Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/376462929

Exploring the Capabilities and Possible Applications of Large Language


Models for Education

Conference Paper · October 2023


DOI: 10.1109/ICETA61311.2023.10344166

CITATION READS

1 207

4 authors, including:

Matúš Čavojský Gabriel Bugar


Technical University of Kosice - Technicka univerzita v Kosiciach Technical University of Kosice - Technicka univerzita v Kosiciach
6 PUBLICATIONS 5 CITATIONS 39 PUBLICATIONS 327 CITATIONS

SEE PROFILE SEE PROFILE

Tomas Kormanik
Technical University of Kosice - Technicka univerzita v Kosiciach
3 PUBLICATIONS 1 CITATION

SEE PROFILE

All content following this page was uploaded by Matúš Čavojský on 21 March 2024.

The user has requested enhancement of the downloaded file.


Exploring the Capabilities and Possible Applications
of Large Language Models for Education
Matúš Čavojský∗ , Gabriel Bugár† , Tomáš Kormanı́k‡ , Martin Hasin§
∗† Department
of Electronics and Multimedia Telecommunications
ठDepartment
of Computers and Informatics
Faculty of Electrical Engineering and Informatics
Technical University of Košice, Letná 9, Košice, Slovakia
matus.cavojsky@tuke.sk∗ , gabriel.bugar@tuke.sk† , tomas.kormanik@tuke.sk‡ , martin.hasin@tuke.sk§

Abstract—This research investigates the possible applications summarization, and even creative writing, demonstrating a
of large language models (LLMs) in education, with a special major improvement in large language models’ capabilities [3].
emphasis on ChatGPT, an OpenAI-developed chat bot model. It When ChatGPT was launched on November 30, 2022, it was
demonstrates the usefulness of ChatGPT as a learning aid and
research assistant while addressing ethical and pedagogical con- considered an unprecedented technological revolution in NLP.
cerns such as cognitive offloading, academic integrity, and critical The chat bot model was developed by OpenAI and attracted
thinking. The study covers a wide range of educational settings, over one million users in just five days, causing site issues,
including content development, learning a foreign language, and generating widespread attention and discussion globally
problem solving, and literature review assistance. Furthermore, [4]. In February 2023, Microsoft unveiled the ChatGPT search
it presents ZeroGPT, a plagiarism detection tool for recognizing
GPT-generated content and summarizes the findings to highlight engine in the Edge browser, as a component of extensive AI-
LLMs’ ability to improve educational experiences. driven enhancements made to both its Bing search engine and
Index Terms—Educational Digital Transformation, Feynman Edge browser. This update allowed users to conduct online
Technique, Large Language Models, ChatGPT, Integration of searches utilizing the chat bot’s input, made possible by the
Educational Technology, Prompt Engeneering incorporation of OpenAI’s GPT-4 model into Bing, resulting
in a ChatGPT-like search experience. Bard is a large language
I. I NTRODUCTION model from Google AI, trained on a massive dataset of text and
code. It was released to the public on March 21th, 2023, and
Chatbots based on Artificial Intelligence-Generated Content can generate text, translate languages, write different kinds of
(AIGC) technology have a long history dating back to Eliza creative content, and answer questions in an informative way.
in the 1960s, which demonstrated the potential of text-based Google Bard’s core UI offers three answer options after every
human-computer interactions despite its simplicity by today’s question, whereas ChatGPT generates responses on-demand.
standards. Natural language processing (NLP) has undergone Regarding the response accuracy, Bard can sometimes provide
a dramatic transformation in recent decades as a result of the truth-looking but misleading information, especially regarding
development of large language models (LLMs). These models, specific dates that are readily available on the internet. The
which can learn directly from raw text, have revolutionized the rapid advancement of these chatbots became critical for effi-
field by allowing a deeper understanding of complex language cient human-computer communication.
structures and nuances that were previously difficult to model. This conference paper focuses on demonstrating ChatGPT’s
More recently, the introduction of the Bidirectional Encoder utility as a learning aid and research assistant. We showcase
Representations from Transformers (BERT) model in 2018 [1] its capabilities as an interactive conversational partner that
marked a significant milestone, establishing new benchmarks facilitates the application of the Feynman Technique, enhanc-
in various NLP tasks. BERT, which was built on Vaswani ing learners’ comprehension of complex concepts. Various
et al.’s [2] transformer architecture, represented a paradigm educational use cases of LLMs are presented, encompassing
shift in NLP by replacing traditional recurrent neural networks tasks like generating summaries, questions, tests, essays, and
(RNNs) with a self-attention mechanism, enabling parallel feedback across different educational levels. Additionally, the
computation of sequences and expanding the model’s ability to ethical and societal implications of these models, including
capture long-range dependencies and contextual understand- concerns related to privacy, bias, and transparency, are ana-
ing. BERT’s success encouraged further research into larger lyzed. We delve into the challenges and risks of employing
and more sophisticated models. GPT-3, the third iteration of LLMs for academic purposes, including plagiarism, cheating,
OpenAI’s Generative Pre-trained Transformer (GPT) model, and cognitive offloading, introducing ZeroGPT, a plagiarism
was released in 2020, boasting an impressive 175 billion detection tool for identifying LLM-generated text. The con-
parameters [3], making it the largest language model of its clusion summarizes our experiments with chat bots, evalu-
time. GPT-3 demonstrated remarkable proficiency across a ates their response generation and use cases across multiple
range of language-related tasks, such as question answering, prompts, and discusses the implications of our findings.
II. R ELATED WORK academic and professional exams in different disciplines, in-
cluding law, medicine, and business [9]. Roivainen [10] in his
The integration of LLMs into educational settings has wit-
article presented exact questions from the IQ test to ChatGPT
nessed a remarkable surge in interest in recent years. Table I by
and estimated its Verbal IQ to be 155, which puts it in the
the Yu [4] compares the significant advantages of ChatGPT’s
top 0.1% of 2,450 people. Gilson et al. [11] show that the
advancements to older chat bot models used for educational
ChatGPT model achieves the equivalent of a passing score
applications.
for a third-year medical student by performing at a greater
than 60% threshold on the NBME-Free-Step-1 (64.4%) and
TABLE I
C OMPARISON OF C HAT GPT TO TRADITIONAL CHAT BOTS FOR 57.8% NBME-Free-Step-2 data sets. As a reaction to such
EDUCATIONAL APPLICATIONS [4] excellent performance, universities and also K-12 schools have
frequently resorted to banning the use of ChatGPT. According
Traditional educational ChatGPT to Lung et al. [6] in an attempt to prevent the use of AI-
chat robots generated assignments, the New York City Department of Ed-
Search mode Keyword based Based on large-scale
ucation passed a rule in January 2023 that prohibited students
retrieval corpus learning from using such tools to conduct plagiarism. RV University in
Bangalore, India has implemented stringent measures which
Response quality Answering questions Similar manual feed-
mechanically back
explicitly forbid students from employing Chat GPT to fulfill
assignments, engage in examinations, or conduct laboratory
Answer scope Limited answerable Significantly tests. Following suit, the Australian state of New South Wales
questions expanding the scope of was the first area to place limitations on students’ use of
answerable questions
ChatGPT, after which measures were implemented in public
Understanding level Context understanding Ability to understand schools in Queensland, Tasmania, and Western Australia with
not supported context a similar goal in mind. Flinders University was among the first
Iterative ability Cannot iterate based Ability to iteratively Australian institutions to implement a policy against computer-
on user feedback optimize based on user generated cheating, which included the unauthorized use of AI
feedback tools such as ChatGPT, Bard, or DALL-E without authoriza-
tion from instructors as well as proper attribution or citations.
As stated in an OpenAI’s overview [5], ChatGPT is a power- Moreover, Australian universities announced the return of
ful language generation model, but it has some limitations. The closed book pen-and-paper exams and a new emphasis of in-
most significant one is that it can only generate text based on class assessment writing [7].
the input given to it; it has no access to external information or
the ability to browse the internet. This means it cannot provide III. L EARNING USING F EYNMAN T ECHNIQUE
accurate or up-to-date information on a wide range of recent The Feynman Technique is a cognitive learning strategy
topics, and it may be unable to generate responses to complex developed by physicist Richard Feynman. It aims to enhance
or unusual questions. The most significant difference between understanding and retention of knowledge through active en-
ChatGPT and Bing Chat/Google Bard lies in the latter two’s gagement and self-explanation. The technique involves a four-
access to internet resources. Access to the internet enables step process.
Bing Chat and Google Bard to be aware of current events 1) Based on knowledge gained from prior experience,
that occur after September 2021, which are not covered by research or lectures, writing down all the information
ChatGPT’s knowledge base. The differences between these about the selected or focused topic or concept.
tools (ChatGPT, Bing Chat, Google Bard) and their respective 2) Using simple language, explaining the concept as if
advantages can be applied in a variety of ways: Conducting the learner was teaching it to someone with no prior
literature review, data and article analysis and other ways knowledge of the subject.
not researched by us like Language translation and Question 3) While teaching the concept, attention should be directed
answering [6]. Rudolph et al. [7] used questions from a wide towards areas where struggles arise in explaining or
variety of academic disciplines to assess the models, including delving deeper into the subject matter. This makes it
a non-trivial inquiry: ”How can one combine eight instances of easier to identify gaps in understanding, indicating the
the digit 8 to obtain the value 1000?”. Surprisingly, almost all need for further research and clarification.
chatbots solved this problem successfully, with the exception 4) The final step is to revisit the learning materials; it is
of Bard, which humorously denied the possibility of reaching critical to thoroughly review the topic, paying special
1000 through addition. attention to areas where difficulties were encountered in
Moreover, it is becoming increasingly difficult to distinguish explaining. The process involves simplifying explana-
whether a text is machine-generated or human-generated, pre- tions, employing analogies, and establishing connections
senting an additional major challenge to teachers and educators with other relevant concepts to facilitate a deeper under-
[8]. Recently newly released version of ChatGPT, GPT-4, has standing, with steps 2-4 forming an iterative or looping
shown human-level performance on the majority of multiple sequence.
In a recent study on Feynman Technique as a Heutagogical fering contextual information, detailing the task at hand,
Learning Strategy for Independent and Remote Learning [12] and specifying the preferred presentation format.
researchers used a true experimental design with pretest and Prompt: For an upcoming assessment from computer
posttest to compare the learning outcomes of students who networking, create a multiple question multiple answer
applied the Feynman Technique (experimental group) and test with bulletpoints. Mark correct answers.
those who did not (control group) in three grade levels (4, 7, • Other, alternative approaches, to prompts can enhance
and 11). The results showed that the experimental group had the quality of outputs by providing detailed step-by-
significantly higher posttest scores and learning gains than the step instructions, specifying required components, and
control group in all grade levels, indicating that the Feynman employing prompt chaining through separate chat inter-
Technique was effective in improving students’ learning. actions. The prompt chaining can be seen in Table II,
Learners may utilize ChatGPT as an interactive conver- where the first prompt requires the ChatGPT to wait for
sational partner to help them explain complex concepts in second input. This methodology allows for the breakdown
their own words. This pedagogical process assists students of complex ideas into smaller tasks, facilitating better
in clarifying their understanding, identifying knowledge gaps, management of prompts and subsequent outputs.
and consolidating their knowledge. The example of ChatGPT
conversation using Feynman Technique is shown in Table In a recent development, ChatGPT has been updated to
II. ChatGPT correctly recognized a knowledge gap in which enable users to provide custom instructions for the model
the user missed to point out that layer one is in control of to consider in its responses. This update includes the ability
signal encoding/decoding, modulation/demodulation, and bit to use RTF as custom instructions, eliminating the necessity
synchronization. The prompt required the use of uppercase of manually writing RTF instructions for each prompt. Self-
”explain to YOU every” and the sentence ”Wait for my input.” directed learning methodologies have proven to be valuable
at the end, as without them, the model started explaining layers [13] during the response to the COVID-19 pandemic by
1-6 instead of waiting for next input. Learners can also use the empowering individuals to take control of their education in
interactive nature of ChatGPT to seek clarification and pose remote and flexible learning environments. This reflects the
probing questions. Engaging in a conversation with ChatGPT growing interest in large language models in the field of edu-
allows learners to obtain additional explanations and gain cation, where they have the potential to significantly enhance
additional insights that improve their overall understanding learning and teaching experiences across various educational
of the subject matter. ChatGPT can also be used for self- levels, from primary and secondary school to higher education
reflection and review by students. By going back and carefully and professional development environments.
examining previous conversations and the responses provided A. Elementary school students
by ChatGPT, learners can identify areas for improvement and
enhance their understanding of the subject. Large language models provide excellent assistance to stu-
dents in primary school in developing their reading and writing
IV. U SE OF LARGE LANGUAGE MODELS IN EDUCATION skills, including activities like identifying syntactic and gram-
When asking a general question without the use of frame- matical corrections. Furthermore, they help to improve writing
work, the ChatGPT responds in a structured answer, in the style and critical thinking skills. These methods encourage
first sentence repeating the question with slight use of syn- students to engage in analytical assessment of their reading and
onyms. Next, the core of the answer is written where most writing materials by producing thought-provoking questions
of the knowledge and typically the answer lies in. Finally and feedback. Furthermore, they assist in the development of
the summary layer is printed where the ChatGPT summarizes reading comprehension abilities by providing students with
it’s core answer in a two to three sentences. In the context brief overviews and elucidations regarding complicated texts,
of AI writing, it is widely acknowledged that the quality boosting their ability to comprehend and digest challenging
of prompts significantly influences the quality of generated subject matter.
outputs, which is also known as prompt engineering. As • Prompt: Check my paragraph for spelling, grammar, and
such, several frameworks have been developed to enhance the punctuation mistakes.
effectiveness of prompts. • Prompt: Provide basic information on the history of
• RTF (Role, Task, Format): An effective framework for steam machines and sources for further reading.
improving AI-generated outputs involves delineating the • Prompt: Suggest arguments for a debate on the merits
AI’s role, providing a clear outline of the task, and speci- and drawbacks of climate change.
fying the desired format. This approach has demonstrated
notable success in enhancing the quality of generated B. Middle and high school students
content. For middle and high school students, large language models
Prompt: As a university student, write an essay about are useful resources because they assist with language learning
large language models in a summary. and proficiency in a wide variety of courses, including math,
• CTF (Context, Task, Format): A framework that, while physics, language, and literature. Despite some limitations,
less effective, still yields significant results involves of- such as ChatGPT’s occasional issues with basic mathematical
TABLE II
A N EXAMPLE OF USING F EYNMAN T ECHNIQUE WITH C HAT GPT 3.5

Using Feynman Technique, I will explain to YOU every layer of model OSI/ISO used in networking. Correct my mistakes in the explanation and suggest
me topics that I missed. Wait for my input.

Of course! I’d be happy to help you explain the layers of the OSI/ISO model and correct any mistakes in your explanation. Please go ahead and explain
each layer, and I’ll provide feedback and suggest any missing topics.

Start with Layer 1 (the Physical Layer) and work your way up to Layer 7 (the Application Layer). Whenever you’re ready, describe each layer’s purpose
and functions.

Physical layer consists of digital and analog signals sent thru the channel (be it a fibre-cable, metalic or wireless). This layer defines hardware equipment,
frequencies and pulses used to represent binary signals. A network disruption can be caused by e.g. a wire cut.
Data link layer is used to communicate via MAC addresses between devices, sending and receiving frames. Is responsible for error detection and
correction using for example a parity bit. Layer 2 can transfer data reliably and unreliably.
Network layer, also known as L3, uses IP addressing to forward network packets to their destination using one or more networks. Typical protocols are
IPv4 and IPv6.
Transport layer is mostly known for its TCP and UDP suites. L4 is responsible for connection-oriented communication. Transport layer uses port numbers
to establish and keep connection relations. Some of the well-known ports are 0 through 1023 which define common services.
Session layer is responsible for keeping and orderly release of connections. In case of a connection loss this protocol may try to recover the connection.
Presentation layer, also called syntax layer, is mainly for encryption and decryption of traffic and ensures the information that the application layer of
one system sends out is readable by the application layer of another system.
Application layer is used by programs to communicate on the internet. They use protocols such as HTTP, FTP or SMTP to deliver information to remote
systems.

You’ve provided a decent overview of the OSI/ISO model layers, but there are some inaccuracies and omissions in your explanation. Let’s correct those
and provide some additional details:
1) Physical Layer:
• You correctly mention that this layer deals with digital and analog signals.
• It defines not only hardware equipment, frequencies, and pulses but also the physical media (e.g., copper cables, fiber-optic cables, wireless
radio waves).
• Mention that this layer is responsible for signal encoding/decoding, modulation/demodulation, and bit synchronization.
• Network disruptions can indeed be caused by issues like wire cuts, interference, or signal degradation.
2) Data Link Layer:
• You’re right that it communicates using MAC (Media Access Control) addresses.
• It’s responsible for framing data into frames, error detection (often using CRC checks), and flow control.
• Mention that it ensures data integrity within a single link or hop.
• Emphasize that it operates on both reliable (e.g., Ethernet) and unreliable (e.g., Wi-Fi) networks.
3) Network Layer:
• You mention IP addressing, but clarify that it handles logical addressing, such as IPv4 and IPv6.
• It’s responsible for routing packets between different networks and subnets.
• Explain the concept of routers, which are devices that operate at this layer and make forwarding decisions based on the destination IP address.
4) Transport Layer:
• Mention that it provides end-to-end communication and is responsible for end-to-end error recovery.
• TCP provides reliable, connection-oriented communication, whereas UDP offers unreliable, connectionless communication.
• Clarify that it uses port numbers to identify specific services on a device.
• Mention that it also handles flow control and segmentation/reassembly of data.
5) Session Layer:
• Describe its role in establishing, maintaining, and terminating sessions between applications.
• Mention that it helps with checkpointing and recovery in case of a connection loss.
6) Presentation Layer:
• Mention that it is responsible for data translation, encryption/decryption, and data compression.
• Explain that it ensures that data sent by the application on one end can be properly understood by the application on the other end.
7) Application Layer:
• Highlight that this is the topmost layer and is where user applications and services directly interact.
• Give examples of common protocols used at this layer, such as HTTP for web browsing, FTP for file transfer, SMTP for email, and many
others.
• Mention that these protocols define how data is formatted, presented, and exchanged between applications.
These corrections and additions should provide a more comprehensive understanding of the OSI/ISO model and its layers. If you have any further questions
or need additional details, feel free to ask!
queries, these models excel at providing detailed step-by- and paper. If more specific input is given, the request and
step explanations, developing student’s abilities to solve prob- searches can be tailored to meet particular needs, making the
lems, providing detailed explanations, and interesting follow- bot accurate to a greater extent.
up questions. By using these models to create personalized
practice questions and tests, students can improve their un-
derstanding, contextualization, and memory of the material.
This approach fosters an understanding of the underlying rea-
soning behind solutions and cultivates analytical and creative
thinking, as it guides users without immediately revealing the
answers. An example of a practise test made by ChatGPT for
computer network student is shown in the Table III.
• Prompt: Suggest how to rework this short story (para-
graph) for clarity and coherence.
• Prompt: Identify key physicists in space exploration,
summarize their theories.
• Prompt: Analyze my essay for weak points, propose
alternative views, make a review.

C. Higher education students


ChatGPT requires users to input data in plain text format
within the chat interface, which may present challenges when
dealing with data that is not inherently structured for such
an interface, such as tabular data. In contrast, Google Bard
utilizes links to access data, facilitating easier retrieval and
summarization of scientific papers. This approach allows for
more comprehensive summaries than abstracts and enables
the explanation of specific points within the article. Our
experimentation with Bing Chat revealed the capability to Fig. 1. A proposal of interactive phishing training.
request in-depth analyses of the current browser tab, including
assessments of the approach’s advantages and disadvantages,
reviewer-like reviews, and potential future research directions. D. Professional development using phishing security trainings
Importantly, these inquiries can be posed for any open browser The utilization of the chat system through an application
window, expanding the scope of analysis beyond traditional programming interface integration offers the opportunity for
text-based inputs. ChatGPT can help to research a topic by interactive engagement of email service users with examples
generating article summaries or generating a list of relevant of phishing emails. With the growing prevalence of phishing
sources based on a specified topic or keyword as stated in attacks, organizations invest significant resources in training
Lund et al. [6]. However, our research indicates that ChatGPT users to mitigate the risks associated with such fraudulent
sometimes generates inaccurate references, unlike Bing Chat emails. Although machine learning-based tools can effectively
and Google Bard, which, due to internet connectivity, can detect anomalies in emails, they have thus far lacked the
perform comprehensive research, offer proper citations, and capability to effectively communicate these findings to users.
provide reliable academic references when requested. By leveraging the communication and explanatory capabili-
Universities should make clear to students the types and ex- ties of ChatGPT, the proposed approach holds potential to
tent of cognitive offloading permitted for specific assessments, enhance security measures and automate the evaluation of
with AI increasingly incorporated into higher level tasks. employee susceptibility to phishing attacks. In this study, an
• Prompt: This paragraph is a rough outline for my es- email sourced from the official website of CSIRT (Computer
say on windmills. Refine my arguments, restructure the Security Incident Response Team Slovakia) was provided as
paragraphs, and generate a draft of the essay. input to the Bing Chat tool, prompting an analysis of the
• Prompt: Develop a research question and hypothesis reasons underlying its fraudulent nature directly from the open
about the correlation of security and ease of use. website.
• Prompt: Propose novel ideas, arguments, theories for my
thesis on security of email agents. Prompt: Is the email on this web page legitimate or
• Prompt: Find relevant research papers for a specific
phishing? Explain why.
topic. Bing Chat effectively identified the provided email as a
An example of fully automatic search with Bing chat bot is phishing attack and proceeded to provide an explanatory
shown in Figure 1. Each link is redirecting to correct website analysis of its typical characteristics and objectives. Further-
TABLE III
A N EXAMPLE OF A COMPUTER NETWORKING PRACTISE TEST GENERATED BY THE C HAT GPT 3.5

Create a test for computer networking student with 3 questions each being multiple-answer (from 1 to all correct) type from EIGRP Cisco course. Do
not write answers, grade my next answer.

Certainly, here’s a test with three multiple-answer questions related to the Enhanced Interior Gateway Routing Protocol (EIGRP) in a Cisco networking
course. Please provide your answers, and I will grade them.

Question 1: Which statements about EIGRP (Enhanced Interior Gateway Routing Protocol) are true? (Select all that apply)
1) EIGRP is a distance-vector routing protocol.
2) EIGRP uses the DUAL (Diffusing Update Algorithm) to calculate the best path.
3) EIGRP supports both IPv4 and IPv6 routing.
4) EIGRP uses hop count as its metric by default.
5) EIGRP exchanges periodic full routing table updates.
Question 2: What are the main benefits of EIGRP? (Select all that apply)
1) Rapid convergence
2) Support for VLSM (Variable Length Subnet Masking)
3) Scalability through route summarization
4) Compatibility with OSPF (Open Shortest Path First)
5) Native support for IPv6
Question 3: In EIGRP, what is the purpose of the ”hello” packets? (Select all that apply)
1) To establish and maintain neighbor adjacencies.
2) To advertise network reachability information.
3) To calculate the composite metric for route selection.
4) To detect and recover from link failures.
Please provide your answers for these questions, and I’ll grade them accordingly.

Question 1 - A, B, C

Question 2 - A, B, C

Question 3 - A, D

Let’s grade your answers:

Question 1: Correct answers: B, C


• B is correct: EIGRP uses the DUAL (Diffusing Update Algorithm) to calculate the best path.
• C is correct: EIGRP supports both IPv4 and IPv6 routing.
Question 2: Correct answers: A, B, C
• A is correct: EIGRP offers rapid convergence.
• B is correct: EIGRP supports VLSM (Variable Length Subnet Masking).
• C is correct: EIGRP allows scalability through route summarization.
Question 3: Correct answers: A, D
• A is correct: ”Hello” packets in EIGRP are used to establish and maintain neighbor adjacencies.
• D is correct: ”Hello” packets also play a role in detecting and recovering from link failures.
You got all the correct answers for these questions! Well done. If you have any more questions or need further clarification, feel free to ask.

Why in Question 1 the A answer was not correct?

In Question 1, option A (”EIGRP is a distance-vector routing protocol”) is not correct because EIGRP is not a pure distance-vector routing protocol. It’s
considered an advanced distance-vector routing protocol or a hybrid routing protocol.

EIGRP does incorporate some features typically associated with distance-vector protocols, such as the use of distance (metric) for route selection, but it
also includes features commonly found in link-state protocols, like maintaining a topology table.

The main reason EIGRP is not categorized as a pure distance-vector protocol is that it uses the DUAL (Diffusing Update Algorithm) to calculate the best
path, which is a significant departure from the simple distance-vector algorithms like RIP. The DUAL algorithm allows EIGRP to converge quickly and
provide loop-free paths, making it more advanced than traditional distance-vector protocols.

So, while EIGRP shares some characteristics with distance-vector protocols, it’s more accurately described as an advanced or hybrid routing protocol.
That’s why option A is not considered correct in this context.
more, the system provided a comprehensive breakdown of the
indicators that signal a phishing attack within the given email:
1) The sender’s address is not a valid domain name for
Microsoft Outlook, but for gmail.sk, which is suspicious.
The correct domain name should be “outlook.com” or
“microsoft.com”.
2) The email claims that the recipient has received a voice
mail, but the attachment is a zip file, not a wav file. This
could be a way to hide malware or ransomware.
3) The link to download the attachment is not from out-
look.com, but from outlook.gmail.sk, which is also
suspicious and could lead to a fake or compromised
website.
4) The email does not address the recipient by name, but
by email address, which is impersonal and generic.
5) The email does not provide any details about the caller,
Fig. 2. A proposal of interactive phishing training.
the message, or the reason for the voice mail, which is
vague and unprofessional.
The initial indicator of a phishing email was accurately
et al. [14] define ”Cognitive Offloading” as any physical
identified, revealing that the email purportedly originated
action to alter the information processing requirements of a
from the Miscorsoft Outlook system as a system message
task to reduce cognitive demand, which can be as simple as
to the email box user, while the actual sender was ”no-
writing something down to avoid the need to remember it later.
reply@gmail.sk”. This scenario assumes the legitimacy of the
While these models can provide vast amounts of information
gmail.sk domain as an email service provider. Changing the
quickly, they may inadvertently encourage users to rely too
Bing Chat setting from ”Precise” to ”Balanced” resulted in
heavily on them for cognitive tasks. This overreliance can
an erroneous classification of the recipient’s email address,
lead to a passive consumption of information without the
”chucknorris@gmail.sk,” as an attack targeting a larger user
engagement of critical thinking and problem-solving skills.
population. Consequently, Bing Chat was solely utilized in
Students in educational environments, for example, may be
”Precise” mode.
The second significant indicator correctly flagged the chat tempted to use these models as a shortcut for learning,
tool and subsequently marked the attachment as a zip file, lowering their active cognitive involvement in the learning
resembling a voice mail. The third indicator of a suspicious process. Therefore, understanding the impact of LLMs on
email was the domain to which the link directed, as a legiti- cognitive offloading and developing strategies to balance their
mate system email should not point to another domain such as use with active cognitive engagement is essential to maximize
outlook.gmail.sk, which may indicate a fake or compromised their benefits while mitigating excessive dependence by using
website. The final two indicators, namely the absence of the anti-GPT tools where student’s critical thinking or problem-
email address and contextual message details, were deemed solving is required.
less significant since it is assumed that a genuine system- ZeroGPT [15] is a plagiarism detection tool that differ-
generated email would not contain such information. However, entiates content created by AI tools from human-authored
a legitimate email might include additional specifics and the text, using perplexity as a metric to assess the generalization
purpose of the message. capabilities of the text’s creator, which can be particularly
Utilizing keywords, the large language model can serve as beneficial for university teachers and educators. To test the
an effective detector and explainer of phishing emails. This accuracy of ZeroGPT, we used the RTF (Role, Task, Format)
capability allows for the identification and highlighting of framework to reduce the likelihood of ChatGPT’s test being
specific keywords, enabling users to comprehend the ratio- detected by the ZeroGPT. However, the essay about proactive
nale behind marking an email as illegitimate. Consequently, and reactive cybersecurity received a score of 62.22% nev-
users can assess the email’s trustworthiness based on the ertheless, as shown in Figure 3. Moreover, the detection of
provided output and offer feedback, thereby facilitating the text translated into foreign language but generated by LLMs
enhancement of the detection and reasoning of the model. An poses a significant challenge for anti-GPT tools, reducing
example of the output transformed to the proposed approach the probability of detection even more. These tools have
is illustrated in Figure 2. the primary goal to identify and flag content generated in
the source language, and they frequently fail to differentiate
V. D ISTINGUISHING BETWEEN TEXT GENERATED BY THE machine-generated translations from human-authored text. The
LLM AND STUDENT reason is that large language models have high linguistic
The proliferation of large language models poses a potential proficiency, allowing them to produce translations that closely
threat in the context of excessive cognitive offloading. Risko resemble human-written content in terms of fluency, grammar,
and context. For example, a text generated by ChatGPT information. Learners should critically evaluate the responses
is translated into Slovak using an online translator, and if provided by ChatGPT and seek additional resources or expert
ZeroGPT detects a significant match, a paraphraser tool can guidance (preferably from humans) when necessary.
be used to reduce the likelihood of raising a suspicion.
ACKNOWLEDGMENT
This work was supported by the Ministry of Education,
Science, Research and Sport of the Slovak Republic, and the
Slovak Academy of Sciences under Grant VEGA 1/0685/23
and by the Slovak Research and Development Agency under
Grant APVV SK-CZ-RD-21-0028.
R EFERENCES
[1] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training
of deep bidirectional transformers for language understanding,” in
Proceedings of the 2019 Conference of the North American Chapter
of the Association for Computational Linguistics: Human Language
Technologies, Volume 1 (Long and Short Papers). Minneapolis,
Minnesota: Association for Computational Linguistics, jun 2019, pp.
4171–4186. [Online]. Available: https://aclanthology.org/N19-1423
[2] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez,
L. Kaiser, and I. Polosukhin, “Attention is all you need,” 2023.
[3] T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal,
A. Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert-
Voss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. M. Ziegler,
J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray,
B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever,
and D. Amodei, “Language models are few-shot learners,” 2020.
[4] H. Yu, “Reflection on whether chat gpt should be banned
by academia from the perspective of education and teaching,”
Frontiers in Psychology, vol. 14, 2023. [Online]. Available:
https://www.frontiersin.org/articles/10.3389/fpsyg.2023.1181712
[5] J. Deng and Y. Lin, “The benefits and challenges of chatgpt:
Fig. 3. An essay about proactive and reactive cybersecurity in a scientfic An overview,” Frontiers in Computing and Intelligent Systems,
researcher words inspected by ZeroGPT. vol. 2, no. 2, p. 81–83, Jan. 2023. [Online]. Available:
https://drpress.org/ojs/index.php/fcis/article/view/4465
[6] B. D. Lund and T. Wang, “Chatting about chatgpt: how may ai and gpt
Despite the precise and relevant information provided by impact academia and libraries?” Library Hi Tech News, vol. 40, no. 3,
large language models, student assessments frequently place pp. 26–29, 2023.
more emphasis on evaluating students’ ability to access and [7] J. Rudolph, S. Tan, and S. Tan, “War of the chatbots: Bard, bing chat,
chatgpt, ernie and beyond. the new ai gold rush and its impact on higher
organize predefined knowledge than on promoting creativity, education,” Journal of Applied Learning and Teaching, vol. 6, no. 1,
critical thinking, and problem-solving abilities. 2023.
[8] D. R. E. Cotton, P. A. Cotton, and J. R. Shipway, “Chatting and
VI. C ONCLUSION cheating: Ensuring academic integrity in the era of chatgpt,” Innovations
in Education and Teaching International, vol. 0, no. 0, pp. 1–12, 2023.
In recent years, with the advent of deep learning techniques, [Online]. Available: https://doi.org/10.1080/14703297.2023.2190148
LLMs have emerged as a powerful tool for NLP. These models [9] OpenAI, “Gpt-4 technical report,” 2023.
can learn directly from raw text, allowing them to capture [10] E. Roivainen, “I gave chatgpt an iq test. here’s
what i discovered,” March 28 2023. [Online]. Avail-
complex language structures and nuances that were previously able: https://www.scientificamerican.com/article/i-gave-chatgpt-an-iq-
difficult to model. The GPT models, in particular, have pushed test-heres-what-i-discovered/
the boundaries of what is possible with LLMs, with the largest [11] A. Gilson, C. W. Safranek, T. Huang, V. Socrates, L. Chi, R. A.
Taylor, and D. Chartash, “How does chatgpt perform on the united
models reaching up to hundreds of billions of parameters. The states medical licensing examination? the implications of large
application of LLMs in education is an interesting field for language models for medical education and knowledge assessment,”
research, offering a wide range of opportunities to enhance JMIR Med Educ, vol. 9, p. e45312, Feb 2023. [Online]. Available:
http://www.ncbi.nlm.nih.gov/pubmed/36753318
students’ educational experiences as well as assist teachers [12] E. P. Reyes, R. M. F. L. Blanco, D. R. L. Doroon, J. L. B. Limana, and
in achieving their objectives. In response to the COVID-19 A. M. A. Torcende, “Feynman technique as a heutagogical learning strat-
pandemic, educational institutions’ paradigm shift toward dis- egy for independent and remote learning,” Recoletos Multidisciplinary
Research Journal, vol. 9, no. 2, pp. 1–13, 2021.
tance learning has highlighted the importance of self-directed [13] L. Gerard, K. Wiley, A. H. Debarger, S. Bichler, A. Bradford, and M. C.
learning methodologies such as ChatGPT along with the Linn, “Self-directed science learning during covid-19 and beyond,”
Feynman Technique. Although LLMs can provide information Journal of science education and technology, pp. 1–14, 2022.
[14] E. F. Risko and S. J. Gilbert, “Cognitive offloading,” Trends in Cognitive
with accuracy and relevance, they cannot replace the essential Sciences, vol. 20, no. 9, pp. 676–688, 2016. [Online]. Available:
human traits of creativity, critical thinking, and problem- https://www.sciencedirect.com/science/article/pii/S1364661316300985
solving, hence teachers should utilize them as supplementary [15] E. Tian, “Zerogpt,” Jan. 2023. [Online]. Available:
https://www.zerogpt.com
tools rather than replacements. While ChatGPT can serve as
a valuable learning aid, it should not be the sole source of

View publication stats

You might also like