Collect qualitative Phillips
feedback from teachers who integrate LLM software into [4]. Comprehension Quiz Generation using Generative
their instruction through surveys, focus groups, and Pre-trained Transformers. Ramon Dijkstra, Zu l̈ ku ̈f
interviews. This feedback will help understand their Genc ̧, Subhradeep Kayal, and Jaap Kamps. Reading .
experiences, challenges, and perceptions of how the [5]. Exploring the impact of artificial intelligence on
software affects classroom dynamics and supports their teaching and learning in higher education. Stefan A.
teaching practices. Insights from this feedback will D. Popenici and Sharon Kerr
provide a comprehensive view of the software's practical
effectiveness and highlight areas for potential
improvement.

IJISRT24AUG1297 www.ijisrt.com 1860

Volume 9, Issue 8, August – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24AUG1297

[6]. Comparing scientific abstracts generated by ChatGPT

to original abstracts using an artificial intelligence
output detector, plagiarism detector, and blinded
human reviewers. bioRxiv, 2022. Catherine A. Gao,
Frederick M. Howard, Nikolay S. Markov, Emma C.
Dyer, Siddhi Ramesh, Yuan Luo, and Alexander T.
Pearson.
[7]. Chatting and Cheating. Ensuring academic integrity in
the era of ChatGPT. EdArXiv, 2023. Debby R.E.
Cotton, Peter A. Cotton, and J.Reuben Ship- way.
[8]. UNESCO. Education 2030 Agenda. https://
www.unesco.org/en/digital-education/ artificial-
intelligence, 2023. Accessed: 22.01.2023.

IJISRT24AUG1297 www.ijisrt.com 1861