1

GPT (Generative Pre-trained Transformer) – A

Comprehensive Review on Enabling Technologies,
Potential Applications, Emerging Challenges, and
Future Directions
Gokul Yenduri, Ramalingam M, Chemmalar Selvi G, Supriya Y, Gautam Srivastava,
Praveen Kumar Reddy Maddikunta, Deepti Raj G, Rutvij H Jhaveri, Prabadevi B, Weizheng Wang, Athanasios V.
Vasilakos, and Thippa Reddy Gadekallu
arXiv:2305.10435v2 [cs.CL] 21 May 2023

Abstract—The Generative Pre-trained Transformer (GPT) rep- TABLE I

resent a notable breakthrough in the domain of natural language L IST OF KEY ACRONYMS ONLY IF IT IS REPEATED
processing, which is propelling us toward the development of
machines that can understand and communicate using language Acronyms Description
in a manner that closely resembles that of humans. GPT is based AI Artificial Intelligence
on the transformer architecture, a deep neural network designed AR Augmented Reality
for natural language processing tasks. Due to their impressive BERT Bidirectional Encoder Representations from
performance on natural language processing tasks and ability Transformers
to effectively converse, GPT have gained significant popularity
BGN Boneh–Goh–Nissim
among researchers and industrial communities, making them one
of the most widely used and effective models in natural language CNN ConvolutionalNeural Network
processing and related fields, which motivated to conduct this DAP Data Access Point
review. This review provides a detailed overview of the GPT, DLT Decentralized Ledger Technology
including its architecture, working process, training procedures, DL Deep Learning
enabling technologies, and its impact on various applications.
In this review, we also explored the potential challenges and DRL Deep Reinforcement Learning
limitations of a GPT. Furthermore, we discuss potential solutions DR Demand response
and future directions. Overall, this paper aims to provide a EC Edge Computing
comprehensive understanding of GPT, enabling technologies, EU End User
their impact on various applications, emerging challenges, and EAPs Energy Access Points
potential solutions.
5G Fifth-Generation
Index Terms—Generative Pre-trained Transformer, Natural 4G Fourth-Generation
language processing, Artificial Intelligence GPT Generative Pre-trained Transformer
GPU Graphics Processing Unit
HPC High Performance Computing
Gokul Yenduri, Ramalingam M, Chemmalar Selvi G, Supriya Y, Praveen HCI Human Computer Interaction
Kumar Reddy Maddikunta, Deepti Raj G, Prabadevi B are with the School IoT Internet of Things
of Information Technology and Engineering, Vellore Institute of Technol-
ogy, Vellore, Tamil Nadu- 632014, India (Emails: { gokul.yenduri, rama- ML Machine Learning
lingam.m, chemmalarselvi.g, supriya.d, praveenkumarreddy, deeptiraj.g2020, NLP Natural Language Processing
prabadevi.b }@vit.ac.in) NPC Non Playable Character
Gautam Srivastava is with the Dept. of Math and Computer Science, Bran-
don University, Canada, and the Research Centre for Interneural Computing, PLM Pre-trained Language Models
China Medical University, Taichung, Taiwan as well as Dept. of Computer PTM Pre-Trained Models
Science and Math, Lebanese American University, Beirut, Lebanon (email: RNN Recurrent Neural Network
[email protected])
Rutvij H Jhaveri is with the Department of Computer Science and Engi- 6G Sixth-Generation
neering, School of Technology, Pandit Deendayal Energy University, India, TL Transfer Learning
(Email: [email protected]).
VU Virtual Reality
Weizheng Wang is with the Department of Computer Science,
City University of Hong Kong, Hong Kong SAR, China, (E-mail:
[email protected]).
Athanasios V. Vasilakos is with the Center for AI Research
(CAIR),University of Agder(UiA), Grimstad, Norway, (Email:
I. I NTRODUCTION
[email protected]).
Thippa Reddy Gadekallu is with the School of Information Technology Language is the cornerstone of human communication and
and Engineering, Vellore Institute of Technology, Vellore 632014, India, plays a vital role in shaping our interactions with the world.
Lovely Professional University, Phagwara, India, Department of Electrical With the advent of NLP, it has revolutionized the way we
and Computer Engineering, Lebanese American University, Byblos, Lebanon,
Jiaxing University , Jiaxing 314001, China, Zhongda Group, China, 314312 interact with machines. NLP has become a game-changer in
(E-mail: [email protected]). the world of communication, enabling humans to interact with
 2

machines in a more natural way. The evolution of NLP has its version [8], and on generative AI’s [9], no existing reviews
been fueled by the exponential growth of textual data in the are dedicated to providing a comprehensive survey on GPT.
internet. Over the years, NLP has witnessed a significant trans- Therefore, there is a need for a comprehensive review that
formation from simple rule-based systems to complex deep focuses on GPT’s architecture, enabling technologies, potential
learning-based models. Despite the advances, natural language applications, emerging challenges, interesting projects and fu-
understanding and generation have long been a challenging ture directions. These limitations motivated us to conduct this
problem in the field of NLP, largely due to the complex review. Hence, this review will not only help researchers and
nature of human language. However, recent advancements practitioners in this field to gain a better understanding of GPT
have paved the way for the new approaches to tackle these but also provide valuable insights into its potential applications
challenges. One such breakthrough in NLP, is the development and major limitations when conducting the research.
of the GPT [1]. GPT became famous after the launch of
ChatGPT by OpenAI, a research company [2] that focuses on
B. Related Surveys and Contributions
developing AI technologies. GPT is a deep learning model that
is pre-trained on large corpora of text data and can be fine- The GPT model is a type of DL model that uses self-
tuned for specific tasks like language generation, sentiment supervised learning to pre-train massive amounts of text data,
analysis, language modelling, machine translation, and text enabling it to generate high-quality language output. The
classification. The transformer architecture used in GPT is recent advancements in GPT model research can be attributed
a significant advancement over previous approaches to NLP, to the continual improvement of its architecture, increased
such as RNN and CNN. It uses a self-attention mechanism to availability of computing power, and the development of novel
allow the model to consider the context of the entire sentence techniques to fine-tune the model for specific tasks. These
when generating the next word, which improved the model’s advancements have led to the creation of larger and more
ability to understand and generate language. The decoder is powerful GPT models, enabling them to perform a wider range
responsible for generating the output text based on the input of NLP tasks with unprecedented accuracy and fluency. These
representation [3]. GPT models have demonstrated great potential in transforming
GPT can perform a wide range of tasks in NLP. One of various industries like healthcare [10], customer service [11],
its key strengths is in natural language understanding (NLU), financial industry [12] and so on. These applications are
where it can analyze and comprehend the meaning of text, in- enabled by the generation of high-quality and diverse data
cluding identifying entities and relationships in sentences. It’s like large-scale corpora of text data with different fast-growing
also proficient in natural language generation (NLG), which enabling technologies [13], [14]. There are numerous survey
means it can create text output, such as writing creative content papers published to provide a comprehensive overview of
or answering questions in a comprehensive and informative the latest developments in GPT models, insights into the
way. Alternatively, GPT is also code generator, where it can different architectures, training methods, evaluation metrics,
write programming code in various languages, such as Python and highlight the challenges and future directions of this field.
or JavaScript. GPT can also be utilized for question answering, This literature survey aims to review and analyze the key
which means it can provide summaries of factual topics or findings and contributions of the most recent survey papers
create stories based on the input text. Additionally, GPT can published on GPT models, to provide a comprehensive and
summarize a piece of text, such as providing a brief overview up-to-date understanding of the state-of-the-art in this exciting
of a news article or research paper, and it can be used for and rapidly evolving field.
translation, which makes it possible to translate text from Lund et al. [4] presents the potential effects of AI and GPT
one language to another. Overall, GPT’s ability to perform models, specifically ChatGPT, on academia and libraries. They
a wide range of NLP tasks with high accuracy and precision, discussed the capabilities of ChatGPT in generating human-
makes it an invaluable tool for various industries, including like responses and its potential applications. They examine
finance, healthcare, marketing, and more. As NLP technology how AI-powered chatbots and virtual assistants based on
continues to advance, we can expect GPT and other language GPT models can enhance student learning experiences, assist
models to become even more sophisticated and powerful, with research tasks, and improve library services. They also
enabling us to communicate with machines more naturally and address concerns regarding data privacy, biases, and the need
effectively. for ethical guidelines. Overall, this survey paper highlighted
the transformative potential of AI and GPT models while
emphasizing the importance of responsible deployment and
A. Motivation human oversight.
GPT has become a transformative technology in the field Kasneci et al. [5] have reviewed the potential opportunities
of NLP, enabling the rapid development and growth of a wide and challenges of using large language models, specifically
range of industries and applications. Despite its wide adoption ChatGPT, for educational purposes. They highlighted the
and numerous potential applications, there is still much to be benefits and limitations of using such models by discussing
explored and understood about GPT’s capabilities. Although their implications for teaching and learning. In addition, a
there are studies on GPT in the literature related to academia defined strategy and pedagogical approach with a heavy focus
and libraries [4], education [5], GPT models [6], banking and on critical thinking and fact-checking are required while using
corporate communication [7], advancements in chatGPT and such large language models in educational institution. Thus,
 3

they concluded the paper by highlighting the key technical conversational AI systems using PLMs and GPTs.
challenges like copyright issues, biased content creation, user Thus, the comparison of existing surveys on GPT models
dependency, privacy and security, and high-cost language highlighting the growing importance of these models in key
models when such language models are used in the educational areas of NLP and other related fields are discussed here.
sector. Hence, this is the first-of-its-kind survey that presents the
Qiu et al. [6] presented an exhaustive survey of various extensive information, by comparing existing surveys with our
types of GPT models by detailing their working architecture. survey and summarized in Table II.
They discussed the evolution of pre-training methods for NLP,
from language modelling to TL and pre-training on large-
C. Systematic Literature Survey
scale corpora. It also reviews the different types of GPT
models, including word embeddings, contextual embeddings, In this review of GPT, we conducted a thorough literature
and transformer-based models, and discusses their applications review using various reputable sources. Our search was pri-
in various NLP tasks such as text classification, Named Entity marily focused on peer-reviewed journals, and high-quality
Recognition, and machine translation. They highlighted the articles from reputed national and international conferences,
benefits of GPT’s models for the NLP domain, such as its seminars, books, symposiums, and journals. To ensure the
ability to improve model performance with limited annotated credibility of our sources, we referred to well-known archives
data, reduce the need for task-specific feature engineering, and such as Google Scholar and arXiv, and publications from top
enable TL across multiple tasks. They discussed the major databases like IEEE, Springer, Elsevier, Taylor & Francis, and
challenges and limitations of PTMs, such as the risk of bias Wiley. To identify relevant GPT references and publications,
and the lack of interpretability. we used keywords such as NLPGPT, GPT architecture, DL for
George et al. [7] studied the potential impact of GPT-4, GPT, Pretraining GPT, Fine-tuning AI GPT and GPT vertical
the next iteration of GPT models, on communication within applications. We then screened all the retrieved articles based
corporate environments. They discussed how GPT-4 can rev- on their titles, excluding any papers with poor-quality material.
olutionize business communication by enabling more efficient Next, we reviewed the abstracts of the remaining articles to
and effective interactions. They explore various applications determine their contributions. In the final step of our literature
of GPT-4 in corporate settings, such as automating customer review, we extracted the necessary data for our analysis. By
support through AI chatbots that can provide personalized following these phases, we ensured that our study was based
responses and resolve queries in real-time. They also ad- on high-quality and credible sources.
dressed potential challenges and considerations associated
with implementing GPT-4 in corporate settings. These include D. Paper Organization
concerns about data security, privacy, and the need for human
The structure of this paper’s organization is illustrated in
oversight to ensure accurate and ethical communication. Thus,
Fig. 1. Section 2 presents the preliminaries of GPT models
they concluded by emphasizing the transformative potential
such as the definition of GPT, its evolution and architecture,
of GPT-4 in revolutionizing business communication to fully
how it works and presents the comparison of various GPT
harness the benefits of GPT-4 while addressing any potential
models. Section 3 discusses the key enabling technologies for
risks or limitations.
GPT models. The impact of GPT models in various applica-
Zhang et al. [8] presents an extensive survey of generative
tions are presented in Section 4. In Section 5, we highlighted
AI and evaluates the capabilities of the ChatGPT models,
some of the exciting GPT projects that are currently developed.
particularly from GPT-4 to GPT-5. They provided an overview
Section 6 includes open issues, other technical challenges
of generative AI, highlighting its significance in generating
and future research directions in the field of GPT. Finally,
realistic and creative outputs across various domains and
we conclude the paper in Section 7, by summarizing the
evaluate their advancements over previous iterations. They
key findings and contributions of this study. The list of key
analyze the architectural improvements, model size, training
acronyms are listed in Table I.
techniques, and dataset considerations employed in GPT-4
and GPT-5. In addition to it, they presented a comprehensive
comparison of ChatGPT with other state-of-the-art generative II. P RELIMINARIES
AI models, such as OpenAI’s DALL-E and CLIP. Finally, In this section, the evolution of GPT models, the architecture
they concluded with valuable insights into the capabilities of GPT, working process of GPT models are discussed and
and limitations of these models and highlights the broader finally, different versions of GPT models are compared.
landscape of generative AI.
Zaib et al. [9] provides a survey on the latest advance-
ments in GPTS and PTMs for conversational AI applications. A. Generative Pre-trained Transformer
They focused on PLMs and their approaches while building The GPT model produces enormous quantities of pertinent
dialogue-based systems. They also highlighted the potential and complicated machine-generated text from a small amount
use of transformer-based models such as BERT and GPT, of text as input. GPT models can be identified as a language
which have demonstrated good performance in understanding model that mimics human text using a DL techniques and it
NLP generation, and dialogue management. Thus, they con- acts as an autoregressive model in which the present value is
cluded with the significant challenges in the field of developing based on the previous value [15].
 4

SECTION-1:INTRODUCTION

RELATED SURVEYS AND SYSTEMATIC LITERATURE PAPER

MOTIVATION
CONTRIBUTIONS SURVEY ORGANIZATION

SECTION-2:PRELIMINARIES

GPT Definitions Definition-1 Definition-2 Definition-3 Definition-4

Evolution of GPT GPT Architecture How does GPT work? Comparisons of GPT versions

SECTION-3:ENABLING TECHNOLOGIES SECTION-4:IMPACT OF GPT ON VARIOUS APPLICATIONS

BIG DATA
Education Healthcare Industry Agriculture
ARTIFICIAL INTELLIGENCE Travel and Transport Entertainment Lifestyle Gaming

CLOUD COMPUTING Marketing Finance E-Commerce

EDGE COMPUTING

HUMAN COMPUTER INTERACTION INTRODUCTION IMPACT OF GPT ON APPLICATION

CHALLENGES SUMMARY
5G & BEYOND NETWORKS

SECTION-5:PROJECTS

SiriGPT AI Dungeon Copy.ai Bond.AI Viable Polyglot AI Meena

AI Channels Fireflies.AI DeepScribe Uber's Plato Research Dialogue System

SECTION-6:OPEN ISSUES AND OTHER TECHNICAL CHALLENGES

Domain Specific GPTs Computational requirements Explainability and interpretability Data Bias

Multimodal support Robustness Multilingual support Model size Limited understanding

Ethical Concerns Security and privacy concerns

SECTION-7:CONCLUSION

Fig. 1. Organization Chart of the survey.

1) Definition 1: GPTs are language models pre-trained on 3) Definition 3: GPT is a language model developed by
vast quantities of textual data and can perform a wide range OpenAI to help give systems intelligence and is used in such
of language-related tasks [16]. projects as ChatGPT [17].

2) Definition 2: A GPT is a language model relying on DL B. Evolution of GPT

that can generate human-like texts based on a given text-based GPT models have evolved through multiple changes and
input. [17]. breakthroughs in NLP technology. These are some significant
 5

TABLE II
C OMPARISON OF THIS SURVEY WITH THE EXISTING SURVEYS

Applications Enabling Technologies

Ref. Remarks

Cloud Computing

Edge Computing

5G and Beyond
Entertainment
E-Commerce
Agriculture

Healthcare
Education

Marketing
Transport

Lifestyle

Finance
Industry

Big Data
Gaming

HCI
AI
[4] X X X X X X X X X X X X X X X X X They conducted a survey discussing capabilities of
ChatGPT on academia and libraries. Although, Key
challenges of Chatgpt were highlighted, practical
implementation challenges and research directions
were missing.
[5] X X X X X X X X X X X X X X X X X Reviewed the potential opportunities and challenges
of using large language models, specifically Chat-
GPT, for educational purposes. Thus, evolution of
GPT and their preliminaries were not discussed in
this survey paper.
[6] X X X X X X X X X X X X X X X X X They presented an exhaustive survey of various types
of GPT models by detailing their working architec-
ture with benefits and limitations of GPTs. However,
[7] X X X X X X X X X X X X X X X X X Studied the potential impact of GPT4 in business
communication and explore various applications of
GPT-4 in corporate settings by highlighting any
potential risks or limitations. But, how GPT archi-
tecture can be used in corporate is not found with
key enabling technologies.
[8] X X X X X X X X X X X X X X X X X Analyzed the architectural improvements, model
size, training techniques, and dataset considerations
employed in GPT-4 and GPT-5. However, prelimi-
nary details are unedr explored.
[9] X X X X X X X X X X X X X X X X X Recent trends in language models, applications of di-
alogue management, question answering NLP tasks
were discussed along with challenges and future
scope of GPT. Although it covered most of the
technical aspects, the integration challenges to over-
comeare not presented.
Our Survey X X X X X X X X X X X X X X X X X Presents the evolution of GPT models, GPT architec-
Paper ture and its detailed working, key enabling technolo-
gies, significant advancements of GPT models and
their potential benefits in real-life applications, GPT
projects, lessons learnt, open challenges and future
research directions.

turning points in the growth of the GPT model: Before GPT, employs a 12-layer decoder architecture transformer with a
NLP models have been trained on large amounts of annotated self-attention system for training. GPT-1’s capacity to execute
data that is related to a specific task. This had a significant zero-shot performance on different tasks was one of its major
drawback because it was difficult to access the quantity of success as a result of its pre-training. This ability demonstrated
labelled data required to train the model precisely. The NLP that generative language modelling can be used to generalize
models were unable to complete tasks outside of their training the model when combined with a successful pretraining idea.
set since they were restricted to a particluar set of data. To With TL as its foundation, GPT models evolved into a potent
get around these restrictions, OpenAI offered a Generative tool for performing NLP tasks with minimal fine-tuning [22].
Language Model called GPT-1 that was created using un- It paved the way for other models to progress even more in
labeled data and then given to users to fine-tune to carry generative pre-training using larger datasets and parameters.
out subsequent tasks like sentiment analysis, categorization, [18].
and question-answering [18]. This indicates that the model To create a better language model later in 2019, OpenAI
attempts to produce an appropriate response based on input created a GPT-2 using a bigger dataset and more parameters.
and that the data used to train the model is not labelled [19]. The model design and execution of GPT-2 are some of the
Fig. 2 shows the timeline of the evolution of several pre-trained key advancements [23]. With 1.5 billion parameters, it has
models from Eliza, which was created in 1960, to the more 10 times the size of GPT-1 (117 million parameters), and it
current 2022-ChatGPT. has 10 times as many parameters and data [21]. By using
GPT-1 was the first ever model that could read the text and only the raw text as input and utilizing little to no training
respond to queries [20]. OpenAI released GPT-1 in 2018. GPT- examples, it is effective in terms of resolving various language
1 was a major move forward in AI development because it tasks related to translation, summarization, etc. Evaluation of
enabled computers to comprehend textual material in a more GPT-2 on various downstream task datasets revealed that it
natural manner than before. This generative language model excelled by substantially increasing accuracy in recognizing
was able to learn a wide variety of connections and gain long-distance relationships and predicting sentences [24].
immense knowledge on a varied corpus of contiguous text The most recent iteration of the GPT model is GPT-3. It is a
and lengthy stretches [21]. This happened after being trained sizable language prediction and production model created by
on a huge BooksCorpus dataset. In terms of design, GPT-1 OpenAI that can produce lengthy passages of the source text.
 6

Pattern matching & Chatbot to provide chatbot to provide original Helps human chat by
replacement psychological concepts replies on its own entertaining

ELIZA-1960 PARRY-1970 Racter-1980 Jabberwacky-1990

ALICE-1995

Analyze human language

SmarterChild-2001 & synthesize
Bot provides the
conversations in a
Siri -2010 fun manner
intelligent
personal assistant

Google Now-2012 Alexa-2014

CHatGPT-2022
Helps to answer the questions,
performs actions, makes Intelligent personal Assist users in generating
recommendations assistant human-like text based on
given input

Fig. 2. GPT Road Map.

GPT-3 eventually emerged as OpenAI’s ground-breaking AI and news items, GPT 3.5 learned the relations between words,
language software. Simply put, it is a piece of software that can sentences, and various components. It was utilized by OpenAI
create lines on its own that are so distinctive they almost sound to develop several systems that have been tailored to complete
like they were written by a human [25]. The GPT-3 program particular jobs [26]. It collected vast amounts of data from the
is presently accessible with limited access via a cloud-based web, including tens of thousands of Wikipedia entries, posts
API, and access is required to investigate the utility. Since its on social media, and news items, and used that information
debut, it has produced several interesting apps. Its capacity, to learn the relationships between sentences, words, and word
which is about 175 billion parameters big and 100 times larger components [29].
than GPT-2, is a key advantage. It is taught using a corpus of The latest version of the GPT model by OpenAI is GPT-4
500 billion words called ”Common Crawl” that was gathered which is a multimodal big language model. It was launched
from a sizable content archive and the internet [26]. Its other on March 14, 2023, and is now accessible to the general
noteworthy and unexpected capability is its ability to carry out public through ChatGPT Plus in a constrained capacity. A
basic mathematical operations, write bits of code, and carry out waitlist is required to gain access to the business API [10].
clever tasks. As a result, NLP models can help businesses by Using both public data and ”data licensed from third-party
responding more quickly to requests and accurately keeping providers,” GPT-4 was pre-trained to anticipate the next coin as
best practices while minimizing human mistakes [27]. Due a transformer. It was then adjusted with reinforcement learning
to its intricacy and size, many academics and writers have based on input from humans and AI for human alignment
referred to it as the ultimate black-box AI method. Due to the and policy conformance. In comparison to GPT-3, which had
high cost and inconvenience of performing inference, as well context windows of only 4096 and 2049 tokens, respectively,
as the billion-parameter size that makes it resource-intensive, the group created two variants of GPT-4 with context windows
it is difficult to put into practice in jobs [24]. of 8192 and 32768 tokens.
GPT-4 was named as the successor of GPT-3. In the meantime,
several AI models built on GPT-3.5, an updated version of C. GPT model’s architecture
GPT-3, have been surreptitiously released by OpenAI [28].
GPT models are based on neural networks that are used for
GPT-3.5 was trained on a mixture of text and code. From the
NLP tasks, such as language modelling, text classification,
vast amounts of data collected from the web, which includes
and text generation.
tens and thousand of Wikipedia entries, social media posts,
The GPT model’s architecture is based on the transformer
 7

model [30]. The Transformer model uses self-attention the parameters of the pre-trained model to optimize
mechanisms to process input sequences of variable length, performance on a specific downstream task, such as text
making it well-suited for NLP tasks. GPT simplifies the classification or text generation [35].
architecture by substituting encoder-decoder blocks with • Language modeling: Language modelling is a key task in
decoder blocks. GPT model takes the transformer model and the GPT architecture. In the case of GPT, the language
pre-trains it on large amounts of text data using unsupervised modelling task is performed during the pre-training phase
learning techniques. The pre-training process involves of the model. In pre-training, the model is trained based
predicting the next word in a sequence given the previous on a large amount of data using a language model
words, a task known as language modelling. This pre-training objective. It is the task of predicting the next word in
process enables the model to learn representations of natural sequence based on the previous words. It allows the
language that can be fine-tuned for specific downstream model to learn relationships between the words and their
tasks [31]. The following are the components of the GPT meaning in the training data [36].
architecture. • Unsupervised learning: Unsupervised learning is an ML
algorithm which enables the model to learn form unla-
belled data without any human intervention. GPT models
• Input Embedding layer: The embedding layer maps the use unsupervised learning in the pre-training phase to
input tokens (e.g., words or subwords) to continuous understand the relationships between the words and their
vector representations, which can be processed by the context in the training data [37].
transformer blocks [32].
• Positional encoding: Since the transformer blocks do not
have any notion of order or position, positional encoding D. How do GPT models work?
is added to the input embeddings to provide information
about the relative position of tokens. Masking: In some GPT models work by using a transformer which is a neural
cases, masking may be necessary to mask certain input network architecture that processes the input sequences of
tokens (e.g., in language modelling tasks, the model natural language text [38]. The GPT model uses unsupervised
should only use tokens that come before the target learning techniques to pre-train this transformer architecture
token). Transformer blocks: GPT models are based on on a significant amount of text input [39]. The model gains
the transformer architecture. It is designed for NLP tasks the ability to anticipate the subsequent word in a sequence
and has been widely used in applications such as ma- based on the preceding words during pre-training. Language
chine translation, text classification, and text generation. modelling is the process that enables a model to discover
Transformers allow the model to focus on different areas the statistical connections between words and their context
of the input while processing [33]. in training data. Fig. 5 shows the various stages of GPT
• Linear and Softmax Functions: In the GPT architecture, operation. The first step entails supervised fine-tuning, the
the softmax function is commonly used for classification second step involves producing optimal responses to input,
tasks. The softmax function is applied to the output of and the third step involves proximal policy optimization and
the final layer of the model. It generates a probability reinforcement learning.
distribution over a set of output classes. The output of The model can be fine-tuned for particular tasks, like text
the final layer is specifically converted into a set of classification or text production, after pre-training. The model
logits before being normalized with the softmax function. is trained on a smaller dataset that is unique to the work at hand
The normalized values obtained from the model can during fine-tuning, and the model’s parameters are changed
be interpreted as the likelihood or probability that a to maximize performance on that task [8]. Fig. 3 shows the
particular input belongs to each of the output classes. The general transformer architecture of GPT.
query, key, and value vectors for each token in the input When used for text creation, GPT models create text by
sequence are frequently calculated using linear functions anticipating the following word in a series based on the
in the attention mechanism. The output of the multi- previously created words. Depending on how it has been
head attention layer is transformed using them in the modified, the model can produce text that is comparable to
feedforward layers as well. The output layer also employs the input text or that adheres to a certain theme or style. Fig.
linear functions to forecast the following token in the 4 projects the GPT model’s transformer architecture and input
sequence [34]. transformations for fine-tuning different tasks.
• Pre-training: Pre-training is a key component of the
GPT architecture. In pre-training, the model is trained
E. Comparisons of GPT Versions
on a large amount of data in an unsupervised manner
even before fine-tuning the model for specific tasks like There are several versions of GPT models each having their
classification and text generation. own features and capabilities. Table III presents a comparison
• Fine-tuning: Fine-tuning is the process of adapting a pre- of various versions of the GPT models. The table presents
trained neural network model, such as GPT, to a new the following details like year of release of the GPT model,
task or dataset by further training the model on that parameters, tokens generated, input type, features of each
task or dataset. Fine-tuning in GPT involves adjusting model, drawbacks of each model, and the size of each model.
 8

TABLE III
C OMPARSION OF DIFFERENT VERSIONS OF GPT MODEL

Model Tokens Size Parameters Dataset Year Features Input Type Drawbacks
GPT-1 - 12-layer decoder 117M parameters Books corpus 2018 Used mostly for lan- A sequence of tokens Limited Capacity, Limited
guage modelling tasks and words Data, Cannot perform complex
and it is transformer tasks, Limited applications
based
GPT-2 - 10 times the size 1.5B parameters Downstream task 2019 Text generation capa- A sequence of tokens Limited Control, Limited Data
of GPT-1 datasets bilities are improved and words Diversity, Expensive computa-
and a chance for misuse tional requirements, Risk of
improper information
GPT-3 4096 and 2049 100 times larger 175B parameters Common Crawl 2020 Good NLP capabilities, A sequence of tokens Limited Control, Limited Data
tokens than GPT-2 language translation, and words and images Diversity, Lack of explanation,
summarization and and tables Ethical concerns
generation of text
GPT-3.5 maximum token 96 layers similar or larger - 2022 Improves user experi- The input type typi- Limited resources to train,Data
limit of 4096 to- number of ence by delivering more cally consists of text Bias,Lack of Explain-
kens parameters like precise and contextu- data ability,Limited Contextual
GPT-3 ally relevant informa- Understanding,High Inference
tion Latency
GPT-4 8192 and 32768 - 100T parameters - 2023 Creative and technical A sequence of tokens -
tokens writing tasks and words and images
and tables

Generative AI (GAI) models are of different types like uni- yielded valuable insights for the use of advanced models like
modal, cross-modal, and multimodal. The first type is uni- GPT in the field of NLP [42]. The GPT models utilize DL
modal which rely on a single type of input, such as text and big data for natural language generation, with GPT-4
or images. The cross-modal, on the other hand, can process being the most advanced model to date [43]. The training
multiple types of inputs and relate them to each other. The data for GPT models typically include millions or even
Multimodal is the most complex type of AI as it can process trillions of data from a diverse range of sources, such as
and integrate information from multiple modalities, such as books, articles, websites, and social media platforms. This
speech, text, images, and even physical interactions with the large and diverse training data helps GPT models capture the
environment. GPT adopts only unimodal and multimodal types variations in language usage, making them more accurate
where ChatGPT is said to be unimodal, while GPT-4 is and effective at NLP tasks. As a result, GPT models may
multimodal. Fig. 6 is an illustration that distinguishes between be used for a variety of tasks, including question-answering,
unimodal, cross-modal, and multimodal Generative AI models. text summarization, and language translation [44]. Moreover,
Overall, GPT models have demonstrated outstanding per- since GPT models can learn from a variety of data sources,
formance with NLP, by enhancing each iteration and its pre- they can be tuned for certain tasks and domains, making them
decessor’ capabilities. Each model, however, also has its own very adaptive and versatile. GPT model has the potential to
restrictions and drawbacks, such as restricted output control, be utilized for a variety of activities, including the creation
lack of diverse data, and ethical concerns. While selecting a of images and videos in addition to its excellent language
GPT model for a particular task, researchers and developers processing capabilities [45].
should carefully take these factors into account [40].
In detail, this section describes the evolution, and architecture While big data presents numerous benefits to GPT, by
of GPT and compares the different versions and types of GPT. enabling the models to get trained with large amounts of data,
it also presents several challenges [46]. GPT is trained on
III. E NABLING T ECHNOLOGIES a variety of data, large amounts of data, and also sensitive
GPT is a convergence of several technologies. It is enabled data. Thus, ensuring data accuracy, privacy concerns, and
by the latest technologies like Big data, AI, Cloud Computing, ethical use of data are some of the challenges that must be
EC, 5G and beyond networks, and HCI. In this section, we considered. However, with the continuous growth of available
provide an overview of enabling technologies related to GPT. data, GPT models will become even more advanced and
The major technologies that constitute the GPT models are capable of performing increasingly complex tasks [47]. The
depicted in Fig. 7. future of big data as an enabling technology for GPT models
is promising, with the potential to revolutionize the field of
NLP. As technology continues to advance, organizations must
A. Big Data
prioritize ethical considerations and data accuracy to fully
Big data refers to the vast amounts of structured and harness the benefits of big data and GPT models.
unstructured data generated by businesses, individuals, and
machines. The proliferation of new technologies, such as the
IoT, has led to an explosion of data production from sources B. Artificial Intelligence
like social media, sensors, and transaction-based systems [41]. AI refers to the simulation of intelligent behaviour in
machines that are programmed to learn from their experience
The emergence of big data has revolutionized the way to reason, understand natural language, and perceive their
organizations approach data analysis and decision-making. environment [48]. AI gives machines the ability to sense their
The training provided by this massive amount of data has surroundings, deal with what they see, handle issues, and
 9

Output documents or medical literature to better grasp and produce

Probabilities language in that field [4]. Considering dialogue generation,
AI techniques such as reinforcement learning and sequence-
to-sequence models can be used to enable GPT generate
Softmax more natural and engaging dialogue in conversational contexts.
Similarly, AI techniques such as semantic parsing and named
entity recognition can be used to help GPT better understand
Linear the meaning of language and the relationships between words
and phrases. This can enable it generate more accurate and
coherent language [14].
The development and enhancement of GPT model language
Add & Norm production capabilities depend heavily on AI, and GPT’s
capabilities will continue to be growing by continuous research
and development in AI.
Feed
As GPT models become more advanced, there are growing
Forward
concerns about the potential for them in reinforcing biases and
propagate harmful or offensive content [50]. Some of these
Add & Norm concerns also include bias which can lead to unintended dis-
Add & Norm
crimination and unfairness, lack of understanding of the con-
text that can lead to misunderstandings or incorrect responses,
Feed
Multi-Head poor data quality can lead to inaccurate or biased models,
Attention ethical concerns like privacy and autonomy [51]. AI models
Forward
like GPT require significant amounts of computational power
to train and run, which can have a significant environmental
Add & Norm
Add & Norm impact due to their high energy consumption [52].
Though AI has a great deal of promise, it’s critical to be aware
Masked Multi- of the underlying issues and make efforts to fix them to ensure
Multi-Head Head that it is utilized responsibly and morally for GPT.
Attention Attention

Positional
Encoding C. Cloud Computing
+ + Cloud computing refers to the on-demand availability
Positional of computer resources, such as storage, processing power,
Encoding
Input Output
and applications, delivered over the internet [53]. The GPT
Embedding Embedding model’s successes are possible not only because of algorithmic
evolution but also increased computational capabilities i.e.
Inputs Outputs
exponential growth in hardware (computational power, storage
capacity), cloud computing, and related operational software
[54]. The applications for cloud and EC working together
Fig. 3. Transformer Architecture.
such as natural language generation, image completion, or
virtual simulations from wearable sensors see that the work
is made more compute-intensive [55].
take action to reach a particular objective. The importance
GPT models need a lot of computational power to analyze
and capability of AI is growing all the time.
a lot of data, and cloud computing offers the scalability
required to cope with demand spikes. Without worrying about
AI enables GPT models to allow machines to comprehend
the constraints of on-premises hardware, GPT models can
and react to human language. There are several ways in which
rapidly and easily scale up or down as needed with cloud
AI can continue to help improve GPT and make it more
computing [56]. Cloud-based platforms like Amazon Web
powerful and effective in its language generation capabilities
Services (AWS) or Google Cloud Platform (GCP) provide
[49].
access to distributed computing resources that can be used
The following are the several ways through which AI can make
to train GPT. Since cloud computing provides web-based
GPT models more powerful:
solutions and thereby does not require the purchase and
1) Fine tuning maintenance of costly hardware, it can be a cost-effective
2) Dialogue generation choice for a GPT model. By utilizing cloud computing, the
3) Natural language understanding GPT model can only pay for the computing resources it uses
GPT’s model performance on particular tasks can be enhanced [57]. The other added advantage of cloud computing in GPT
by utilizing AI approaches. For instance, it can be trained on is, it gives GPT models the freedom to access computing
a large corpus of text from a particular field such as legal resources whenever it wants, from any location in the world.
 10

Text Task
Prediction Classifier
Classification Start Text Extract Transformer Linear

Entailment Start Premise Delim Hypothesis Extract Transformer Linear

Layer Norm

+ Start Text 1 Delim Text 2 Extract Transformer

Feed Similarity
Linear
Forward Start Text 2 Delim Text 1 Extract Transformer

Layer Norm

Start Context Delim Answer 1 Extract Transformer Linear

+
Multiple Start Context Delim Answer 2 Extract Transformer Linear
Masked Multi
Choice
Self Attention

Start Context Delim Answer N Extract Transformer Linear

Text & Position

Embedded

Fig. 4. Transformer Architecture and Input Transformations for Fine-Tuning on Different Tasks.

1 Supervised fine tuning (Step1) 2 Reward Model (Step 2)

A pretrained GPT model is used, The labelers then wrote an Supervised fine tuning (SFT) model is used, inputs
fine-tuned with labelers by creating appropriate response to the are fed to the finetuned model, several responses
a supervised dataset. input prompt’s were generated 2.1
1.1 1.2
Labeler provides a reward for each of This data is used to
these outcomes, ranks the outcomes train a reward model
GPT model will be fine-tuned using new supervised dataset 1.3 from best to worst 2.3
2.2
After the model is trained in step 1, In this step it
generates optimal responses to input
Proximal Policy Optimization (PPO) Reinforcement Learning (Step 3) 3

A new prompt is sampled from the The PPO model is initialized from
The policy generates the output
trained dataset the supervised policy
3.1 3.2 3.3

The reward model calculates a The reward is used to update the

reward for the output policy using PPO
3.4 3.5
In this step unseen input sequences are passed to the clone SFT model, pass the responses
to the reward model for understanding, how high quality was this response for that input

Fig. 5. How does GPT Work.

This makes GPT models more accessible to users by enabling possess the knowledge and tools necessary to effectively
smooth operation across a variety of gadgets and platforms address security problems and stop data leaks. Cloud-based
[58]. Cloud computing providers offer high security and storage services, such as Amazon S3 or Google Cloud
compliance standards, which can protect the GPT model and Storage, provide scalable and reliable storage for GPT’s data.
its data from online dangers. Cloud service providers also
 11

Data Data
Pre- Pre-
training training

HUMAN HUMAN VISUAL

GPT TEXT RESULTS GPT
INSTRUCTION Input INSTRUCTION Input RESULTS
Output Output

UNIMODAL (CHATGPT) CROSS MODAL(STABLE DIFFUSION 2.1)

INSTRUCTION 1 TEXT RESULTS

Data

Output
Input Pre- VISUAL
INSTRUCTION 2
training RESULTS

GPT
INSTRUCTION 3 AUDIO RESULTS

Human MULTI-MODAL Results

Instructions

Fig. 6. A comparison between unimodal, cross-modal, and multimodal modal GPTs.

Improved input quality

Enhanced user experience
Enhances the usability of
GPT models

Massive data for training

Improves model accuracy
for training
Fine tuning
Faster data transmission
ta
Da

Improved connectivity
HC

Lower latency
Big
I

Good computational
Key power
5G & enabling Artificial Model Optimization
Beyond technologies Intelligence Helps in deployment of
various application
Co
ng

Cl utin
mp e

mp
ou
uti
Co Edg

d g

Reduced latency Provides scalability for GPT

Cost effective models
Security and Cost effective option
compliance Security and compliance
Low bandwidth usage Flexibility

Fig. 7. Enabling technologies of GPT models.

Despite the advantages of cloud computing where it can GPT model needs a sizable amount of computing power and
help GPT models to operate more efficiently, effectively, and data storage. These resources can be accessible online with
securely, there are also a few technological aspects where it cloud computing. As a result, continued operation of the GPT
creates a drawback for GPT [59]. To function properly, the model requires a robust and dependable internet connection,
 12

and any breakdown in connectivity may result in delays or E. 5G and beyond networks
even data loss. There are some security concerns when storing 5G networks represent the latest generation of cellular
sensitive data, such as personal information or trade secrets, in networks that promise faster data speeds, lower latency, and
the cloud which can be risky if proper security measures are the ability to connect a vast number of devices simultaneously
not in place [54]. While cloud computing can be more cost- [66]. 5G and beyond networks enable faster data transmission
effective than building and maintaining an in-house computing speeds than previous generations of cellular networks,
infrastructure, it can still be expensive for long-term use. It also which can help in training and deploying larger and more
suffers issues like performance variability, limited availability complex language models. This can result in faster training
etc., times and better performance. 5G and beyond networks
can provide lower latency than previous generations of
cellular networks, which can reduce the time required for
D. Edge Computing communication between GPT and other devices, such as
servers or other language models [67]. This can improve
The rapid growth of IoT, a large amount of data from the real-time response of the GPT model for applications
several IoT devices, and also cloud services have necessitated that require quick and accurate language processing. 5G and
the emergence of a concept called EC. EC is an open AI beyond networks offer improved connectivity options, such
and distributed design with decentralized computational as increased capacity and more reliable connections, which
power. In EC, there is a lesser need for clients and servers to can help in scaling up the deployment of the GPT model for
communicate over long distances, which lowers latency and large-scale language processing tasks. With the deployment
bandwidth utilization. of 5G and beyond networks, EC is becoming more prevalent.
Instead of depending on centralized data centers, EC entails This means that a GPT model can potentially be deployed
bringing computing capacity and data storage closer to the closer to the end-user, reducing the latency and improving the
consumer [60]. response time for applications that require real-time language
processing [68]. Ultra-Reliable Low-Latency Communication
In GPT, where there is a need for real-time data analysis, EC (URLLC) is a key feature of 5G networks. In the context
plays a major role in faster processing and better efficiency in of GPT language models, URLLC can enable real-time and
producing good results [61]. GPT models are typically large reliable communication between multiple devices, such as
and complex, requiring significant processing power to run. edge devices, cloud servers, and end-users [69].
By deploying GPT models on the edge devices, closer to the
source of data, latency can be reduced in replying to users Though 5G and beyond technology offers potential advan-
who seek information through the GPT models by eliminating tages to GPT models, it is also important to note that the actual
the need to move data back and forth from end devices to impacts of this technology may change depending on how it’s
the cloud. Since EC maintains data near the periphery and implemented and used. 5G enables the access to uncontrolled
away from centralized servers, it can offer improved security access to the Internet,it may attract cybersecurity risks and
and more privacy protections in the case of the requests privacy concerns [70]. Also, as GPT uses a large amount of
made by users through GPT [62]. GPT models utilize a lot data for analysis it could also cause privacy concerns. 5G
of data for learning and thereby the cost of data transfer and beyond networks in GPT models need high infrastructural
also increases with data volume. EC can aid in controlling requirements which is a costly process.
data transfer expenses. EC can also help in lowering the
amount of bandwidth by pre-processing the data even before
transferring it to the cloud. Particularly when analyzing photos F. Human Computer Interaction
or videos, GPT models can produce a lot of data [63]. EC HCI, which is multi-faceted, concentrates on the design of
accelerators, such as graphics processing units (GPUs) and computer technology and, in particular, on how people and
field-programmable gate arrays (FPGAs), can be used to speed computers communicate with each other [71].
up GPT model inference and training. These accelerators can
be integrated into edge devices or edge servers, providing more HCI has a greater influence over GPT models. As a
efficient processing of GPT models. language model, GPT is designed to interact with humans
EC and GPT models make a great combination. Comparative by generating natural language responses to input text. HCI
to cloud data centres, edge devices may have constrained research can help designers create more effective input
computation and storage capabilities [64]. This might limit the mechanisms for the GPT model, such as natural language
scope of GPT models that can be installed on edge devices in interfaces, that allow users to communicate more easily and
terms of size and complexity. Since GPT models handle large accurately with the model [72]. HCI also helps in enhancing
and varied data, EC can also increase security risks and data the GPT model’s user experience by creating interfaces that
privacy concerns. Implementing EC in existing infrastructure are more intuitive and user-friendly. This makes it easy for
can be difficult and require significant investment in hardware, the users to interact with GPT models and understand their
software, and networking components. This can be a barrier responses [73]. HCI also estimates the performance of GPT
for many organizations which are using the GPT model and models by evaluating their responses with real-time users
EC [65]. and identifies the areas where the model needs improvement,
 13

thereby improving its reliability and accuracy. HCI enhances on a vast amount of data, GPT can generate human-like text
the usability of GPT models by reducing the time and effort that is coherent and informative, making it a valuable tool
required for the users to interact with [74]. in developing educational content such as textbooks, study
guides, and course materials. Furthermore, GPT can be used to
While HCI can be incredibly helpful in improving the de- analyze and summarize complex text, which can help educa-
sign and usability of GPT models, there are also some potential tors and students save time and increase comprehension. With
drawbacks to consider. If the research is not conducted with its ability to support NLP applications and create intelligent
a diverse group of users, HCI can introduce biases into the tutoring systems, GPT has the potential to revolutionize the
design of the GPT model. HCI techniques can be expensive way we learn and teach. In this context, following section
and time-consuming. As GPT models become more complex, will explore the different ways in which GPT can contribute
it may become more difficult to design interfaces and input to the education industry and transform the future of learning.
mechanisms that are both effective and user-friendly [75]. HCI
may not always be able to provide the necessary insights or • Intelligent Tutoring: Intelligent tutoring is a teaching ap-
feedback to drive improvements in GPT models. There are proach that uses AI and ML to provide personalized and
also ethical concerns around the use of GPT models, including adaptive instruction. It analyzes student performance data,
issues related to privacy, bias, and the potential misuse of understands their strengths and weaknesses, and gener-
the technology [76]. As GPT models become more complex, ates customized learning paths. It provides immediate
it may become more difficult to design interfaces and input feedback, personalized guidance, and remedial support.
mechanisms that are both effective and user-friendly. It is effective in improving learning outcomes, increasing
student engagement, and reducing learning time. With
IV. I MPACT OF GPT MODELS ON VARIOUS A PPLICATIONS advanced natural language processing capabilities, GPT
GPTs have made significant progress, and its impact is can enhance the personalized and adaptive instruction
being felt across various industries like education, healthcare, provided by intelligent tutoring systems. It can analyze
industry, agriculture, travel and transport, e-commerce, en- natural language input from students, enabling intelligent
tertainment, lifestyle, gaming, marketing, and finance. This tutoring systems to better understand and respond to
section provides valuable insights on the impact of the GPT their queries, needs, and preferences. It can also gen-
models in the aforementioned applications as depicted in Fig. erate personalized feedback and assessment based on
8. the individual learning progress of each student, helping
them to identify and address their knowledge gaps and
improve their performance. GPT can also analyze student
A. Education performance data and generate adaptive learning paths
1) Introduction: Education has been around for centuries, that provide customized instruction and remediation, en-
with traditional education being the most common form. suring that each student learns at their own pace and
Traditional education involves a teacher imparting knowledge achieves their learning objectives. Additionally, it can
to a group of students in a physical classroom. While suc- create interactive dialogue systems that simulate natural
cessful, traditional education can be restrictive and inflexible, conversations between students and virtual tutors, making
limiting students’ ability to learn at their own pace and in learning more engaging, interactive, and personalized
their preferred style. It can also be limited by geography, as [77]. The authors in [76] have identified that GPT-4 model
students need to be physically present in a classroom to learn. outperforms general-purpose GPT-3.5 model as well as
Technology has emerged as a solution to some of these issues, GPTs (Med-PaLM, a prompt-tuned version of Flan-PaLM
allowing for personalized learning experiences and more en- 540B) specailly trained on medical data. The authors
gaging, accessible resources. Online learning platforms, digital have tested GPT-4 models’ ability to explain medical
textbooks, and multimedia tools offer students access to a vast reasoning, personalize explanations to students, and in-
array of resources from anywhere in the world. Technology teractively craft new counterfactual scenarios around a
can also facilitate collaboration and communication among medical case.
students and teachers, leading to a more dynamic and inter- • Learning assistance and material development: Learning
active learning experience. Distance learning, hybrid learning materials are critical in education as they provide a
models, and online classes are examples of how technology structured way for students to acquire knowledge and
can help break down the barriers of traditional education, skills. They can be tailored to meet the needs of diverse
making learning more flexible, efficient, and effective. By learners and make learning more engaging and effective,
integrating technology into traditional education, we can create supporting teachers to create a more dynamic and interac-
a more personalized and effective learning experience, bene- tive learning environment. GPT can contribute to creating
fiting students worldwide. learning materials by automating content generation, pro-
2) Impact of GPT in Education: The field of education is viding multilingual content creation, language correction,
constantly evolving, with advancements in technology playing personalized content creation, conducting topic research,
a significant role in shaping the way we learn and teach. and generating assessments. It saves time and effort for
One such technology that has the potential to transform the educators and publishers, improves the accuracy and
education industry is GPT. As a large language model trained readability of material, and makes learning more engag-
 14

Content Creation
Automated Assessment
Benefits of Integrating GPT Writing Assistance
Intelligent Tutoring
Education
Drug Discovery
Patient Diagnosis
Diesease Prediction
Personalized Medicine
Big Data Healthcare

Improving Crop Yield

Pest Control Assistance
Identifying Diseases, Data
Analysis and Prediction
Agriculture
Diet Planner
Travel Guide and Trip Advisor
Personalized Cook Book
Artificial Hobby Curator
Intelligence Lifestyle
Content Creation
Customer Service
Personalized Advertising
Forecast Analysis
Marketing
Enabling Technologies

Cloud

Applications
Chatbot Development
Computing
Game Content Creation
Non Playable Character
Gaming

Solitude
Enhanced Customer Interaction
Personalized Content Creation Entertainment
Edge Computing
Sentiment Analysis
Financial Forecasting
Trading Strategies
Risk prediction and Management Finance

5G & Sustainability
Beyond Customer Service
Automated Assitance
Industry

Logistical Management
Intelligent Fleet Management and
Tracking
Real-Time Inventory Trackiing Travel & Transport

Human Computer
Interaction Proof Reading
Order Processing
Data Analysis
E-Commerce

Fig. 8. The impact of GPT models on various applications.

ing and effective. GPT can generate high-quality content potential bias in grading. It can also help teachers iden-
such as summaries, quizzes, and lesson plans based on tify areas where students may need additional support,
specific learning objectives, making learning accessible enabling them to tailor their teaching methods to better
to a wider audience. It can analyze written content and meet individual needs. GPT with its advanced natural
provide suggestions to improve grammar, punctuation, language processing skills, can help in automated as-
and readability. GPT can also assist in research writing by sessment by analyzing and grading student responses
suggesting ideas for structure, rephrasing and organizing to various types of assessment questions, including es-
content, and identifying gaps in research [78]. Moreover, says and short answer questions. It can also provide
GPT can also provide personalized feedback based on in- feedback to students [79], such as highlighting areas
dividual learning progress, enhancing the development of for improvement and suggesting further reading or re-
more comprehensive and informative learning materials. sources. GPT’s natural language processing capabilities
• Automated Assessments: Automated assessment in ed- can help to identify the meaning and context of students’
ucation uses technology to evaluate students’ learning responses, making automated assessment more accurate
outcomes, providing immediate feedback and reducing and effective. Additionally, GPT can generate personal-
 15

ized learning materials and exercises based on students’ tential for producing unreliable information. Further research
assessment results, supporting educators to create more is needed to explore human-computer interaction and user
tailored and effective learning experiences. The authors interface design to integrate GPT into educational workflows
in [80] have used Chat GPT in evaluating the students’ while ensuring that the information they provide is accurate
assignments such as quiz style questions, and also in and reliable.
generating relevant practice problems to improve content
retention and understanding. The results were promising
B. Healthcare
in the classroom. The authors believe that Chat GPT has
the significant ability in reducing the load of instructor 1) Introduction: Before technology became widespread
without compromising students’ learning outcomes. in healthcare, healthcare services were primarily delivered
• Fostering Creativity: Creativity thinking plays a vital role through face-to-face interactions between healthcare profes-
in education by encouraging students to think beyond sionals and patients. Traditional healthcare faced several chal-
traditional boundaries and develop innovative solutions to lenges, including limited medical instruments, paper-based
complex problems. It helps students to approach learning health records, patients receiving care mostly in hospitals
with an open mind and a willingness to explore new ideas, or clinics, physical travel requirements to receive medical
leading to greater engagement and motivation. GPT’s attention, and limited medical research. Despite these chal-
ability to generate human-like responses and creative lenges, traditional healthcare still provided valuable medi-
writing can aid in improving creativity. It can help cal services to patients. However, with the introduction of
improve creativity by generating new and innovative technology, healthcare has become more efficient, accessible,
ideas based on vast amounts of data and information. and personalized, resulting in improved patient outcomes and
By analyzing patterns in language and identifying con- better overall healthcare services. Technology has become
nections between different concepts, GPT can suggest an essential aspect of society, as reflected in the significant
novel approaches to teaching and learning. Additionally, investments made in this sector. Despite the advancements
GPT can also generate creative prompts or challenges for in technology, the healthcare industry still faces various new
students, encouraging them to think outside the box and challenges, including access to healthcare, high costs, per-
approach problems in unique ways [81]. GPT can also sonalized medicine, data privacy and security concerns, and
analyze and evaluate students’ creative work, providing an aging population. However, technology has the potential
feedback and suggestions for improvement. So, GPT can to address these challenges and improve the efficiency of
be a valuable tool for promoting and enhancing creativity healthcare services.
among students and faculty members. 2) Impact of GPT in healthcare: Recent years have seen
3) Challenges: There are several advantages to incorpo- significant advancements in technology, including in the
rating GPTs in education, but it is essential to acknowledge healthcare industry. Biotechnology, medical devices, and phar-
the potential limitations. While GPTs can quickly generate maceuticals have undergone transformations through the use
information, they may impede students’ critical thinking and of cutting-edge technologies like DL [82] and ML [83].
problem-solving skills. Furthermore, learners who benefit from Currently, the healthcare sector is utilizing various forms of
personal interaction with instructors may find the lack of AI techniques for medical research and providing medical
human involvement disadvantageous. GPTs rely on statistical services. One such technique is the GPT features of NLP,
patterns, so they cannot provide a comprehensive understand- which hold immense potential for the healthcare industry.
ing of the material being taught [79]. Privacy concerns arise GPT can help to overcome several challenges in healthcare
when using sensitive student data in GPTs for educational in multiple ways. For instance, it can be used to develop
purposes. Additionally, since GPTs cannot provide citations, it intelligent systems that assist doctors in making accurate
is challenging to identify the source of information generated. diagnoses and providing clinical assistance [84] [85]. GPT
The cost of maintaining GPT may be prohibitive for schools can also analyze large volumes of medical data and generate
and educational institutions with limited resources. Finally, reports. Furthermore, it has potential applications in drug
distinguishing between reliable and unreliable information discovery [86] [87], personalized medicine, patient diagnosis,
generated by GPTs can be difficult, so it is necessary to have medical image analysis, analyzing electronic health records,
human oversight to ensure data accuracy and regulate access. clinical decision support systems, and disease prediction.
4) Summary: GPT offers numerous advantages in the edu- • Drug Discovery: Recent AI and machine learning tech-
cation sector, including personalized and adaptive instruction, niques [88] [89] are having the potential to contribute
automated assessment, creative writing support, and research to the growth and development of drug discovery. GPTs
writing assistance. They have the potential to revolutionize are capable of learning new patterns and relationships
teaching by creating lesson plans and activities, responding [90] in the dataset they were trained on. This capability
to natural language queries, and integrating multiple digital can be used in drug discovery to aid in the identification
applications. However, there are also challenges to consider, and design of potential new drugs with desired properties
such as the potential negative impact on critical thinking and [91]. One of the key challenges in drug discovery is
problem-solving skills, lack of human interaction, data security finding compounds that can interact with specific parts
and privacy concerns, inability to provide full comprehension, of the body. GPT can help in this process by learning
lack of citations or sources, high cost of maintenance, and po- the patterns and relationships from large databases of
 16

known compounds [86]. GPT can be trained on large genomic, and nutritional data of patients. The dietician
sets of chemical databases to analyze chemical reactions and the author in [98] have observed that the utilization
and their outcomes. This can help suggest potential of Chat GPTs has significantly decreased obesity rates
combinations of new drugs using the analyzed data. These among patients by offering personalized recommenda-
new drugs can also be analyzed using GPT to test their tions regarding nutrition plans, exercise programs, and
efficacy and toxicity. psychological support. This approach allows for the de-
• Diagnosis: GPT can be used in medical diagnosis by ana- velopment of customized treatment plans that cater to the
lyzing patient data. It can help to analyze medical records specific needs of individuals, leading to a more efficient
and extract information such as patient demographics, method of treating obesity with the assistance of Chat
symptoms, and medical history. This can help medical GPT.
professionals provide effective patient care and improve 3) Challenges: While GPT is a powerful language model
outcomes. The recent release of GPT-4 has the ability to with numerous applications in healthcare, it is not without
support multimodal information, allowing it to analyze its challenges. The primary challenge is data bias. As GPT
images as input and produce text results as output [92]. models are also learning models, the significant drawback of
It is recommended to use AI systems such as a GPT, biasing is also applicable to GPT. GPT can be susceptible
as clinical support tools to assist medical professionals to bias. If the data used to train the model is biased, the
in diagnosing and treating patients, but they should not model will learn from it and replicate the bias. This leads to
be relied upon as the sole source of medical advice or incorrect treatment and predictions. Another challenge is the
decision-making. GPT can also be used to identify rare transparency of the model. GPT is complex to understand and
diseases by analyzing patient’s complete information. The interpret. This lack of transparency in technology can make
authors in [93] have used a general-purpose GPT based doctors and medical personnel not believe in the predictions,
on GPT-3 model for patient diagnosis and triage. The which may result in a hesitancy to trust and adopt technology
model has given a triage accuracy of 70% which was [99]. Another important concern is security and privacy issues.
worse than a physician. But, in next subsequent weeks, As it is a model to be trained on data, there is a huge
the accuracy has improved to 92% which is close to the amount of sensitive information about the patients to be used
performance of a physician. In diagnosis, GPT-3 model to improve the algorithm and its performance. This results in
has given 88% accuracy. For emergency cases, GPT-3 has significant security and privacy concerns related to the use of
given 75% accuracy whereas physician has given 94%. GPT in healthcare. The final and important challenge is limited
• Disease prediction: GPT has great potential in disease clinical validation. GPT are showing promising improvement
prediction [94]. By analyzing large amounts of medical in various fields of healthcare, such as drug discovery, and
data, including patient records, medical images, and clin- disease prediction. But still, their effectiveness and accuracy in
ical trials, these pre-trained language models can learn medical research and clinical settings have yet to be validated.
patterns and make predictions about the likelihood of More research and clinical trials are required to prove that GPT
a patient developing a particular disease. For instance, can transform the medical industry with full trust.
trained healthcare GPTs can be used to predict the 4) Summary: GPT have the potential to revolutionize the
occurrence of diseases such as diabetes, heart disease, healthcare industry by contributing to drug discovery, per-
and cancer by analyzing various parameters, including sonalized medicine, clinical support in making decisions,
the patient’s medical history, age, family history, and diagnosis support, and disease prediction. This can be helpful
lifestyle. It can also be used to predict the likelihood of for human beings to predict the disease in advance and treat
a rare disease This helps in the early detection of high- it through proper medicine. However, there are significant
risk patients so that medical personnel can take necessary challenges that are to be addressed, such as technology adop-
measures and suitable medicines to reduce the risk of tion, data bias, regulatory challenges, and security and privacy
developing the disease. The medical practitioner and issues. It is so important to analyze and evaluate the benefits
author in [95] have recommended using GPT-4 models’ and risks of using GPT in healthcare and to continue to
ability of NLP in bariatric surgery. monitor their development and implementation.
• Personalized medicine: The COVID-19 pandemic has
highlighted that not all body systems are clinically sim-
ilar. For instance, during the pandemic, medicines like C. Industry
Remdesivir and Tocilizumab have been effective for one 1) Introduction: An important economic transition from
category of patients but do not affect another category agriculture and handicrafts to large-scale industry and auto-
of patients with similar clinical metrics, as they progress mated production was achieved by the industrial revolution.
from a mild or moderate level of infection to a severe Efficiency and productivity were raised as a result of modern
stage [96]. This highlights the need for personalized equipment, energy sources, and labour arrangements. New
medicine in today’s world. GPT can be used to identify opportunities and challenges have been created as a result
variable patterns of data to predict or classify hidden of the quick development of new technologies in both the
or unseen patterns, which can be used for exploratory workplace and other industries [100]. The utilization of big
data analysis. GPT provide the possibility of identify- data is a well-known technology-driven trend. Nowadays,
ing personalized medicines [97] based on the clinical, companies have access to enormous volumes of data that
 17

may be examined to uncover insightful information. Big data ized the fashion business. The amount of personaliza-
can help businesses make wise decisions and discover areas tion has been further increased in this context by the
for development. AI is another innovation that is changing incorporation of GPT, dramatically altering the purchas-
industries. AI systems have the ability to analyse complex ing experience for customers [107]. Fashion platforms
data, automate procedures, and make wise conclusions [101]. may analyse a significant quantity of user data, such as
This improves production by increasing its dependability, browsing history, purchasing behaviour, and style prefer-
adaptability, and efficiency. The process of ”digitalization,” ences, using the advanced capabilities of GPT to produce
which includes incorporating digital technologies into every tailored recommendations. Fashion platforms can direct
element of business, is creating industries to become more consumers towards clothing options that fit their desire
flexible, efficient, and valuable. Businesses may automate for sustainable fashion by including eco-friendly fabric
tedious work, improve client experiences, and streamline oper- selections into the system. GPT improve users’ general
ations by implementing digital solutions. In today’s digitally- fashion knowledge and confidence while enabling users
driven world, adopting technological advancements is essential to keep up with the most recent trends. The image-text
for maintaining competitiveness and promoting growth. retrieval skills of GPT significantly improve visual search
2) Impact of GPT in Industry: In industrial scenarios, GPT capability in fashion platforms [108]. Users may make
has the potential to be applied as a sustainability tool, assisting more confident shopping decisions and minimize the need
businesses in evaluating and enhancing their sustainability for returns by visualizing how various clothing items
goals. Companies can improve supply chain tracking and and accessories would appear on them without physically
query response by integrating pre-trained transformer models trying them on. The model may recommend the proper
like ChatGPT with supply chain management platforms [102]. size for various brands and apparel products by taking
Additionally, GPTs can offer modifications to the production into account a user’s measurements, preferred fit styles,
process that might increase efficiency [103]. GPT can also and historical data. The overall purchasing experience is
help users make knowledgeable decisions about how to use enhanced and the frustration of wrong size is decreased.
resources, allowing businesses to remain competitive while • Sustainability: Sustainable development means address-
reducing their environmental effect. For example, the GPT- ing current demands without sacrificing the capacity of
2 model has demonstrated efficacy in sentiment analysis, future generations to address their own needs. Goals for
providing insightful data for numerous applications [104]. sustainable development can be attained by implementing
• Hospitality sector In the hospitality industry, hotels place GPTs in a variety of sectors, including manufacturing
a high focus on providing satisfying guest experiences. and corporate operations [109]. The models can estimate
To ensure that every tourist is satisfied during their where energy saving measures would be most useful
stay, this necessitates adapting to their requirements and by analyzing past data and patterns to provide insights
preferences. Hotels may improve the guest experience into energy usage, pinpoint problem areas, and recom-
in a number of ways by integrating GPT into their mend opportunities for improvement. GPTs can aid in
website or mobile application. Hotels may respond to identifying sustainability-related problems, creating plans
consumer inquiries in a timely and precise manner by and strategies to solve them, investigating brand-new
utilizing GPT [105]. Customers do not have to wait for sustainable activities, keeping track of advancements,
human assistance when looking up information about and conducting routine reviews. Companies can choose
facilities, booking procedures, or room availability. Cus- activities that will have the biggest positive impact by
tomers’ overall satisfaction with the hotel’s services is grading tasks and actions according to their impact on
increased as a result of the large reduction in client wait sustainability [110]. The models can optimize supply
times. GPTs can also make it easier for visitors who chains for decreased carbon emissions, minimized waste,
speak multiple languages to communicate [106]. Hotels and improved resource efficiency by assessing elements
can offer a more inclusive and welcoming experience including transportation routes, packaging materials, and
for visitors from other countries by removing linguistic supplier practises [111]. This results in more environ-
obstacles. Hotels may provide their visitors with immer- mentally friendly production, distribution, and sourcing
sive and engaging experiences by combining GPT with procedures.
AR technologies. For instance, customers can use their
mobile devices to get AR guided tours of the hotel or 3) Challenges: There are many different industrial fields
nearby attractions, offering a distinctive and entertaining where GPT models can be applied; the three areas mentioned
way to explore the surroundings and learn more about the above are only a few. However, for optimal use, the industrial
hotel’s amenities.GPTs integration into various aspects of sector needs to be ready to adapt to a constantly changing
the hospitality industry gives hotels the ability to deliver environment. Public and corporate policies must be developed
streamlined, tailored, and effective services, increasing over the long term to promote the use of sustainable production
client happiness and loyalty. techniques. For enterprises, deploying pre-trained GPT models
• Fashion: By providing highly customized user recommen- can be a costly task. Continuous development and training are
dations based on personal style, brand preferences, and also required to accommodate new and evolving inquiries as
particular clothing or accessory demands, collaborative client expectations change. Companies have to carefully assess
filtering and AI algorithms have undoubtedly revolution- the benefits and costs before implementing the GPT model
 18

because these continuing efforts raise the deployment cost crop yields. By examining historical yield data, weather
[112]. For industries to fully benefit from GPT models, it is patterns, soil conditions, and crop management practices,
crucial to address issues with interpretability, data reliance, and GPT identifies valuable patterns and correlations, pro-
ethical considerations. Industry may therefore take advantage viding insights and suggestions for optimal crop man-
of these GPT models’ advantages, make wise decisions, and agement techniques [114]. It enables precision farming
promote sustainable development. by integrating data from sensors, satellites, and IoT
4) Summary: GPTs have the ability to have a positive devices, granting timely guidance on resource allocation
impact on society and business operations. They can speed for improved efficiency. Additionally, GPT aids in the
up operations like accounting, sales, and marketing, increasing early identification and management of crop diseases and
productivity. But before they are widely used, ethical problems pests, minimizing yield losses through precise and prompt
need to be fully investigated. Technology products will change recommendations. Moreover, GPT supports crop breeding
as GPT models develop. To reap the benefits and reduce dan- and genetic optimization by analyzing genetic data and
gers, it is essential to solve interpretability and data concerns. plant characteristics, expediting the development of high-
GPTs can have a tremendous positive impact on businesses, yielding and resilient crop varieties. Therefore, GPTs data
society, and the economy when they are used responsibly. analysis and decision support capabilities significantly
contribute to enhancing crop yields and maximizing agri-
cultural productivity [115].
D. Agriculture
• Pest Control:
1) Introduction: Traditional agriculture, a time-honored GPT offers significant support in the realm of pest control
practice passed down through generations, sustains civiliza- in agriculture. By analyzing extensive data on pests,
tions with its crop cultivation and livestock rearing methods. including their behavior, life cycles, and characteristics,
Rooted in a deep connection to nature, it emphasizes sustain- GPT can provide valuable insights for effective control
ability and local ecosystem understanding. Beyond providing measures. It aids in early pest detection by analyzing
sustenance and livelihoods, traditional agriculture preserves sensor data and satellite imagery, enabling proactive
cultural heritage. However, it also faces challenges such as interventions to prevent pest spread and minimize damage
labor-intensive processes and shortages, inefficient resource [116]. GPT also assists in determining suitable pest
utilization, vulnerability to pests and diseases, and limited control methods tailored to specific crops and pests,
access to real-time data and environmental impact. Today, by considering factors like environmental impact and sus-
merging tradition with modernity, we have the opportunity to tainability. Additionally, it contributes to precision pest
leverage technological advancements to enhance productivity, control by leveraging real-time data to optimize timing
sustainability, and resilience while honoring the profound and dosage of interventions, reducing chemical usage
legacy of traditional agriculture for future generations. and resistance risks. It also aids in identifying natural
2) Impact of GPT in Agriculture : GPTs have the ability enemies and beneficial organisms, promoting natural pest
to overcome the challenges of agriculture. It offers valuable control mechanisms such as habitat diversification and
advantages to the agriculture sector. It acts as a comprehensive companion planting. Through GPT’s data analysis and
knowledge source, providing information on crop cultivation, recommendation capabilities, it empowers farmers with
pest management, and soil health. By analyzing real-time data, informed decisions, leading to more effective and sus-
GPT assists farmers in making informed decisions regarding tainable pest management strategies, ultimately reducing
optimal planting times and resource allocation. It plays a crop losses and enhancing agricultural productivity.
crucial role in identifying and addressing crop diseases and • Identifying Diseases and Soil analysis:
pests accurately. Moreover, GPT enables precision farming GPTs offer valuable assistance in disease identification
practices by utilizing sensor data and satellite imagery, en- and soil analysis within the field of agriculture. With
suring precise irrigation, fertilization, and pest control. Ad- its ability to analyze extensive data sets, GPT can ac-
ditionally, it provides market analysis and price prediction, curately identify crop diseases by processing information
empowering farmers to navigate market conditions and opti- such as symptoms, historical data, and disease patterns.
mize pricing strategies. GPT also supports farm management This enables timely and effective disease management
and planning, optimizing crop rotation and resource usage. strategies [114]. Additionally, It plays a significant role
By facilitating agricultural research and innovation, GPT con- in soil analysis by analyzing diverse soil-related data,
tributes to advancements in crop breeding and sustainable including nutrient levels, pH, organic matter content, and
practices. Embracing GPT in agriculture enhances decision- soil composition. By interpreting this data, It provides
making, efficiency, and sustainability, ultimately promoting insights into soil health and fertility, empowering farmers
improved productivity and food security. For instance, GPT-4 to make informed decisions regarding nutrient manage-
can educate farmers about new methods and goods and warn ment, soil amendments, and cultivation practices. More-
them of potential issues or possibilities by analyzing data from over, GPT can identify complex interactions between soil
many sources [113]. conditions and crop diseases, helping farmers understand
• Improving Crop Yields: the relationship and take preventive measures accord-
With its data analysis capabilities and real-time rec- ingly. It also supports precision agriculture practices by
ommendations, GPTs plays a crucial role in enhancing integrating sensor data and satellite imagery to assess
 19

soil variations across fields, allowing for site-specific Inquiries and problems can be handled quickly and efficiently
management strategies and optimized resource allocation. by chatbots and customer support systems powered by AI,
Furthermore, it also facilitates knowledge sharing and improving the entire customer experience [14].
collaboration by analyzing and disseminating research 2) Impact of GPT in Travel and Transport : Companies
findings, best practices, and disease outbreak information can learn about customer preferences in real time by us-
among agricultural communities. This collective intelli- ing GPTs in logistics and transportation, which results in
gence enhances disease monitoring and control efforts on better personalization and more customer satisfaction. GPTs
a broader scale. leverage NLP approaches to interpret customer requirements
3) Challenges: While GPT, provides significant benefits and preferences, enabling customized suggestions as well as
to agriculture, there are challenges to its implementation. guidance in the logistics and transportation processes. The
GPT’s effectiveness depends on the availability and quality most effective routes and forms of transportation can be
of data, making insufficient or biased data a limitation. The recommended using GPTs, which can analyse a large amount
interpretability of GPT’s decision-making process is challeng- of data, including traffic patterns, weather conditions, and
ing due to its black-box nature, hindering trust and under- delivery requirements [118]. In addition, GPTs can be used as
standing. GPT’s computational requirements and infrastructure travel planners, allowing visitors to enter their travel budget,
can be demanding, posing difficulties for resource-constrained duration, and destination to create customized itineraries. For
farmers. Language and domain-specific nuances can affect travel agencies, this personalized approach increases consumer
its performance, impacting accuracy and relevance. Ethical satisfaction and revenue.
considerations surrounding data privacy and ownership need • Logistical Management: GPTs can be quite important

careful attention to ensure responsible use. By addressing in the context of shipping logistics. They can automate
these challenges, researchers and practitioners can unlock the creation of shipping labels, eliminating up manual
GPT’s potential while ensuring its practicality and ethical entry and lowering the possibility of mistakes. Addi-
implementation in agriculture. tionally, GPTs can have access to real-time tracking
4) Summary: GPT holds immense potential in agriculture, data and can integrate GPS data and sensors to provide
offering numerous benefits alongside notable challenges. Its businesses and customers with precise and up-to-date
data analysis capabilities empower farmers with informed shipment status information. Companies can successfully
decision-making in disease identification, soil analysis, and monitor shipments with the use of GPTs, geographic
precision farming, leading to improved crop yields and sus- information systems (GIS), and routing algorithms [14].
tainable practices. However, the effectiveness of GPT relies Organizations can track shipments in real-time and ensure
on data availability and quality, while its interpretability re- visibility throughout the supply chain by utilizing GPS
mains a challenge due to its black-box nature. Additionally, data and sensor technology [119]. Customers can receive
computational requirements, language nuances, and ethical precise updates on their shipments using this real-time
considerations require careful attention. By addressing these information, which will improve their experience overall.
challenges, the agricultural sector can harness the full potential Overall, the use of GPTs into shipping logistics results in
of GPT, paving the way for more productive, efficient, and increased automation, efficiency, and client satisfaction.
responsible farming practices. • Intelligent Fleet Management and Tracking: Companies
can get real-time fleet updates by utilizing GPT models,
which enables them to track vehicles quickly and pre-
E. Travel and Transport cisely. GPT models’ underlying technology also supports
1) Introduction: Historically, animals have been used by proactive fleet management. GPTs can identify possible
people as their main source of transportation. But as the problems or maintenance needs before they develop into
world’s population increased, the demand for more effec- expensive breakdowns or accidents by analyzing data
tive transportation systems increased. Transportation-related from a variety of sources [120]. With this knowledge,
technological advancements have fundamentally changed the companies may take preventative measures, such as plan-
sector in several ways. Business operations like order tracking, ning maintenance or quickly fixing developing problems,
freight management, and customer support can be streamlined ultimately saving time and money by preventing unin-
by automation employing AI-driven technologies. Companies tended delays. Additionally, GPTs can provide clever
can enable their employees to concentrate on more benefi- alerts and notifications. Businesses can receive alerts
cial and profitable duties by automating these tasks [117]. when vehicles arrive at specified areas by setting up
With better transportation networks and logistics management specific triggers, which enables better coordination and
systems that optimize routes and reduce transit times, tech- customer service [121]. For instance, businesses can alert
nological developments also enable speedier delivery times. clients or storage facilities in advance of a truck’s arrival,
In terms of product development, technical advancement has allowing for effective unloading and loading procedures.
paved the way for the development of innovative vehicles, • Real-Time Inventory Tracking: GPTs enable businesses
infrastructure, and logistics systems, leading to the production to manage their inventory levels while on the road
of more sophisticated and effective transportation choices. with a cloud-based platform that makes it simple to
Another noteworthy benefit of technology advancement in access inventory data from anywhere in the world. Better
logistics and transportation is increased customer service. inventory management and decision-making are made
 20

possible by this real-time accessibility. This ensures that truthful, unbiased, and free from harmful presumptions or
the appropriate quantity of stock is accessible when discriminatory procedures [126]. Although the models contain
needed to fulfil consumer requests, while minimizing advanced features, they are difficult to tailor for specific use
carrying costs and preventing lost sales as a result of cases, need a lot of data to train, and have built-in limitations.
stockouts. GPTs can streamline inventory management 4) Summary: Emerging GPTs have the potential to enhance
procedures by eliminating the need for human data entry productivity, communication, and the calibre of goods and
into spreadsheets, saving time and cutting overhead costs services, which will benefit many aspects of people’s life.
[117]. With the advent of 5G technology, the cost of GPTs can offer real-time updates, effective route optimization,
connected devices has dramatically decreased, making it and customized recommendations in the travel and transporta-
more practical and affordable for businesses to set up and tion industries, enhancing the overall travel experience and
operate connected inventory monitoring systems. This increasing operational effectiveness. Adopting them, however,
may make real-time inventory tracking solutions more comes with some difficulties. As specific roles are replaced
widely adopted, thereby increasing the effectiveness and by automation, GPTs may result in job displacement [127].
precision of inventory management [122]. Additionally, the computational and memory requirements
• Streamlining Delivery Operations: GPTs are able to es- for GPTs make their deployment on compact or low-power
timate traffic trends and improve routes for both drivers devices difficult. GPTs may not be accessible to growing
and passengers using real-time data [14]. These models businesses due to the high costs associated with obtaining and
can produce effective routes that reduce travel times and using them. Despite these obstacles, attempts are being done to
enhance overall delivery performance by taking into ac- overcome them and improve the usability and value of GPTs
count aspects like traffic congestion, road conditions, and for a larger range of users.
delivery schedules. Route optimization not only reduces
travel time but also benefits the environment. In order to
F. E-Commerce
improve air quality and create a more sustainable delivery
process, it is possible to cut down on idle times and trip 1) Introduction: Electronic commerce, commonly referred
distances. Businesses may streamline operations, improve to as e-commerce, is a way for conducting economic trans-
the overall customer experience, and contribute to a actions and create relationships between groups of people
more sustainable and environmentally friendly approach and entities using digital information processes and electronic
to logistics by automating procedures, optimizing routes, communications [128]. Globally, this type of trade has experi-
and utilizing real-time data [123]. enced substantial growth, particularly in the retail sector. The
• Tourism: GPTs have the potential to significantly improve preference for internet shopping, especially among younger
a number of tourism-related aspects. GPTs can offer millennials, is a prominent trend in consumer behaviour.
customized solutions that suit the individual’s preferences Mobile devices have consequently taken over as the main
by understanding their needs and interests, resulting in a method for carrying out internet transactions [129]. Therefore,
more pleasurable travel experience. GPTs are excellent it is crucial for e-commerce companies to give the customer
at understanding and creating text that is human-like [4]. experience in their mobile applications top priority. The pro-
This functionality can be used in the travel and tourism vision of brief text summaries for titles and reviews is an
sector to enable chatbots or virtual travel assistants to essential component of this. These summaries are essential
communicate with users in natural language [124]. Trip for optimizing search results, helping consumers identify ap-
planning and information retrieval are made more sim- propriate items, and ultimately raising customer happiness in
ple and user-friendly by the ability of travelers to ask the online purchasing space [130].
questions, look for advice, and obtain full details about 2) Impact of GPT in E-Commerce Realms: The e-
destinations, modes of transportation, customs, and more. commerce sector could significantly advance with the intro-
GPTs are capable of producing in-depth and interesting duction of GPTs. GPTs can be accessed by users or customers
descriptions of tourist sites, attractions, lodging, restau- and are intended to answer commonly asked questions and
rants etc. GPTs can provide time-efficient routes that give in-depth details about many elements of the e-commerce
guarantee a complete travel experience [125]. Including process, such as products, delivery, refunds, and more [107].
advice on local legislation, emergency contacts, medical One of the main benefits of GPTs is their capacity for quick
facilities, and potential risks, GPTs can offer helpful responses, which decreases the amount of time customers must
information and direction regarding travel safety. wait to hear back from businesses [131]. By taking care of an
important number of client inquiries, this function not only
3) Challenges: Privacy issues may occur when using sen- increases customer happiness but also lessens the workload
sitive data in travel GPTs. It is essential to manage user data on support workers. Customers will ultimately have a better
sensibly and putting up strong security measures to safeguard purchasing experience as a result of being able to quickly
private data. The quality of the model’s outputs is directly acquire the information they require and interact with GPTs
influenced by the correctness and completeness of the data uti- [106].
lized during the training phase. Ethical considerations should • Proofreading: To improve the calibre and accuracy of
be taken into account when creating AI-powered applications written content in e-commerce, GPTs can be used for
employing GPTs. It’s crucial to check that the models are proofreading. Written content is essential for product
 21

descriptions, marketing materials, customer reviews, and desired outcomes. GPTs can help with email writing
other text-based components in the e-commerce sector that encourages readers to become partners, investors,
[105]. For the purpose of projecting professionalism, or customers [138]. To increase the likelihood of a
fostering trust, and delivering a satisfying user experi- favourable response or interaction, these emails can be
ence, this text must be devoid of errors, well-written, customized to address the needs and potential benefits
and grammatically correct. E-commerce companies can for the receivers. To improve their comprehension and
automate the process of identifying and correcting these production of appropriate material, GPTs can be trained
problems by using GPTs for proofreading, which saves on domain-specific knowledge bases, such as e-commerce
time and effort as comparison to manual proofreading [107]. The models can offer more precise and situation-
[132]. This can be especially helpful in situations when specific recommendations for advertising strategies, prod-
users are writing product reviews or interacting with uct positioning, and target audience interaction because
customer service. An improved user experience is facil- of this specialized training.
itated by the early detection and rectification of errors, • Data analysis: There are numerous ways to use GPTs for
which also helps to avoid potential misunderstandings or data analysis in e-commerce. E-commerce data prepa-
miscommunications. ration can be aided by GPTs [24]. Data normalization,
• Order Processing: GPTs are useful in many areas of cleansing, and formatting are a few of the duties in-
order management and customer service because they volved. GPTs can produce summaries, identify significant
can comprehend and produce text that looks like human topics, and extract appropriate data by studying textual
speech. GPTs can help with handling consumer questions descriptions, reviews, and consumer feedback. This helps
about orders [133]. GPT is capable of interpreting the you know the data more thoroughly, identify trends, and
queries, providing important details like order status, find insightful information. Customer reviews and social
tracking information, and expected delivery time, as well media comments from e-commerce can be analyzed for
as suggesting corrections for frequent problems [134]. sentiment using GPTs. The sentiment expressed in text
By delivering real-time information, GPTs can assist cus- can be evaluated using GPTs and categorized as either
tomers in tracking their orders. Customers can customize positive, negative, or neutral [107]. Understanding client
their purchase with the help of GPTs. GPTs can help in views, recognizing product strengths and deficiencies,
the identification of possibly fraudulent orders by exam- and making data-driven decisions all benefit from this
ining past transaction data, consumer behaviour patterns analysis, which also helps to increase customer happiness.
etc [135]. Based on a customer’s past purchases, browsing Segmenting consumers based on preferences, behaviours,
habits, and preferences, GPTs can offer tailored product or past purchases can be aided by GPT models [139].
recommendations. When a consumer puts a purchase, For the purpose of detecting fraud in e-commerce trans-
the model can examine the information and produce actions, GPTs can be used. GPTs can support the identifi-
recommendations for related or supplementary products cation of potentially fraudulent actions by examining past
that the customer might find interesting. transaction data, user behaviour patterns, and recognized
• Generating titles for products: Companies can use GPTs fraudulent tendencies [140].
to produce interesting and educational material to im- 3) Challenges: While GPTs have a lot of potential for
prove the appeal of their product listings [117]. Based on numerous e-commerce applications, they also have several
a product’s category, brand, and special characteristics, drawbacks. In order to produce responses, GPTs mostly rely
GPTs can come up with attractive titles for it. The model on the context given in the input text. They could, however,
can produce imaginative and memorable names that aid in find it difficult to fully understand the broader context or
brand awareness and differentiation by receiving relevant details that are unique to e-commerce. GPTs provide replies
information such as the characteristics of the product using training data and prior knowledge. They are unable to
and the target market. GPTs are trained to produce in- access real-time data or carry out real-time calculations [141].
depth and interesting product descriptions [113]. These They might not be appropriate for giving current information,
summaries can offer a thorough summary that aids clients such as pricing, product availability, or dynamic promotional
in selecting products wisely. GPTs are capable of coming offers. GPTs gain their knowledge from a wealth of training
up with clever and appealing captions for product images. data, which includes text taken from the internet, which
GPTs can be adjusted to better reflect the tone and may be biased, stereotyped, or otherwise offensive [142].
aesthetic of a certain brand [136]. As a result, the brand The models may unintentionally provide biased or unsuitable
identity is consistent and unified throughout all product replies if they are not rigorously managed and monitored,
listings. which could be harmful to the customer experience and brand
• Strategy Planning: GPTs have the ability to come up reputation. The use of ethical principles and the training data
with original and distinctive concepts for marketing must both be given careful thought.
campaigns [137]. The model can provide recommenda- 4) Summary: The conversational interface offered by GPTs
tions for different campaign aspects, such as slogans, customizes the purchasing process and makes interactions with
taglines, themes, contests, social media strategies, and clients more interesting and appropriate to their individual
more by taking into account relevant information about requirements. GPTs can also be utilized to get insightful
the product, target audience, marketing objectives, and customer feedback. Businesses can learn about customers’
 22

preferences, issues, and opinions regarding their products and • Solitude with GPT: As the GPT itself is an entertainer,
services by conversing with them. In order to better serve one can feel better alone with the GPT, which helps to
their target audience, organizations can use this information to come out of loneliness by exploring its savors [147].
discover areas for improvement, increase customer happiness, GPTs assist in providing soothing poems, mental healing
and make data-driven decisions. It’s essential to recognize quotes, and funny riddles. People with loneliness may feel
that GPTs might occasionally make mistakes or give poor anxiety, especially with older ones at home. In this case,
answers, particularly when dealing with complicated or am- GPT-4 helps people with its Voice Technology feature,
biguous queries. This highlights the necessity of continual enabling users to input their audio [147]. In turn, the
model training, thorough testing, and modification to guarantee GPT model responds to user-specific speech output using
that they consistently meet consumer needs. To confirm the NLP algorithms embedded with it. The elderly can feel
efficacy and dependability of using GPTs specifically in the e- safe and attentive at home. GPT-4 is multilingual and can
commerce area, more research and testing are required. When understand various dialects and accents for personalized
implementing GPTs, it’s critical for businesses to take into user experience.
account the particulars of their own e-commerce businesses, • Enhanced Customer Interaction: The advent of Chat-
their target market, and the type of the client enquiries. Regular GPT and Bard has improved customer interactions on
monitoring and feedback analysis, along with a systematic content such as movies, Over-the-Top (OTT) platforms
and iterative approach, can help make sure that the outcomes like Netflix, Hulu, Disney+ Hotstar, and prime video,
of using GPT models are in line with the objectives of e- sound recordings, song lyrics, pictorial works, comics,
commerce enterprises. jokes, memes, viral videos, and other entertaining factors.
Further, GPTs provide human-like recommendations on
user-specific fun activities based on user interactions for
G. Entertainment
an immersive experience. This has dramatically improved
1) Introduction: In the ancient days, the Entertainment the interactions in the engagement industries. User en-
meant about playing games with neighbors covers all outdoor gagement can be further improved by providing dynamic
games, indoor games and chatting with neighbours through and more realistic responses to user queries, such as
telephone. As digitization has bought greater advancements creating virtual actors for interacting with real actors
in computation and communication, in turn access to internet [148].
is also much easier. This has changed the way people are • Personalized Content Creation: GPTs can help generate
entertained. as people are connected and fully engaged in user-specific personalized content by analyzing the user
completing the target for the day. And there was a radical shift preferences and generating content like predicting future
from traditional employment to employment in the Industrial scenarios tailored to the user’s interest. GPTs can be
Revolution age. Stress and pressure are common factors hin- used for creating personalized, engaging, and high-quality
dering people of different age groups. The different forms of content for online business advertising, ideas for content
entertainment serve as stress busters. Entertainment and mental generation, marketing messages for attracting customers,
health are interrelated; the former transfers happiness, bringing descriptions for selling products, and captions for social
harmony and peace to mental health. Some common forms media [149]. In addition, it can be used for optimizing
of entertainment include playing games, watching TV series the contents for search engines, i.e., GPTs will provide
or movies, or funny videos, shopping, debugging, coding, relevant terms for search, thereby avoiding traffic to the
browsing the internet, listening to music, dancing, chatting, web sources.
painting, crafting, reading books, cooking, and many more, • For the Film and TV industry: GPT-powered virtual
which can lessen the stress carried [143]. Entertaining and assistants assist users in booking tickets and generat-
getting entertained is the biggest motivation and medicine for ing content and personalized recommendations using AI
all mental illnesses. Entertainment helps to improve the motor models. The evolution of GPT-4 with advanced NLP
skills of humans, thereby inducing a positive cognitive effect and DL algorithms helps the scriptwriter to generate AI-
towards the work. driven content without the human author named virtual
2) Impact of GPT on the Entertainment Industry: GPT is storytelling [150]. GPTs create interactive stories, dia-
a potential game-changer in the entertainment field, delivering logues, and characters, recommending suitable characters.
endless entertainment. Since its evolution, GPT models have Furthermore, GPTs can be used to create content for
been adopted as an entertainer crosschecking their ability video games, voice-enabled applications, AR applica-
to produce content on funny and illogical questions. GPTs tions, and other VR experiences in virtual worlds [151].
entertain people in many ways, and of course, using GPT itself • For Social media influencers: GPTs can generate person-
an entertainment as it reduces the burden of overthinking by alized marketing ads for each customer based on their
providing immediate feedback to queries in seconds [144]. previous interactions and provides relevant suggestions
The results are amazing and have been utilized for many for customer viewing experiences. Youtubers and other
purposes today. When the GPT model was probed to complete social media content creators will potentially benefit from
a scene from the movie “Frozen,” it responded with an generating channel content based on demand and realistic
entertaining writeup [145], [146]. Some of the impacts of societal activities.
GPTs on Entertainment applications are given below:
 23

• Realistic Gaming Interactions: GPT helps to generate the H. Lifestyle

players, gaming narratives, dialogues, user interface, and
1) Introduction: Lifestyle, the way of our living, is one
user-specific gaming recommendations and new game
of the prominent areas most people in today’s digital era of
creation. Powerful HCIs can render a better user ex-
AI, are bound to and look for constant improvement. The
perience for game developers and players. Assistance
“modus vivendi” is a Latin expression that semantically means
to the game developers in debugging and enhancing
a way of living and should be understood in terms of values
the code developed. GPT uses various NLP and AI
and attitudes. These two terms manifest self, influenced by
algorithms trained with massive data to predict the next
family, society, and global media. Directly or indirectly, these
phrase/movements and provide human-like experiences in
influence an individual’s lifestyle. Adopted from a sociological
3D gaming environments. ChatGPT has been integrated
perspective, an individual expresses oneself through different
with AR and VR to provide an immersive gaming expe-
practices, viz., eating ways, drinking behaviours, mode of
rience.
travel, travelling places, costume designs, body-shaping cloth
3) Challenges: Latency is the major issue connected with to wear, media preferences, education choices, job preferences,
rendering the voice-based response to the voice input. As entertainment modes, managing leisure time, means of com-
well, plausible misinterpretations may mislead the responses, munication and so [152]. In all these practices, individuals
and interruptions to the relayed output are difficult. Enabling would like to explore and learn about what, where, how, and
technologies like EC and 5G can help overcome this issue. when factors for reading sustainable development [153]. The
Also, GPTs must be capable of storing the facts with audio concept of lifestyle is all about “how one wants to live one’s
conversations to relay them while conversing the other day. life.” Consumerism is the act of purchasing artifacts for soci-
Furthermore, the AI system must be built in such a way etal status and is one of the thriving lifestyle factors. Certain
that it can continuously learn (lifelong machine learning) and standard indicators like job, wealth, and physical and mental
enhance over time. The major ethical concern with virtual health determine the quality of one’s life. Also, the choice
storytelling is the bias exposed in the training data and the of a healthy lifestyle moderately determines the health of an
obscurity of reproduced content on the generated stories. individual [154]. Furthermore, few people believe that lifestyle
Another issue with the generated content is plagiarism (i.e., reflects their socioeconomic status. Many epidemiologic stud-
producing content similar to the content in the published ies state that better lifestyles have dramatically reduced the
articles or books), raising disputes with intellectual property risk of various chronic diseases and are the primary cause
rights. In addition to this, the source of the content generated for their prevention [155]. The lifestyle has been defined
remains unexplored. The language barriers in using GPT must on different societal levels from individual, positional, and
be lessened to improve user experience and utilize the features national to global [153]. At the global level, lifestyle is adopted
of GPT [113]. The implication of the user to provide inputs in by general world-class influencers. In contrast, at the national
a certain format to GPTs can be further improved by providing level, the influencing factors will be the government and dif-
different options in addition to voice-based inputs GPT4, ferent cultural patterns across the country. The positional level
like braille screen input for visually disabled people. The concerns influence from different status groups, age categories,
user authentication can also be further enhanced to safeguard gender groups, and social classes. And the individual level is
user-specific content generation and avoid repeated content influenced by a closely moving group of individuals concerned
generation for users with similar requests. One of the primary about self-identity. The major source of information about
concerns with the GPTs adoption is job loss. Content creators, these influencers is the Internet through social media networks
bloggers, and poets may lose their jobs. and personal development advertisements.
4) Summary: The entertainment industry is the one which 2) Impact of GPT in Lifestyle: The most remarkable ap-
will be in demand always, as it is a lifeline for many indi- plication of AI, the GPT, paves the way for the betterment
viduals leading a stressful work environment or personal life. of mankind in offering human-like intelligent conversation
Despite the stress, entertainment has become part of routine on all whereabouts. People will always prefer to interact
life due to its immersive nature, creating harmony in the mind with other peers to learn their attributes and tweak them for
and the environment. GPTs have made a major contribution to societal status. Various GPTs have flourished for different
enhancing the entertainment industry, but the job security of lifestyle indicators, and they provide human-like assistance
many professionals in this field remains unanswered. GPTs to all queries on fine-tuning the lifestyle by harnessing the
must be trained on unbiased data and ensure transparency power of AI [156]. The advanced reasoning capability of GPT-
in source content generation to provide a secure, robust, and 4 serves the purpose better [147].
efficient contribution to the entertainment industry. To attract • Diet Planner: Free GPT applications for maintaining
all types of users, the multilingual capability and content a balanced diet, helping the individual with a weight
rendering of GPTs can be further enhanced. The issues con- loss diet plan, followed by a brief list of meal plans,
strained by providing user inputs to GPTs can be alleviated to required shopping lists, physical activity plans targeting
all extent. Furthermore, safer user content generation without particular body parts [157], motivational messages, and
plagiarism and relating facts with previous conversations can personalized sleeping patterns. These apps act more like
be guaranteed by abiding by the storage requirements to deal personalized training assistance and help to track progress
with a more personalized user experience. with visualization charts or graphs. Fitness level, available
 24

free hours, medications taken, and available exercise sometimes require users to update information in a specific
equipment will be given as input to GPT. format and may have outdated databases. GPTs cannot access
• Travel Guide and Trip Advisor: Harnessing AI models, specific job openings’ websites but can still provide insights
GPT provides an individual’s travel plan itinerary based into acquiring them. At times, it can produce nonsensical infor-
on information like the place(multiple cities), budget, and mation [162]. Therefore, before adopting the GPT recommen-
the number of days. These GPTs provide local recom- dations fully, further instigation is recommended. Furthermore,
mendations on restaurants, hotels, and other attractions. developing a large multimodal learning model abiding huge
RoamAround, Roamr, and VacayChatbot are some of the and dynamic datasets will be costly.
travel planning GPTs [158]. 4) Summary: GPT is a personalized assistant for improving
• Personalized Stylist and Beauty Advisor: GPTs can an individual’s lifestyle from various prospective influencers.
act as personalized stylists for an individual by gen- Generating personalized recommendations alleviates an indi-
erating occasion-specific clothing and costume prefer- vidual’s fear of survival in the digitized society. Individuals
ences. GPTs can assist in organizing wardrobes based will be personally trained to adapt to different cultural and
on seasonal outfits and provide recommendations on e- technological shifts in the sustainable development of them-
commerce fashion stores for purchasing favourite brands. selves and the economy as a whole. On the other hand, more
GPTs can provide tailoring design options, fabric choices, stringent recommendations may incur huge budget overruns
and design materials. Furthermore, GPTs can provide and sometimes provoke the individual to misinterpret, leading
updates on a stock based on the preferred searches and to dreadful consequences. GPTs provide both positive and
provide insight into fabric types suitable for weather negative recommendations based on the input fed. So, for the
conditions that suit personal style. effective adoption of a GPT for lifestyle practices, adverse
• Personalized CookBook: GPTs can serve as cooking training and testing on extreme behaviours must be carried out.
assistants by recommending new curated recipes suiting GPTs must be trained in the realistic and dynamic perception
the family dietary plan, ingredients available, time, in- of individuals in real life.
dividual’s cooking skills, and new flavoured ingredients.
ChefGPT, PantryChef, and MacrosChef are some GPTs
that generate unique and delightful recipes [159]. Con- I. Gaming
sequently, GPTs can assist in shopping list recommenda- 1) Introduction: Before the advent of technology and the
tions and the nutritional value of the recipe generated. gaming industry, entertainment was primarily centred around
• Hobby Curator : GPT assists an individual in identifying activities such as reading, listening to music, watching plays
one’s enjoyable leisure time activity by learning new and movies, participating in sports and physical games, and
skills [160]. Having a list of interests and ideas ready, socializing with friends and family. People also engaged in
the GPT helps narrow down various options, instructional traditional board games and card games, which were often
videos to proceed, chatting and sharing with online played in groups and provided a fun and social way to pass
communities, and researching the cited hobby to explore the time. After technology stepped into the gaming industry,
more fun. Budget will also be an important factor in the way games are created, and the experience it has given
this perspective, as learning new hobbies may require users have transformed tremendously. Technology has enabled
joining paid classes or courses. GPT provides step-by- developers to create more immersive and engaging experiences
step instructions and guidelines to learn a new skill faster. for players. It has contributed in various ways, like improving
• Dream Maker: GPTs with multimodal learning helps graphics, performance, online play, and mobile gaming. Im-
to search for a job based on one’s qualifications and proved GPUs and other technologies allow for more detailed
experience. In turn, it assists in preparing the job-specific and realistic graphics, making games more visually stunning.
resume, cover letters, training for the interviews (coding Faster processors and higher amounts of RAM allow for
and technical queries), and grooming sessions and can smoother gameplay and faster loading times, reducing lag and
redirect to the training place where knowledge can be improving overall performance. Technologies like AI, AR, and
acquired [161]. The futurist GPT models can assist in VR have created a new dimension of game development and
phase-by-phase questionnaires in the interview process experience. Players can now immerse themselves in gaming
worlds in a way that was not possible before. With the help
3) Challenges: The recent version of the GPT uses both of advanced AI techniques, game developers can create more
reinforcement and supervised learning models so it can learn sophisticated and challenging opponents for players, as well as
based on the interaction with the user and can use existing data NPCs with more realistic behaviours. Technology has greatly
to derive personalized decisions. In the context of lifestyle, expanded the possibilities of gaming and enabled developers
GPTs offer the most promising solution for almost all lifestyle to create more immersive, visually stunning, and engaging
influencers, but the still challenging part is the trustworthiness experiences for players.
of the data and copyright issues. Also, relying more on GPT 2) Impact of GPT in Gaming : GPTs have the ability to
as it solves all our problems may insipid human intelligence contribute to all sectors, including the gaming sector. GPT
in upcoming generations. Though the GPT provides weight are not specifically designed for creating and playing games,
advice, it can never be a substitute for the medical practitioner, but they have the potential to improve the gaming experience
as some information can be misleading. Travel planning GPTs by improving enhanced dialogue and story telling, creating
 25

dynamic and personalized gaming worlds, generating more designed for creating NPCs, but they can be used to
realistic and engaging characters [163], game content creation, generate dialogue and other character interactions that
chatbot development. can be incorporated into NPCs. Additionally, It can be
used to generate character backstories and personalities,
• Chatbot development: GPTs have been used in gaming which can inform the development of NPCs. The authors
through the development of chatbots that use NLP to in [166] have trained and used GPT-2 for text generation
communicate with players [164]. Because it allows the of video games. They have trained GPT-2 on a large
chatbot to understand and respond to a wide range of user corpus of video game quests and used a GPT model
inputs and queries related to the game. GPTs have been to generate the dialogue for quest-giver NPCs in role-
pre-trained on a large corpus of text data, which makes playing games. The output has shown that GPT can learn
them adept at NLP. It can understand and respond to user the structure and linguistic style of the games, and the
queries in a way that feels natural and intuitive. It can also quality of the content it has generated is high, making it
understand the context of a user’s query, which means a good alternative to writing new RPG quests by hand.
they can provide relevant and useful responses even when 3) Challenges: GPTs are computationally expensive and
the user’s query is ambiguous or incomplete. It can require high computing resources to do their purpose. This
also generate game-related content, such as descriptions means that implementing them in a game would require
of game characters or settings, that can help to enrich powerful hardware and this could have an impact on the
the user’s gaming experience. Furthermore, it can also performance of the games. Lack of training data: GPTs require
personalize the user’s experience by learning from their large amounts of high-quality training data to be effective.
previous interactions with the chatbot and tailoring its In the gaming industry, this could be difficult to obtain, as
responses accordingly. gaming data are likely to be fragmented and less structured
• Game content creation: GPTs are used in game design. than the kind of data used to train GPT models [131]. In
They are used to create game content such as levels, addition, GPTs can perform content creation based on patterns
items, and quests. If the game designer is working on a they have learned from their training data, which means that
new role-playing game, GPT can be used in creating char- they can be unpredictable. The content generated by GPT
acters to be used in the games. To generate new character may be nonsensical or inappropriate content to the game.
classes in the games, the developer has to give inputs that In the context of gaming, this lack of control could lead to
contain information about the game environment, game undesirable or even offensive game content. GPTs can generate
settings, player abilities, and game play mechanics. GPTs text based on user input, they can’t interact with the game
has the ability to analyze the text and expectations given environment in the same way a human player can. This limits
by the developer, and it can generate a list of potential their usefulness in gaming and may make them less effective
character classes based on the expectations given as text. than other AI technologies.
The designer then refines the ideas and chooses a more 4) Summary: GPTs can transform the gaming industry by
suitable character to develop further with unique abilities contributing to improved game dialogue creation, enhanced
and game mechanics. The authors in [165] have used non-player characters, personalized gameplay, procedural con-
GPT2 and GPT3 to procedurally generate role-playing tent generation, chatbot generation, and analyzing players’
game with video game descriptions. The resultant quest abilities. However, it also has potential challenges that are to
was evaluated by 349 online RPG players. The results be addressed, such as the need for high computing resources, a
concluded that one of the five quest descriptions was lack of control over content creation, and restricted interaction
accepted for game development. with the game environment. In addition, the most important
• Analyze player’s ability and skill: GPTs can detect challenge in adopting a GPT model in gaming is a lack of
and analyze players’ abilities and skill levels and tailor training data. If the challenges are addressed and the gaming
the game accordingly. This analysis helps in making industry evolves with properly structured data to train a GPT
dynamic modifications to the game environment based model, then GPTs can revolutionize the field of gaming.
on the player’s abilities and skill levels. This feature
helps achieve dynamic difficulty balancing. GPTs can J. Marketing
also assist in identifying the player’s intent. Thus, when 1) Introduction: Traditional marketing primarily relied on
players ascend to higher levels, it can assist in making traditional media channels, such as television, radio, newspa-
the games more challenging based on the player’s abilities pers, and magazines, to reach consumers. Companies used to
and skill levels in the previous levels develop marketing campaigns based on demographic data, and
• NPCs: NPC stands for ”Non-Player Character.” In AI mass media channels were used to broadcast these campaigns
games, NPCs refer to characters or entities in a game to a broad audience. However, the advancements in technology
that are not controlled by a player. NPCs can take on have brought about significant changes in the marketing indus-
a variety of roles within a game, such as enemies to try, and companies are increasingly integrating new marketing
fight, quest givers, merchants, or friendly characters that strategies evolved through technologies to reach and engage
provide helpful information. They are often controlled with customers. One of the significant transformations has
by AI algorithms that determine their behaviour and been the rise of digital marketing channels such as social me-
actions within the game world. GPTs are not specifically dia, search engines, Email, and mobile applications that allow
 26

companies to target specific populations with precision and the context of the mail. The authors have concluded that
provide real-time feedback on campaign performance, allow- applying GPT-3 to rationalize email communication is
ing for more effective and efficient marketing. Technology has feasible both technically and economically.
also given rise to marketing automation tools such as customer • Personalized advertising: GPTs can generate personalized
relationship management systems, chat-bots, and personalized content such as product descriptions, blog posts, and
email marketing, which have made marketing more efficient social media captions tailored to individual customers’
and effective. Another significant transformation has been the preferences and interests. This can help businesses create
use of big data and analytics to better understand customer content that resonates with their target audience, leading
behaviour and preferences. This has allowed companies to to higher engagement and conversion rates. By analyzing
create more personalized and targeted campaigns based on customer data, GPTs can segment customers according
specific customer needs and preferences. to their behaviour, interests, and preferences. As a result,
2) Impact of GPT in Marketing : The marketing industry businesses can tailor their marketing campaigns to each
has evolved with various AI-powered techniques. This revo- segment and provide personalized messaging and offers
lution started in marketing by providing businesses with pow- that are more likely to connect with each customer group.
erful tools for generating insights, automating processes, and The authors in [169] have proposed a generative model
improving customer experiences. GPTs are also being used to identify the name of the product from the product text
in marketing to generate engaging and personalized content. and use this information filter to improve the product
Some of the applications of GPT in marketing include content recommendation based on the product retrieval model.
creation, customer service, and personalized advertising. This method has been implemented in the dynamic prod-
• Content creation: GPTs can contribute to marketing in uct advertising system of Yahoo. It is observed that the
various ways, such as by improving speed and efficiency recommendation system has recommended the product
in content creation, ensuring consistency and quality of based on the user’s interest, and it was evaluated using an
content, generating personalized content, creating multi- A/B test to serve similar products in an ad carousel, which
lingual content, and repurposing existing content. It can can help the system to explore more products efficiently.
be trained on a company’s existing marketing materials • Forecast analysis: Using customer data analysis, GPTs
and customer data, allowing it to create new content, can forecast future behaviour and buying patterns. This
such as blog posts, social media updates, and product allows businesses to customize their marketing campaigns
descriptions, in a fast and efficient manner. Despite its to each customer’s desires based on their purchase pat-
speed, it maintain high standards for quality and consis- terns, increasing the likelihood of conversion or purchase.
tency. Moreover, GPTs [162] can generate personalized The authors in [170] have used chatGPT to perform pre-
content based on customer data, such as search history dictive modelling based on past data. They have used the
and past purchases. This helps create content that is rel- GPT model to predict the future based on the customer’s
evant to the users’ desires, leading to better engagement behaviour and buying pattern. This primarily helps the
and conversion rates. GPTs can also generate content in system to recommend the products to the customers as
various languages, allowing marketers to expand their per their desires.
reach across regions. Copy.ai [167] has used GPT-3 to 3) Challenges: GPTs are designed to generate content that
generate human-like text that is optimized for marketing imitates human writing, but the content generated may not
purposes such as website copy, social media posts, ad- align with the brand’s image or message. This lack of control
vertisement copy, and email campaigns. This means that can be a potential challenge for marketers. Another challenge
marketer personnel no longer focus on content creation. that applies to all learning technologies is that data bias is
Instead, they can spend productive time improving the possible in GPTs [138]. Based on the large dataset of text used
other aspects of marketing. for training, if the data is biased, it will affect the generated
• Customer service: GPTs can be trained on customer ser- content, which may also exhibit the same biases. GPT is
vice conversations and chat logs to generate more natural complex and difficult to interpret, making it challenging to
responses, like humans. This can help business person- explain how the model arrived at its conclusions. This lack
nel provide better customer service 24/7 and save time of transparency can lead to a lack of trust in adopting GPTs,
and resources. It can be trained to generate automated and marketing teams may struggle to make improvements in
responses for frequently asked questions, providing faster their strategies. As like every AI technology, there are ethical
responses to customers and ensuring consistency in the concerns associated with GPT models. For instance, the use
quality of replies. GPTs can also analyze customers’ emo- of GPT in marketing could raise concerns about the use of
tions and sentiments, enabling businesses to proactively personal data and privacy, particularly if the model is used to
address negative feedback. This is particularly helpful in generate targeted advertising or personalized content. To avoid
maintaining customers’ trust. The authors in [168] have any negative consequences, companies must ensure they use
used GPT-3 model for automated drafting of responses these models ethically and transparently.
for incoming mails. They used it to understand the mail, 4) Summary: Using GPTs in marketing can provide various
and then software engineering and business studies were benefits, such as better content creation, personalized messag-
used to understand the challenges encountered and finally, ing, increased efficiency, competitive advantage, and enhanced
the response generated after a thorough understanding of customer experience. However, this strategy also involves
 27

potential challenges, such as limited control, data bias, lack of market, exchange rates, and other financial metrics. This
transparency, and ethical considerations. Therefore, companies can help investors and financial organizations make more
must consider the advantages and drawbacks of GPT adoption accurate predictions and reduce their risk exposure. With
in marketing, and implement these models ethically and trans- the ability to analyze and process the natural language,
parently to avoid negative outcomes. Successful integration GPTs can be used to analyze and interpret financial data,
of GPTs in marketing requires proper planning, a skilled news, and other related information. Financial analysts
workforce, and continuous monitoring to ensure the desired and researchers can use the ability to analyze natural
results and mitigate any potential risks. language to extract insights from unstructured data like
news articles, social media content, and other information
that is relevant to forecasting. This can help improve the
K. Finance
accuracy of financial forecasting models by providing
1) Introduction: The finance industry, also known as the a more comprehensive view of market trends and sen-
financial sector, is a broad term that encompasses a wide range timents. This analysis may help improve the accuracy
of institutions and businesses that provide financial services to of prediction. Financial analysts can use the model to
individuals, businesses, and governments. The finance industry identify the relationship between the financial parameters
plays a critical role in the global economy, facilitating the flow that could change the market conditions in advance. This
of funds between savers and investors, managing risk, and prediction may be helpful for investors as they make
providing financial services and products to support economic investment decisions.
growth. The finance industry has been the leader in technology • Trading strategies: GPTs can also be used to analyze
adoption in recent years, with a focus on improving efficiency, market trends and historical data to develop trading
reducing costs, and delivering better customer experiences. strategies. This can help traders make better decisions in
The adoption of technologies like big data and analytics, terms of trading to increase their profitability. GPTs can
mobile and digital payments, blockchain and distributed ledger be used to identify the potential risks in trading portfolios.
technology, AI and ML, and cloud computing make the sector By analyzing the large volume of information related to
more flexible, scalable, trustworthy, transparent, secured, and trading, GPT will get the potential to identify the risk
easier to access. parameters and provide insights into how to mitigate these
2) Impact of GPT in Finance: GPT has greatly influenced risks. The authors in [173] have used a popular GPT for
finance by automating customer support using chatbots and stock market trend prediction. The results show that the
virtual assistants, enhancing fraud detection, offering invest- method used is simple but the efficiency and accuracy of
ment insights and recommendations based on financial data the method are very effective. The prediction it has made
and news, assisting with risk assessment for investments and is very close to the reality.
loans, impacting algorithmic trading strategies, simplifying • Risk prediction and management:The adoption of GPT
compliance with regulations by analyzing legal documents, can enhance the process of risk prediction and manage-
improving credit scoring and loan processes, and emphasizing ment in several ways. It can improve data analysis by
the importance of handling sensitive financial data securely detecting patterns that may pose a risk. It can also help
and transparently. in enhancing fraud detection by analyzing transaction
• Sentiment analysis: Sentiment analysis is a technique data and identifying fraudulent activity based on patterns.
used in the finance industry to evaluate the sentiment of Additionally, GPT can be utilized to make better portfolio
investors [171] and the general public towards specific management decisions by analyzing historical industry
companies, industries, or markets by analyzing news ar- data, company financial statements, and news articles, as
ticles, social media posts, and other text-based sources of well as social media feeds. This portfolio management
information. GPT has the potential to improve sentiment process can provide valuable information about the in-
analysis in finance by providing more accurate and de- vestment risk of a given organization, enabling informed
tailed analyses of financial data. With sentiment analysis, investment decisions and effective risk management.
the industry can predict stock prices by assessing the
sentiment of news articles, social media posts, and other 3) Challenges: GPTs have more challenges in the finance
sources of information about a particular company or in- sector. Primarily, they demand significant computational re-
dustry to make informed investment decisions. By utiliz- sources to train and deploy, which can be expensive and time-
ing sentiment analysis, GPTs can aid financial institutions consuming for financial organizations to implement. Another
in identifying potential risks and taking appropriate action challenge is that, even though GPTs are capable of producing
to mitigate them. The authors in [172] have investigated precise predictions, they can be challenging to interpret, which
how incorporating a lexicalized ontology can enhance can present a problem for financial institutions seeking to
the performance of aspect-based sentiment analysis by comprehend the reasoning behind specific predictions [174].
extracting indirect relationships in user social data. The This lack of interpretability can harm risk management objec-
investigation results show that the analysis has given 98% tives.Implementing GPT in finance sector can be vulnerable
accuracy. to adversarial attacks, which are designed to manipulate the
• Financial forecasting: GPTs can be trained on past fi- model’s output by injecting false data. This can be particularly
nancial market data to predict future trends in the stock problematic for financial institutions that rely on GPTs for risk
 28

management and investment decisions. It also require large SiriGPT for voice commands and ChatGPT for generating
amounts of training data to achieve high accuracy. However, text. SiriGPT utilizes a tokenizer exclusively developed by
in some cases, financial institutions may not have access to Apple that has been optimized for processing natural language
sufficient data to train the model effectively. GPTs can also tasks. SiriGPT’s training data is not publicly available as it is
be biased if the training data used to develop the model is exclusive to Apple. However, the language model is trained
biased. This can lead to inaccurate predictions and unintended on diverse text data from various sources such as books,
consequences. news articles, web pages, and other text data sources. This
4) Summary: The use of GPTs in the finance industry ensures that SiriGPT can handle different natural language
has promising benefits such as improved risk management, tasks accurately and efficiently. It has been reported that
enhanced fraud detection, better portfolio management deci- SiriGPT is one of the largest language models available, with
sions, and increased efficiency. However, it also has potential over a trillion parameters.
challenges that need to be addressed, such as high computa-
tional requirements, the complexity of implementation, limited B. AI Dungeon
interpretability, vulnerability to adversarial attacks, limited
training data, and bias in training data. So, the use of GPTs Latitude, a startup based in Utah, created a groundbreaking
in the finance industry presents significant benefits but also online game called AI Dungeon [178], which showcased a
requires careful consideration of the challenges involved to novel type of collaboration between humans and machines.
ensure the effective and secure deployment of these models. It is a free-to-play, single-player, and multiplayer adventure
game that caught traction within the gaming community. It
combines fantasy and AI to create endless possibilities, e.g.,
L. Summary On Impact of GPT models in Applications one can take charge of a military operation to defend against
The impact of GPTs in various applications and challenges aliens or become a famous detective investigating an attempted
was highlighted. GPT with its varied usage has changed murder of the queen of the fairies. Unlike games with prede-
the way people perceive facts such as content creation, en- termined storylines, AI Dungeon allows you to guide the AI
hanced user interfaces, personalized learning, item tracking, to generate unique characters and scenarios for your character
self-awareness, market risk analysis, business forecasts and to interact with. The game boasted about incorporating the
introspection. However, there are concerns about the potential GPT-3 text generator, but then the algorithm began producing
negative impact of GPTs, such as the spread of fake news, unsettling narratives, including graphic depictions of sexual
bias in data and decision-making, not domain specific, ethical encounters involving minors [179].
issues, data reliability, the complexity of implementation,
multimodal and multilingual support, security and privacy C. Copy.ai
concerns, vulnerable to data attacks, limited input data, ex-
plainability of results, large model size, high computational Copy.ai [180] is a mighty AI startup founded by Paul
requirements and job loss. Despite these concerns, it is clear Yacoubian in 2020. This project is created using GPT-3,
that GPTs will continue to be a powerful tool for industries mainly targeting business and marketing campaigns. It has the
seeking to leverage the power of NLP and generative AI. As following use cases: (i) For Teams: It assists with producing
the technology improves and new applications emerge, it will customized sales copy, composing long-form articles and
be interesting to see how GPTs continue to shape the future pages on a large scale, reusing content on various platforms,
of industries around the world. and creating product descriptions; (ii) For Emails: The AI-
powered email writer takes care of the most challenging parts
of marketing by creating email campaigns that are highly
V. PROJECTS effective at converting leads, all with just a few clicks of
This section presents the exciting projects developed using a button; (iii) For Blogs: By generating content briefs and
GPT model technologies for different applications mentioned crafting one-of-a-kind SEO-focused blog articles every month,
in the above sections. Table. IV, Table. V shows the different it can save a significant amount of money for the business.
levels of such projects along with different parameters to In addition, it’s feasible to create briefs, outlines, and even
compare their characteristics leveraging the capabilities in initial drafts in mere minutes, which can be utilized as an
many real-life applications. excellent source of inspiration for writers to create high-quality
content; (iv) Social Media: It aids in generating social media
A. SiriGPT posts quickly and efficiently, allowing for a rapid expansion
of the social media following. Additionally, Copy.ai includes
Siri [175] is an intelligent digital assistant that enables a suite of other tools, such as a headline analyzer, a language
Apple device users to complete tasks more efficiently and translator, and a content rephrase.
with ease, often anticipating their needs even before they make
requests. SiriGPT [176] [177] is a voice assistant powered by
a GPT model and developed entirely using shortcuts. Apple D. Bond.AI
device users can utilize ChatGPT, fueled by GPT-3, by using Bond.AI [181] is a company focused on AI for financial
an API key provided by OpenAI. This novel combination institutions, which has a headquarters in Little Rock, Arkansas.
offers the best of both worlds, allowing users to utilize It was established by Uday Akkarajuin in 2016 and prided
 29

TABLE IV
P ROJECT S UMMARY TABLE .
Project DeepScribe Meena Jukebox Uber’s plato research dialogue system Polyglot AI SiriGPT
Application Healthcare Lifestyle Entertainment Transport Education Lifestyle
widely used for
Purpose Medical documentation Personalized product rec- Enables the original mu- Enhances user experience using Uber enables absolute communication irre- Assist with voice-based assistants
and to improve doctor- ommendation sic creation both artisti- rides, helps drivers and riders in spective of the language barrier across
patient association cally compelling and com- scheduling rides, navigating routes, pro- different regions and cross-culturalism
mercially viable in a vari- viding real-time updates on traffic and
ety of styles and genres weather conditions.
GPT Adoption Customized version of Google’s seq2seq GPT-2 extension called GPT-2 GPT-0, GPT-1,GPT-2,GPT-3 GPT-3
GPT’s transformer-based neural ”Multi-Scale Transform-
network architecture ers for Music Modeling”
similar to Open AI’s GPT (MST) model
Dataset Not Disclosed Meena dataset over 40 bil- 1.2 million songs, 600,000 Persona-Chat with 160,000 conver- CONLL-2003, Sentiment140 dataset, Information not publisized
lion words , 341 GB cap- pieces of sheet music, sational dialogues, Cornell Movie- Reuters Corpus, 20 Newsgroups
tured from public domains 45,000 MIDI files Dialogs Corpus with 200,000 movie dataset, WMT (Workshop on Machine
like Reddit and social me- conversation, DailyDialog over 13,000 Translation) datasets and SQuAD
dia platforms dialogues, and (Stanford Question Answering Dataset)
Building Blocks Recurrent Neural Network Seq2Seq Transformer- Transformer-based Language modeling, Dialogue model- Language Identification, Named Entity Transformer-based neural network ar-
and Attention mechanism based Architecture Language Model and ing, Discrete latent variabe modeling Recognition (NER), Sentiment Analy- chitecture
fueled by NLP techniques Autoregressive model and response ranking sis, Text Classification, Machine Trans-
lation, Question Answering
Evaluation Met- Bleu score, perplexity Bleu score, perplexity Frechet Audio Distance Bleu score, Perplexity and Distinct n- accuracy, precision, recall, , F1-score, Perplexity, BLEU score, F1 score,
rics (FAD) and Pitch and gram Bleu score as well as cross-entropy loss ROUGE score, Human evaluation
Rhythm Similarity or perplexity
Addressed Chal- Reduced Transcription er- Natural and Engaging Fresh orginical music con- customer service, user experience, and Multilingualism and Sentiment Analy- Language understanding and genera-
lenges rors and enhanced patient conversations tent creation and drasti- operational efficiency sis are the key challenges in NLP and tion, Data scarcity, Contextual under-
care cally reducing the cost Polyglot AI solved this problem by standing, Text summarization, Senti-
and time by creating high- offering a tool for supporting morethan ment analysis, Named entity recogni-
quality music contents, 40 languages and pre-trained sentiment tion
and also to preserve and analysis model
advance musical heritage.
Input data Audio Text Audio Text Text Audio
Owned By DeepScribe Google OpenAI Uber Uizard Technologies Apple

TABLE V
P ROJECT S UMMARY TABLE ( CONTINUED ).

Project AI Dungeon Copy.ai Bond AI Viable AI Channels Fireflies.ai

Application Gaming Business and marketing Finance Business Analytics AI Industry Business
widely used for
Purpose Interactive and engaging help clients create written To enhance the financial provide businesses with intelligent in- provide a platform for developers, data to simplify the meeting process and
storytelling experience for content more quickly and well-being of clients sights to help them make better deci- scientists, and machine learning prac- reduce the time and energy required for
players easily sions titioners to create, deploy, and manage note-taking and collaboration
their AI models
GPT Adoption GPT-3 GPT-3 GPT-3 GPT-4 GPT-3 GPT-4
Dataset Common Crawl, OpenAI books, articles, and web- likely use of a combina- Information not publisized Users’ own dataset Possible datasets: the Common Voice
GPT-2, and various text sites tion of publicly available dataset from Mozilla having over 9,000
datasets from Kaggle financial datasets, propri- hours of speech data in multiple lan-
etary data, and client data guages
Building Blocks Machine Learning NLP, Language Models, NLP, Personalization, Unsupervised learning, Contextual un- Pre-built models, Model training, Data Speech-to-Text Technology, NLP,
Models, Text Input Neural Networks Conversational User derstanding, Sentiment analysis, Topic preparation, Collaboration Cloud Computing, Integration
Interface, Game Engine, Interface, Data Analytics modeling, Entity recognition technologies
Content Database, Player
Feedback System, Cloud
Infrastructure
Evaluation Met- Response Coherence, Re- Perplexity, BLEU score, Intent recognition Perplexity, Accuracy, F1 score, Word Accuracy, Precision and Recall, F1 Speech Recognition Accuracy, NLP
rics sponse Diversity, Player ROUGE score, F1 score accuracy, entity extraction similarity Score, Perplexity, User satisfaction Performance, Integration Performance,
Satisfaction, Engagement, accuracy, and language Task Completion Time, User Satisfac-
Realism, Novelty model perplexity tion
Addressed Chal- Narrative Generation, Lack of writing skills, In- Personal financial man- Understanding unstructured Natural language understanding, Scala- Time-consuming manual note-taking,
lenges Content Creation, consistency, Multilingual agement, Customer en- data, Contextual understanding, bility, Personalization, Integration with Difficulty in capturing important de-
Personalization, content creation gagement, Fraud detection Visualization and exploration of data, other systems, Maintenance and up- tails, Lack of visibility and accountabil-
Replayability, and prevention Customization and integration dates ity, Communication barriers
Accessibility, Creative
Expression
Input data Text Text Audio and Text Text Text Audio
Owned By Latitude Copy.ai Bond.AI Viable AI MiroMind AG Fireflies AI

itself on providing AI technology centred around human needs. data sources, such as social media posts, customer reviews,
This innovative project offers a product named BondBot, and survey responses, by employing Viable. GPT assists in
which is powered by Empathy Engine 3.0 and ChatGPT, to comprehending the sentiment and context behind the data,
enhance the financial health of clients. It assists financial insti- resulting in valuable insights that can enhance a company’s
tutions and employers in promoting interconnected finance by services, products, and customer experience. Viable’s ”Insight
offering various tools to improve the institution’s profitability Explorer” is a distinctive feature that enables users to interact
and the financial health of its clients on a single network. with and visualize their data via a user-friendly interface. In
It uses customer data to create individual personas for every addition, the platform offers advanced analytics capabilities,
bank customer or small business, considering their behaviours, including entity recognition, topic modelling, and sentiment
strengths, and potential needs. This approach enables the plat- analysis. The GPT-based technology of Viable is continually
form to develop multiple customized pathways to holistically evolving and advancing, which allows the platform to deliver
enhance clients’ financial well-being. more precise and insightful data. Moreover, Viable provides
customized integration and solutions to cater to the specific
E. Viable requirements of each business.
Viable [182] is a platform powered by GPT-4 that utilizes
the latest advancements in NLP and AI to offer businesses F. AI Channels
intelligent insights to aid their decision-making processes. AI Channels [183] is a platform that provides a comprehen-
Companies can extract actionable insights from unstructured sive set of tools for developers, data scientists, and machine
 30

learning practitioners to develop, launch, and manage their AI movie conversations, DailyDialog over 13,000 dialogues, and
models. The platform offers an all-in-one solution for creating EmpatheticDialogues over 25,000 user dialogues. The main
personalized AI models, starting from data preparation and components in developing the GPT-powered PLATO project
model training to deployment and monitoring. Users can train are language modelling, dialogue modelling, discrete latent
their models on their data or, on pre-trained models provided variable modelling and response ranking. The Plato Research
by AI Channels. These models can be deployed as APIs Dialogue System was trained on a massive corpus of text data
or Docker containers on various infrastructures, including consisting of over 40 GB of uncompressed text while Bleu
different cloud platforms. It also provides a dashboard for score, Perplexity and Distinct n-gram are the evaluation met-
tracking model performance and managing configurations. It rics used for training and testing the PLATO project. Uber’s
covers various use cases, including computer vision, NLP, and AI PLATO has addressed many key challenges like customer
speech recognition. The platform includes pre-built models for service by personalizing user feedback with conversational AI
tasks such as image and text classification, object detection, agent, user experience using the Uber platform for scheduling
and sentiment analysis. Additionally, users can create their rides, navigating routes, and providing real-time updates, and
models using popular frameworks. The main objective of AI increasing operational efficiency by reducing the need for
Channels is to make building and launching AI models more human customer service representatives and enabling faster
accessible to developers and businesses without specialized AI and more accurate communication between riders, drivers, and
skills. the Uber app.

G. Fireflies.ai I. Jukebox
Fireflies AI [184] is a privately held company based in San Jukebox, a GPT-powered music creation, was developed
Francisco, California, founded by Krish Ramineni and Sam in 2020 as an extension of Open AI’s GPT language model
Udotong. Fireflies AI software is powered by GPT-4 to au- [186]. Jukebox’s goal is to push the boundaries of what AI can
tomate notes-taking tasks and collaborations during meetings. accomplish in the world of music creation and to investigate
It is compatible with various video conferencing platforms, fresh applications for AI. A variation of the GPT architecture,
including Zoom, Google Meet, and Microsoft Teams, and it the ”Multi-Scale Transformers for Music Modeling” (MST)
can transcribe meeting audio and video content in real time. model, was created specifically to handle the intricate and
Its primary function is based on speech-to-text technology, multi-scale nature of musical data. Additionally, Jukebox can
which enables it to generate a searchable transcript of the produce lyrics that match the music’s tone and style. A sizable
meeting, which can be used for later review and to recall and varied dataset of musical recordings, lyrics, and related
essential points and action items. Additionally, the software metadata was used to train Jukebox such as 1.2 million songs
utilizes NLP capabilities that can identify significant keywords sourced including Lakh MIDI Dataset, Free Music Archive,
and phrases within the conversation. Apart from note-taking, Spotify and Tidal, 600,000 pieces of sheet music were sourced
Fireflies AI includes collaboration tools such as assigning tasks from IMSLP (International Music Score Library Project),
and sharing notes with other team members. It can integrate and 45,000 MIDI files from Lakh MIDI Dataset and the
with project management and task tracking tools to auto- MIDIworld collection. Faster training times and more effective
matically generate tasks based on the identified action items use of computational resources were made possible by the
during the meeting. Fireflies AI provides several customization distributed computing setup with 2048 TPU( Tensor Process-
options to suit particular use cases and workflows. Users can ing Unit) cores used to train the Jukebox model. Training the
configure the software to automatically join specific meetings model required significant computational resources demanding
or capture audio only from specific speakers. It allows users faster training times by the distributed computing setup with
to specify particular words and phrases to highlight in the 2048 TPU (Tensor Processing Unit) cores used to train the
transcript, making it easier to identify critical points during the Jukebox model. A combination of subjective and objective
later review. Thus, Fireflies AI aims to simplify the meeting metrics was used to assess and test Jukebox. In a large-scale
process and reduce the time and energy required for note- subjective assessment, more than 1,000 participants listened
taking and collaboration. and rated each one individually determining the overall score
for each song produced. On the other side, objective assess-
H. Uber’s Plato Research Dialogue System ments were conducted by evaluating Frechet Audio Distance
(FAD) and Pitch and Rhythm Similarity. Overall, Jukebox
Uber’s AI Lab introduced Uber’s Plato Research Dialogue
revolutionizes with its significant advancement in the music
System in 2020 developed by a team of researchers and engi-
industry through creative inspiration, music production, music
neers to enable the intelligence in riding experience. PLATO -
education and preservation of music heritage.
Pre-trained Dialogue Generation Model with Discrete Latent
Variable [185]. Uber’s Plato Research Dialogue System uses
GPT-2, a large-scale language model developed by OpenAI J. Meena
in 2019. Uber’s Plato Research Dialogue System project Google’s Meena project was developed by Google Research
used several datasets to train and evaluate their conversa- Team in 2020 for providing personalized product recom-
tional agents such as Persona-Chat contains 160,000 conversa- mendations [187]. The primary goals of the Meena project
tional dialogues, Cornell Movie-Dialogs Corpus with 200,000 empowered the lifestyle sector to enhance the user experience
 31

and customer service by recommending goods and services BLEU score (Bilingual Evaluation Understudy), METEOR
on a personalized basis. The project designed a GPT using (Metric for Evaluation of Translation with Explicit ORdering)
the seq2seq transformer-based neural network architecture, or F1-score. Remarkably, Facebook used new Polyglot AI
in particular for open-domain conversational agents. The ar- to translate between 100 languages [189]. Thus Polyglot AI
chitecture was pre-trained over 341 GB of text captured enables absolute communication irrespective of the language
from Reddit and other social platforms containing over 40 barrier across different regions and cross-culturalism.
million words and called this massive collection as ’Meena Thus, this section focused on several exciting real-life
Dataset’. Meena was tested using the automated performance projects which are developed and used for humankind. These
metrics known as Bleu score and perplexity on a cluster projects were discussed by presenting Table ?? highlighting
of HPC nodes with a total of 2048 NVIDIA V100 GPUs. the details of the project with model architecture, datasets
One of the biggest challenges solved Meena was building used, training and testing, and evaluation metrics involved with
trust and generating reliable engaging human-like conservation the challenges addressed. The next section will discuss the
that typically enhances user satisfaction and personalization. open research issues and future directions for the potential
Meena has achieved state-of-the-art performance compared to benefits of GPT models.
other open-domain chatbots and revolutionized the wide range
of applications in the lifestyle industry and a way beyond VI. OPEN RESEARCH ISSUES AND FUTURE DIRECTIONS
by providing natural and engaging responses through virtual This section highlights the various open research issues con-
assistants, customer service bots and personal shoppers. cerned with the implementation and adoption of sustainable
GPT models. It also provides insights into future research
K. DeepScribe directions for the betterment of researchers in the field of
DeepScribe was a GPT-based medical project developed GPT development. Fig. 9 outlines the many issues that can
in 2019 by the student team at the University of California develop while using GPT models, as well as the various future
by partnering with giant US-based healthcare providers such approaches that need to be considered for the effective usage
as One Medical, Stanford Medicine, Mount Sinai and Sutter of GPT models.
Health [188]. The DeepScribe’s technology aims at transcrib-
ing medical conversation allowing doctors to treat the patients A. Domain Specific GPT models
rather than noting down the patient’s history, enhancing the
Domain-specific GPT models are mandated in almost all
doctor-patient relationship and targeting the overall quality
applications; developing these models is still challenging and
of patient care. Although DeepScribe used the customized
an open issue within GPT. While the current GPT models have
variants of Open AI, the technical details of the GPT model
been developed to understand natural language and generate
used for customizing the model were not disclosed which was
content effectively, their performance may not be equally
optimized for medical transcription tasks.
effective when handling specific domains, such as medicine,
agriculture, etc. One of the key challenges in adapting to a
L. Polyglot AI particular domain is the availability of domain-specific data.
Polyglot AI is a communication platform designed to gen- It is well known that the performance of GPTs is directly
erate text in multiple languages and process the data by proportional to the quality and quantity of data used for train-
performing several tasks such as advanced NLP techniques, ing the model. So, obtaining such quality data for a specific
text translation, and sentiment analysis. The potential features domain is expensive and time-consuming, as the data are
of Polyglot AI have been exploited in the following application heterogeneous. Also, these data accumulations may even make
areas such as language translation, chatbots, language learning these models much larger, sometimes catastrophic too, leading
tools, content creation, customer support, and data analysis to forgetting the knowledge attained during the process. To
across different languages and regions. Polyglot AI is built overcome this issue, pre-training tasks and domain-specific
based on different variants of GPT models, and state-of-the- model generation are integrated by data augmentation [190].
art language model architecture for NLP tasks, which uses the Another challenge is fine-tuning the model to accustom to the
self-supervised learning approach. unique characteristics and vocabulary of the domain. A few
The Polyglot AI was pre-trained using a large amount domain-specific GPT models have been developed and imple-
of textual data on multiple languages simultaneously in an mented despite these challenges. There is a growing interest
unsupervised environment using a shared architecture, Multi- in creating more domain-specific GPTs for various domains.
lingual Universal Sentence Encoder (MUSE). MUSE devel- Moreover, these models will be trained using the knowledge
oped by Google, is a pre-trained DL model used for cross- acquired from large language models specific to domains.
lingual TL, that encodes the text into common vector space Therefore, these models can be fine-tuned for specific tasks
for multiple languages. Thus, the Polyglot language model or domain-specific requirements with gradually improving
was created with the following pre-training techniques as performance. GPT models have the potential to be trained in
Masked Language Modeling (MLM), Translation Modeling any context, and researchers are exploring new approaches
Language (TML), sequence-to-sequence modelling and cross- and methods to address these challenges. Furthermore, these
lingual TL. The pre-trained language model is fine-tuned and models will be more efficient, enhanced interpretability, and
evaluated by standard benchmarks and metrics such as the domain generability than the existing Large language models
 32

Fig. 9. Challenges and Future Directions.

as they are customized to specific domain concerns and can sides these challenges, domain-specific models require higher
provide more concise and informative solutions. TL can be computation costs for the resources and time spent in pre-
used for developing domain-specific GPT models. Domain- training and relearning in downstream tasks during fine-tuning
specific GPT models were developed to summarize products of pre-trained domain-specific models. Therefore, domain-
based on customer reviews on an E-commerce site, where the specific model development must focus on optimizing resource
language model is pre-trained on the Chinese-short summa- consumption and fine-tuning the pre-trained model to alleviate
rization dataset and has obtained fine-tuned results [130]. Be- the forgetting problem involved in existing models.
 33

B. High Computational requirements C. Explainability and interpretability

Explainability and interpretability are currently major chal-
lenges for GPTs for specific applications. Explainability refers
As the Transformer model utilizes varied heterogeneous
to providing a clear and understandable explanation of how
datasets for training and learning from the knowledge ac-
the model has arrived at any output. Interpretability, on the
quired, one of the key challenges of GPT models is high
other hand, refers to the ability to understand the internal
computational resources for pre-training and inference. The
processes of the model. GPT models are highly complex
computational requirement continuously increases as the mod-
and difficult to understand and interpret due to their size
els become more complex and larger. Depending on the size
and architecture. The outcomes and decisions of the model
and complexity of the model and the available resources, the
are based on previous learning and training, and the models
time required to train the model can take days, weeks, or
learn from vast amounts of data to make decisions. These
even months. Moreover, the inference time for these models
decisions may not be easily explainable to humans. This lack
is typically slower, making it challenging to use them for
of transparency and interpretability raises concerns about the
real-time applications. This poses a significant obstacle to
reliability and safety of the model, particularly in critical ap-
adopting GPT models for many practical applications. Despite
plications such as healthcare and finance. Researchers are cur-
these challenges, significant efforts are underway to overcome
rently conducting much research to make GPT models more
them. To accommodate the increasing data size and pre-
explainable and interpretable [192] by utilizing EXplainable
training computational requirement, data enhancement-based
Artificial Intelligence (XAI) to provide explanations for the
GPT models were developed [190] by joining the downstream
decisions arrived at, specifically to different users at stake. As
tasks and pertaining process by reconstructing the domain-
well, XAI models enable interpretability by providing detailed
specific text before proceeding for pre-training and utilizing
explanations for the internal process. As GPT can generate
the empirical knowledge rather than learning for falsy domain
any type of unconstrained output for instance code generation
data. Researchers are exploring various ways to optimize and
for the given problem, it requires proper justifications and
speed up the training and inference process, such as using
explanations for the output. So, to assure these codes by
specialized GPUs and TPUs. They are also developing more
GPT are reliable, a metric model to evaluate and validate
efficient algorithms and attempting to reduce the model size
this GPT code was developed using NLP metrics and XAI
without sacrificing performance. In addition, ChatGPT has
for model interpretability [193]. Also, some domain-specific
evolved to include plugins [191] that enable statistical analysis
GPT models of GPT-3 have evolved with solutions [194] [195]
for real-time applications. By integrating these plugins with
to ensure that the GPT model’s decisions are understandable,
the help of third-party services, ChatGPT can now be used
explainable, and trustworthy enough to be used for critical
for analyzing real-time applications as well.
applications like healthcare and finance.
1) Increasing Model size and Space Constraints: Devel-
oping and training large language models, such as a GPT D. Data Bias
model, can be a challenging task due to significant technical
Data bias is an open issue concerned with the adoption
and computational difficulties as discussed. The size of GPT
of any advancements in AI, till GPT [196]. This is also
models presents a major challenge, as the computational
a prominent challenge for GPT and other machine-learning
resources required for training and inference increase with
models. It refers to patterns or relationships in the data that
the number of parameters that need to be trained. As the
do not accurately reflect the true distribution of the target
model size increases, it also requires more memory to store
population or domain. GPT models are trained on vast amounts
and manipulate parameters during training and inference [76].
of text data which may contain bias in language use or cultural
Acquiring and processing vast amounts of high-quality training
assumptions. Still, the source of data remains undeclared, con-
data is another challenge in training large language models like
sidering GPTs are trained using internet data which may have
GPT. For instance, GPT-4, which is the largest GPT model to
faulty, fake, and error data, GPTs may generate biased texts or
date with 1 trillion parameters, demands a massive amount
information imitating the training data [197]. Such biases can
of computational resources, such as specialized hardware like
be amplified in the model’s output, resulting in false or unfair
GPUs and TPUs, spatial requirements, and high-speed network
results. Data bias can arise from various sources, such as selec-
connections to transfer data between different parts of the
tion bias, labelling bias, concept drift, confounding variables,
system. Model evaluation and interpretation are also critical
and changes in input data distribution over time. For example,
challenges. Since large language models like GPT are trained
suppose a dataset used to train a GPT model is dominated
on a massive scale, understanding how the model makes
by a particular demographic group. In that case, the resulting
predictions and why it generates specific outputs is difficult.
model may be biased in its predictions towards that group,
Evaluating the quality and accuracy of the model’s output and
leading to inaccurate or unfair predictions when applied to new
identifying and addressing biases or errors in its performance
data. This bias can have serious consequences, especially in
can also be challenging.
healthcare, finance, and law enforcement, where biased results
As these and other significant efforts continue, we can can significantly impact human lives. To mitigate these issues,
expect the challenge of computational resource requirements researchers have developed strategies such as diversifying the
for GPT models to transform into a strength in the future. training data, debiasing the training data, modifying the model
 34

architecture, and using post-processing methods to normalize result, they may be influenced by subtle patterns or biases in
the data and create more fair and inclusive GPT models. the training data. If such biases or patterns exist in the data,
The authors in [198] have made an in-depth analysis of the the GPT model may amplify or perpetuate existing biases,
most downloaded text generation model GPT2. By examining leading to unfair outcomes. A few techniques may be used,
the intersections of gender with religion, sexuality, ethnicity, such as adversarial training [204] [205], defensive distillation
political affiliation, and continental name origin, the authors [206], and regularization techniques [207] such as dropout,
evaluated prejudices associated with occupational associations weight decay, and batch normalization, to mitigate and handle
among various protected categories. These biases may have adversarial inputs. Therefore, GPT development must focus on
inaccuracies in climatic data prediction or global warming developing models with more robustness, enabling them to be
[199]. Therefore, data bias must be of greater concern in GPT tolerant of various vulnerabilities, and thus to be used reliably
model development as the data quality of the internet is limited and susceptible in a wide range of applications.
to avoiding producing disturbing content.

E. Multimodal support G. Multilingual support

The challenge of developing multimodal learning ability in While GPT models have demonstrated remarkable profi-
the GPT model remains unsolved. Multimodal support refers ciency in NLP tasks for individual languages, achieving mul-
to the GPT model’s ability to process and generate text along tilingual support remains a significant challenge. The primary
with other modalities, such as audio, images, and videos. GPT difficulty in developing multilingual GPT models lies in the
models have shown impressive results in generating high- significant differences in syntax, grammar, and vocabulary
quality text and NLP tasks, but it was primarily designed for across various languages. As the number of internet users
text-based tasks and cannot handle other modalities. However, day by day increasing irrespective of literacy rate, multilingual
due to its success in text processing, users expect its integra- support will target all types of end users. To create models that
tion with other modalities, such as speech recognition, video can effectively process multiple languages, researchers need to
summarization, and image or video captioning [200]. Several train GPT models on extensive, diverse datasets that span a
research initiatives have been proposed to integrate multimodal broad range of languages and language families. Addition-
support to address this issue. One approach is to feed the visual ally, designing language-specific pre-processing techniques
and audio information with the corresponding text to the model to prepare input data for the model is another obstacle to
as input. The other is to handle this input modality process as a overcome. Various languages possess distinct writing systems,
separate model and use the output as input to GPT. Multimodal word orders, and linguistic features, necessitating specialized
video captioning is done using GPT in the unlabelled videos pre-processing techniques to ensure that the model can process
[200]. Multimodel learning has been applied for information the input data effectively. Despite the challenges, researchers
retrieval [201] and image generation for illustrating the news continue to explore new methods to improve the multilingual
[202] to assist the GPTs. However, the primary challenge in capabilities of GPT models. Some techniques involve training
both approaches is effective integration, requiring architectural separate models for each language or developing language-
changes and techniques to handle various modalities. Recently, specific fine-tuning techniques. Others include developing
OpenAI’s GPT4 has launched with multimodal support, en- cross-lingual TL techniques that allow the model to transfer
abling it to read images, analyze the input, and generate text as knowledge and skills learned in one language to another.
output. It cannot create images as output, though. Nevertheless,
the field of multimodal processing is still an active area of
research, and much work must be done to effectively and H. Limited understanding
efficiently process and understand multimodal data.
GPT models have a limited understanding of context and
meaning, despite their ability to generate coherent text. This
F. Robustness problem arises due to issues such as a lack of semantic
The robustness is a major requirement to be imposed by any understanding, bias, stereotyping, and handling nuances and
type of GPT model, and it is a global problem for all learning- figurative language. As a result, the outputs generated by the
based prediction technologies. Robustness refers to the ability model may contain errors or inaccuracies, even if they are
of the model to maintain high performance and accuracy even grammatically correct. Researchers are exploring various tech-
in the face of unexpected or adversarial inputs. Although GPT niques to enhance the model’s contextual understanding. Un-
models have shown impressive performance in a wide range of derstanding GPTs will be more reactive and may attract more
NLP applications and have set a benchmark for high-quality users for accurate results [208]. These methods include incor-
text generation, they are still vulnerable to certain types of porating external knowledge sources like knowledge graphs
errors and attacks. In particular, handling adversarial inputs is a and ontologies into the training process, developing common
challenging task in GPT models. GPT models are particularly sense reasoning capabilities, and improving the model’s ability
susceptible to adversarial attacks [203]. Adversarial inputs are to handle nuances and idiomatic expressions. By enhancing the
specifically designed to make a learning model collapse and contextual understanding of GPT models, their outputs will
misbehave. GPT models can be highly prone to these attacks be more accurate, relatable, sequential, less biased, and more
because they are trained on a large volume of text. As a useful for a variety of applications.
 35

I. Ethical Concerns to ensure the ethical and safe use of GPT models before using
them in various applications.
The ethical concerns in GPT models are an active area of
Therefore, GPT model development must focus on devel-
discussion and debate due to the potential negative impacts that
oping more robust, reliable, safest, multi-lingual, multimodal
the use of GPTs could have on society. Although GPT models
support-enabled solutions for delivering domain-specific or
have demonstrated remarkable abilities in generating coherent
human-specific solutions with optimal resource utilization.
and realistic text, there are concerns about the perpetuation
of biases and stereotypes, the possibility of malicious use, VII. C ONCLUSION
and the effects on employment and economic inequality.
The impact of GPT and other large language models is
Some of the ethical characteristics to be possessed by GPT
far-reaching and profound. As these technologies continue to
include functional Morality, operational morality, abiding by
evolve and improve, they have the potential to transform the
the right for explanation law, improved transparency with
way we interact with technology and each other. From per-
human involvement, unbiased data, and adhering to govern-
sonalized recommendations and customer service to language
ment regulations on data usage [196]. The responsibility of
translation and text generation, the possibilities are endless.
developers and companies to address these ethical concerns
However, as with any technology, there are potential ethical
and ensure the ethical use of a GPT model is also a topic
and societal concerns that must be addressed. As we continue
of debate. The ethical implications of GPT models are being
to rely more heavily on these language models, we must
actively researched and discussed in the fields of AI, computer
ensure that we are using these tools responsibly and with
science, and philosophy.
consideration for their impact on society as a whole. These
include challenges related to biases in the data used to train
the models, safeguarding privacy and security, understanding
J. Security and privacy concerns
the implications of human creativity, and the potential impact
GPT models raise concerns about security and privacy, on employment and job displacement. We need to continue to
particularly as they become more widespread. One of the main evaluate and reflect on the impact of GPT and other language
concerns is that GPT could be used for harmful purposes, such models, to ensure that they are being used in a way that
as creating fake news or deep fakes, as it can generate text that benefits society as a whole. By doing so, we can help to ensure
looks real and convincing, making it difficult to distinguish that these technologies are used to their fullest potential while
between genuine and fake content. Another concern is the po- minimizing any negative impact that they may have.
tential for privacy violations when using a GPT model. Large
language models like GPT require a significant amount of R EFERENCES
training data, which could contain sensitive or personal infor- [1] X. Han, Z. Zhang, N. Ding, Y. Gu, X. Liu, Y. Huo, J. Qiu, Y. Yao,
mation. This raises concerns about privacy and data protection A. Zhang, L. Zhang et al., “Pre-trained models: Past, present and
future,” AI Open, vol. 2, pp. 225–250, 2021.
as per European Union’s General Data Protection Act [209], [2] ”Introducing OpenAI”. [Accessed on 23.03.2023]. [Online]. Available:
particularly if the training data is not properly anonymized https://openai.com/blog/introducing-openai
or if the models are used to generate text based on user data [3] L. Dong, S. Xu, and B. Xu, “Speech-transformer: a no-recurrence
sequence-to-sequence model for speech recognition,” in 2018 IEEE
without their explicit consent. Some of the problems concerned international conference on acoustics, speech and signal processing
with confidentiality related to the pre-training dataset are (ICASSP). IEEE, 2018, pp. 5884–5888.
Data tracing, Membership Inference Attacks, reconstruction [4] B. D. Lund and T. Wang, “Chatting about chatgpt: how may ai and
gpt impact academia and libraries?” Library Hi Tech News, 2023.
attacks, and property inference attacks and the vulnerabilities [5] E. Kasneci, K. Seßler, S. Küchemann, M. Bannert, D. Dementieva,
concerned with a model encoder are hyperparameter stealing F. Fischer, U. Gasser, G. Groh, S. Günnemann, E. Hüllermeier et al.,
attacks and encoder parameter stealing attacks. Poisoning, “Chatgpt for good? on opportunities and challenges of large language
models for education,” Learning and Individual Differences, vol. 103,
Backdoor, and evasion attacks are the vulnerability related to p. 102274, 2023.
the integrity of self-supervised learning. Resource depletion [6] X. Qiu, T. Sun, Y. Xu, Y. Shao, N. Dai, and X. Huang, “Pre-trained
attack is one major issue with data availability, which may lead models for natural language processing: A survey,” Science China
Technological Sciences, vol. 63, no. 10, pp. 1872–1897, 2020.
to tremendous effects incorrect results, and may cause greater [7] A. S. George, A. H. George, T. Baskar, and A. G. Martin, “Revolu-
deviations too [209]. Additionally, the GPT model’s ability to tionizing business communication: Exploring the potential of gpt-4 in
generate text based on user input could inadvertently disclose corporate settings,” Partners Universal International Research Journal,
vol. 2, no. 1, pp. 149–157, 2023.
sensitive information, such as personal or financial details, or [8] C. Zhang, C. Zhang, S. Zheng, Y. Qiao, C. Li, M. Zhang, S. K. Dam,
trade secrets. This could happen if a GPT model is used in an C. M. Thwal, Y. L. Tun, L. L. Huy et al., “A complete survey on
insecure environment or if it is targeted by malicious actors generative ai (aigc): Is chatgpt from gpt-4 to gpt-5 all you need?”
arXiv preprint arXiv:2303.11717, 2023.
seeking to obtain sensitive information. Researchers and devel- [9] M. Zaib, Q. Z. Sheng, and W. Emma Zhang, “A short survey of pre-
opers should focus on assuring authenticity in using users’ data trained language models for conversational ai-a new age in nlp,” in
in case of interactive information generation based on privacy Proceedings of the Australasian computer science week multiconfer-
ence, 2020, pp. 1–4.
data shared. These include using differential privacy to protect [10] Z. Liu, X. Yu, L. Zhang, Z. Wu, C. Cao, H. Dai, L. Zhao,
training data privacy [210], implementing secure hardware or W. Liu, D. Shen, Q. Li et al., “Deid-gpt: Zero-shot medical text de-
software protocols to protect models from cyberattacks, and identification by gpt-4,” arXiv preprint arXiv:2303.11032, 2023.
[11] P. Rivas and L. Zhao, “Marketing with chatgpt: Navigating the ethical
developing techniques to detect and prevent the malicious use terrain of gpt-based chatbot technology,” AI, vol. 4, no. 2, pp. 375–384,
of GPT models. It’s crucial to adopt and follow these measures 2023.
 36

TABLE VI
VARIOUS LESSONS LEARNED AND FUTURE RESEARCH DIRECTIONS .

Sl.No Lessons Learned Open Issues Future Directions

1. ◦ Huge volume of data usage ◦ Data privacy - may unknowingly reveal sensitive ◦ Optimized architecture and algorithms
is critical information ◦ Cloud-based computing
◦ Varied data quality - Inconsistency in quality of ◦ Hardware advancements
data used for training
◦ Scalability - Models should be able to handle an
increase in data set size and complexity
2. ◦ Importance of Proper Pre- ◦ Data bias - Overrepresentation of certain groups ◦ Continuous monitoring
processing of data or perspectives ◦ Testing model for potential biases
◦ Poor model performance ◦ Diversifying the training data
◦ Reduced efficiency of the model
3. ◦ Importance of explainability ◦ Complexity of models ◦ AI governance models can be used
and interpretability ◦ Inability to explain predictions ◦ Model Summaries can be provided
◦ Need of User-tailored Explanations generation ◦ Techniques like LIME(Local Model-Agnostic Ex-
◦ Developing Interpretable models planations can be used
◦ Lack of transparency in the data source ◦ Uncertaining estimates can be obtained from a
model

4. ◦ Ethical concerns ◦ Data privacy and data protection ◦ Counterfactual analysis can be used
◦ Misuse of data ◦ Federated learning can be used
◦ Accountability and transparency concerns ◦ Ethical guidelines, Legal frameworks and regula-
◦ Societal implications - displacing jobs and exacer- tions can be developed to avoid harmful use
bating equalities

5. ◦ Lack of contextual under- ◦ Possibility for ambiguous, contradictory, incorrect ◦ Incorporation of knowledge graphs and semantic
standing in AI systems results leads to misunderstandings embeddings into the training process
◦ Inconsistency in responses or outputs ◦ Usage of attention mechanisms to focus on relevant
◦ Lack of ability in distinguishing true and false parts of the input
information ◦ Imparting reasoning and inference capabilities
◦ Task or domain-based fine-tuning

6. ◦ Pre-trained models may not ◦ Possibility for ambiguous, contradictory, incorrect ◦ Incorporation of knowledge graphs and semantic
perform well for Domain- results leads to misunderstandings embeddings into the training process
specific task ◦ Inconsistency in responses or outputs ◦ Usage of attention mechanisms to focus on relevant
◦ Lack of ability in distinguishing true and false parts of the input
information ◦ Imparting reasoning and inference capabilities
◦ Task or domain-based fine-tuning

[12] M. Leippold, “Thus spoke gpt-3: Interviewing a large-language model Intelligent System and Industrial Applications (CITISIA), 2021, pp. 1–
on climate finance,” Finance Research Letters, vol. 53, p. 103617, 2023. 7.
[13] M. Trajtenberg, “Ai as the next gpt: a political-economy perspective,” [24] V. Taecharungroj, ““what can chatgpt do?” analyzing early reactions
National Bureau of Economic Research, Tech. Rep., 2018. to the innovative ai chatbot on twitter,” Big Data and Cognitive
[14] D. Haluza and D. Jungwirth, “Artificial intelligence and ten societal Computing, vol. 7, no. 1, p. 35, 2023.
megatrends: An exploratory study using gpt-3,” Systems, vol. 11, no. 3, [25] G. Spitale, N. Biller-Andorno, and F. Germani, “Ai model gpt-3 (dis)
p. 120, 2023. informs us better than humans,” arXiv preprint arXiv:2301.11924,
[15] L. J. Quintans-Júnior, R. Q. Gurgel, A. A. d. S. Araújo, D. Correia, 2023.
and P. R. Martins-Filho, “Chatgpt: the new panacea of the academic [26] J. Ye, X. Chen, N. Xu, C. Zu, Z. Shao, S. Liu, Y. Cui, Z. Zhou,
world,” pp. e0060–2023, 2023. C. Gong, Y. Shen et al., “A comprehensive capability analysis of gpt-3
[16] Q. Zhu and J. Luo, “Generative pre-trained transformer for design and gpt-3.5 series models,” arXiv preprint arXiv:2303.10420, 2023.
concept generation: an exploration,” Proceedings of the Design Society, [27] J. Liu, D. Shen, Y. Zhang, B. Dolan, L. Carin, and W. Chen,
vol. 2, pp. 1825–1834, 2022. “What makes good in-context examples for gpt-3?” arXiv preprint
[17] ”GPT”. [Accessed on 25.03.2023]. [Online]. Available: https: arXiv:2101.06804, 2021.
//aidungeon.io/ [28] M. Bommarito II and D. M. Katz, “Gpt takes the bar exam,” arXiv
[18] M. Kosinski, “Theory of mind may have spontaneously emerged in preprint arXiv:2212.14402, 2022.
large language models,” arXiv preprint arXiv:2302.02083, 2023. [29] T. Hagendorff, S. Fabi, and M. Kosinski, “Machine intuition: Uncov-
[19] ”GPT-1, GPT-2 and GPT-3 models explained”. [Accessed on ering human-like intuitive decision-making in gpt-3.5,” arXiv preprint
27.03.2023]. [Online]. Available: https://360digitmg.com/blog/types- arXiv:2212.05206, 2022.
of-gpt-in-artificial-intelligence [30] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N.
[20] B. Ghojogh and A. Ghodsi, “Attention mechanism, transformers, bert, Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”
and gpt: Tutorial and survey,” 2020. Advances in neural information processing systems, vol. 30, 2017.
[21] ”Generative Pre-trained Transformer 3 by Ope- [31] W. Hou and Z. Ji, “Geneturing tests gpt models in genomics,” bioRxiv,
nAI”. [Accessed on 23.03.2023]. [Online]. Avail- pp. 2023–03, 2023.
able: https://medium.com/@shripad.kulkarni18/generative-pre-trained- [32] S. Edunov, A. Baevski, and M. Auli, “Pre-trained language model
transformer-3-by-openai-4abe6614c8ef representations for language generation,” in Proceedings of the 2019
[22] N. Williams, S. Ivanov, and D. Buhalis, “Algorithmic ghost in the re- Conference of the North American Chapter of the Association for
search shell: Large language models and academic knowledge creation Computational Linguistics: Human Language Technologies, Volume
in management research,” arXiv preprint arXiv:2303.07304, 2023. 1 (Long and Short Papers). Minneapolis, Minnesota: Association
[23] M.-T. Nguyen, P.-T. Nguyen, V.-V. Nguyen, and Q.-M. Nguyen, “Gen- for Computational Linguistics, Jun. 2019, pp. 4052–4059. [Online].
erating product description with generative pre-trained transformer 2,” Available: https://aclanthology.org/N19-1409
in 2021 6th International Conference on Innovative Technology in [33] A. Rahali and M. A. Akhloufi, “End-to-end transformer-based models
in textual-based nlp,” AI, vol. 4, no. 1, pp. 54–110, 2023.
 37

[34] J. R. Stevens, R. Venkatesan, S. Dai, B. Khailany, and A. Raghunathan, [59] S. Balkus and D. Yan, “Improving short text classification with
“Softermax: Hardware/software co-design of an efficient softmax for augmented data using gpt-3,” arXiv preprint arXiv:2205.10981, 2022.
transformers,” in 2021 58th ACM/IEEE Design Automation Conference [60] H. Hua, Y. Li, T. Wang, N. Dong, W. Li, and J. Cao, “Edge computing
(DAC). IEEE, 2021, pp. 469–474. with artificial intelligence: A machine learning perspective,” ACM
[35] M. Bangura, K. Barabashova, A. Karnysheva, S. Semczuk, and Computing Surveys, vol. 55, no. 9, pp. 1–35, 2023.
Y. Wang, “Automatic generation of german drama texts using fine tuned [61] J. Yao, S. Zhang, Y. Yao, F. Wang, J. Ma, J. Zhang, Y. Chu, L. Ji,
gpt-2 models,” arXiv preprint arXiv:2301.03119, 2023. K. Jia, T. Shen, A. Wu, F. Zhang, Z. Tan, K. Kuang, C. Wu, F. Wu,
[36] J. Savelka, A. Agarwal, C. Bogart, and M. Sakr, “Large language J. Zhou, and H. Yang, “Edge-cloud polarization and collaboration: A
models (gpt) struggle to answer multiple-choice questions about code,” comprehensive survey for ai,” IEEE Transactions on Knowledge and
arXiv preprint arXiv:2303.08033, 2023. Data Engineering, pp. 1–1, 2022.
[37] H. Liu, Y. Cai, Z. Lin, Z. Ou, Y. Huang, and J. Feng, “Variational latent- [62] Y. Yu and S. Lee, “Measurements of the benefits of edge computing
state gpt for semi-supervised task-oriented dialog systems,” IEEE/ACM on autonomous driving,” in 2022 13th International Conference on
Transactions on Audio, Speech, and Language Processing, vol. 31, pp. Information and Communication Technology Convergence (ICTC).
970–984, 2023. IEEE, 2022, pp. 2155–2159.
[38] ”GPT-3, explained: This new language AI is uncanny, funny — [63] D. Yuan, L. Cui, M. Xie, Z. Su et al., “Paratra: A parallel transformer
and a big deal”. [Accessed on 23.03.2023]. [Online]. Available: inference framework for gpus in edge computing.”
https://www.computerhope.com/jargon/g/gpt.html [64] X. Zhou, H. Liu, C. Shi, and J. Liu, Deep Learning on Edge Computing
[39] A. Hendy, M. Abdelrehim, A. Sharaf, V. Raunak, M. Gabr, H. Mat- Devices: Design Challenges of Algorithm and Architecture. Elsevier,
sushita, Y. J. Kim, M. Afify, and H. H. Awadalla, “How good are 2022.
gpt models at machine translation? a comprehensive evaluation,” arXiv [65] K. Li, K. Chen, S. Luo, H. Zhang, and P. Fan, “Ubinn: A commu-
preprint arXiv:2302.09210, 2023. nication efficient framework for distributed machine learning in edge
[40] ”OpenAI GPT-n models: Shortcomings & Advantages in 2023”. computing,” IEEE Transactions on Network Science and Engineering,
[Accessed on 23.03.2023]. [Online]. Available: https://research. 2023.
aimultiple.com/gpt/ [66] C. Benzaı̈d, T. Taleb, and M. Z. Farooqi, “Trust in 5g and beyond
[41] S. Pramanik and S. K. Bandyopadhyay, “Analysis of big data,” in networks,” IEEE Network, vol. 35, no. 3, pp. 212–222, 2021.
Encyclopedia of Data Science and Machine Learning. IGI Global, [67] J. S. Wey and J. Zhang, “Passive optical networks for 5g transport:
2023, pp. 97–115. Technology and standards,” Journal of Lightwave Technology, vol. 37,
[42] A. Zaremba and E. Demir, “Chatgpt: Unlocking the future of nlp in no. 12, pp. 2830–2837, 2019.
finance,” Available at SSRN 4323643, 2023. [68] S. Zhang, W. Y. B. Lim, W. C. Ng, Z. Xiong, D. Niyato, X. S. Shen,
[43] N. Aleksić, M. Arnelid, D. Lisy, A. Balachandran, M. Belfrage, and C. Miao, “Towards green metaverse networking: Technologies,
A. Brännström, M. Busarello, S. A. Carretta, K. Cotton, S. de Heer advancements and future directions,” IEEE Network, 2023.
et al., “Deep learning as a gpt and disruptor in innovation processes.” [69] Z. Li, M. A. Uusitalo, H. Shariatmadari, and B. Singh, “5g urllc:
[44] S. Bubeck, V. Chandrasekaran, R. Eldan, J. Gehrke, E. Horvitz, Design challenges and system concepts,” in 2018 15th International
E. Kamar, P. Lee, Y. T. Lee, Y. Li, S. Lundberg et al., “Sparks of Symposium on Wireless Communication Systems (ISWCS), 2018, pp.
artificial general intelligence: Early experiments with gpt-4,” arXiv 1–6.
preprint arXiv:2303.12712, 2023. [70] A. Gonzalez Fanfalone et al., “The road to 5g networks: experience to
[45] D. M. Katz, M. J. Bommarito, S. Gao, and P. Arredondo, “Gpt-4 passes date and future developments,” 2019.
the bar exam,” Available at SSRN 4389233, 2023. [71] Y. Rogers, H. Sharp, and J. Preece, Interaction design: beyond human-
[46] D. O. Beerbaum, “Generative artificial intelligence (gai) ethics computer interaction. John Wiley & Sons, 2023.
taxonomy-applying chat gpt for robotic process automation (gai-rpa) [72] Y. Liu, M. Yu, M. Jiang, and Y. Huang, “Creative research question
as business case,” Available at SSRN 4385025, 2023. generation for human-computer interaction research,” 2023.
[47] H. A. Dida, D. Chakravarthy, and F. Rabbi, “Chatgpt and big data: [73] P. Hämäläinen, M. Tavast, and A. Kunnari, “Evaluating large language
Enhancing text-to-speech conversion,” Mesopotamian Journal of Big models in generating synthetic hci research data: a case study,” in ACM
Data, vol. 2023, pp. 33–37, 2023. SIGCHI Annual Conference on Human Factors in Computing Systems.
[48] V. Pereira, E. Hadjielias, M. Christofi, and D. Vrontis, “A systematic ACM, 2023.
literature review on the impact of artificial intelligence on workplace [74] A. Shafeeg, I. Shazhaev, D. Mihaylov, A. Tularov, and I. Shazhaev,
outcomes: A multi-process perspective,” Human Resource Management “Voice assistant integrated with chat gpt,” Indonesian Journal of
Review, vol. 33, no. 1, p. 100857, 2023. Computer Science, vol. 12, no. 1, 2023.
[49] M. Trajtenberg, “Artificial intelligence as the next gpt: A political- [75] J. Zhang, J. Pu, J. Xue, M. Yang, X. Xu, X. Wang, and F.-Y.
economy perspective,” in The economics of artificial intelligence: An Wang, “Hivegpt: Human-machine-augmented intelligent vehicles with
agenda. University of Chicago Press, 2018, pp. 175–186. generative pre-trained transformer,” IEEE Transactions on Intelligent
[50] S. Tu, A. Cyphert, and S. Perl, “Limits of using artificial intelligence Vehicles, 2023.
and gpt-3 in patent prosecution,” Tex. Tech L. Rev., vol. 54, p. 255, [76] H. Nori, N. King, S. M. McKinney, D. Carignan, and E. Horvitz,
2021. “Capabilities of gpt-4 on medical challenge problems,” arXiv preprint
[51] A. Mathew, “Is artificial intelligence a world changer? a case study of arXiv:2303.13375, 2023.
openai’s chat gpt,” Recent Progress in Science and Technology Vol. 5, [77] A. Tack and C. Piech, “The ai teacher test: Measuring the pedagogical
pp. 35–42, 2023. ability of blender and gpt-3 in educational dialogues,” arXiv preprint
[52] R. Reed, “The theology of gpt-2: Religion and artificial intelligence,” arXiv:2205.07540, 2022.
Religion Compass, vol. 15, no. 11, p. e12422, 2021. [78] A. Lecler, L. Duron, and P. Soyer, “Revolutionizing radiology with gpt-
[53] H. Akbar, M. Zubair, and M. S. Malik, “The security issues and based models: Current applications, future possibilities and limitations
challenges in cloud computing,” International Journal for Electronic of chatgpt,” Diagnostic and Interventional Imaging, 2023.
Crime Investigation, vol. 7, no. 1, pp. 13–32, 2023. [79] D. Baidoo-Anu and L. Owusu Ansah, “Education in the era of gen-
[54] K. D. Gupta et al., “A review of generative ai from historical perspec- erative artificial intelligence (ai): Understanding the potential benefits
tives,” 2023. of chatgpt in promoting teaching and learning,” Available at SSRN
[55] R. Sajja, Y. Sermet, D. Cwiertny, and I. Demir, “Platform-independent 4337484, 2023.
and curriculum-oriented intelligent assistant for higher education,” [80] A. O’Cain, B. D. Fedoruk, Z. Masri, R. Frost, and A. Alahmar, “A
arXiv preprint arXiv:2302.09294, 2023. system for the improvement of educational assessment using intelligent
[56] F. Etro, “The economic consequences of the diffusion of cloud comput- conversational agents,” Available at SSRN 4393234, 2023.
ing,” Dutta, Soumitra; Mia, Irene. The Global Information Technology [81] A. Alam, “Possibilities and challenges of compounding artificial intel-
Report, vol. 2010, 2009. ligence in india’s educational landscape,” Alam, A.(2020). Possibilities
[57] K. YAMAOKA, K. WATANABE, K. KISE, A. DENGEL, and S. ISHI- and Challenges of Compounding Artificial Intelligence in India’s
MARU, “Experience is the best teacher: Personalized vocabulary Educational Landscape. International Journal of Advanced Science and
building within the context of instagram posts and sentences from gpt- Technology, vol. 29, no. 5, pp. 5077–5094, 2020.
3,” 2022. [82] H. Chen, O. Engkvist, Y. Wang, M. Olivecrona, and T. Blaschke, “The
[58] R. Ressmeyer, S. Masling, and M. Liao, ““deep faking” political twitter rise of deep learning in drug discovery,” Drug discovery today, vol. 23,
using transfe r learning and gpt-2,” 2019. no. 6, pp. 1241–1250, 2018.
 38

[83] N. Pillai, A. Dasgupta, S. Sudaskorn, J. Fretland, and P. D. Mavroudis, [108] K. Goei, M. Hendriksen, M. de Rijke et al., “Tackling attribute fine-
“Machine-learning-guided early drug discovery of small molecules,” grainedness in cross-modal fashion search with multi-level features,”
Drug Discovery Today, 2022. in SIGIR 2021 Workshop on eCommerce. ACM, 2021.
[84] Y. Kim, J.-H. Kim, J. M. Lee, M. J. Jang, Y. J. Yum, S. Kim, U. Shin, [109] S. G. Bouschery, V. Blazevic, and F. T. Piller, “Augmenting human
Y.-M. Kim, H. J. Joo, and S. Song, “A pre-trained bert for korean innovation teams with artificial intelligence: Exploring transformer-
medical natural language processing,” Scientific Reports, vol. 12, no. 1, based language models,” Journal of Product Innovation Management,
pp. 1–10, 2022. vol. 40, no. 2, pp. 139–153, 2023.
[85] K. S. Kalyan, A. Rajasekharan, and S. Sangeetha, “Ammu: a survey [110] R. Bommasani, D. A. Hudson, E. Adeli, R. Altman, S. Arora, S. von
of transformer-based biomedical pretrained language models,” Journal Arx, M. S. Bernstein, J. Bohg, A. Bosselut, E. Brunskill et al., “On
of biomedical informatics, vol. 126, p. 103982, 2022. the opportunities and risks of foundation models,” arXiv preprint
[86] Z. Liu, R. A. Roberts, M. Lal-Nag, X. Chen, R. Huang, and W. Tong, arXiv:2108.07258, 2021.
“Ai-based language models powering drug discovery and develop- [111] J. Cowls, A. Tsamados, M. Taddeo, and L. Floridi, “The
ment,” Drug Discovery Today, vol. 26, no. 11, pp. 2593–2607, 2021. ai gambit: leveraging artificial intelligence to combat climate
[87] J. Vamathevan, D. Clark, P. Czodrowski, I. Dunham, E. Ferran, G. Lee, change—opportunities, challenges, and recommendations,” Ai & So-
B. Li, A. Madabhushi, P. Shah, M. Spitzer et al., “Applications of ciety, pp. 1–25, 2021.
machine learning in drug discovery and development,” Nature reviews [112] H. Benbya, T. H. Davenport, and S. Pachidi, “Artificial intelligence in
Drug discovery, vol. 18, no. 6, pp. 463–477, 2019. organizations: Current state and future opportunities,” MIS Quarterly
[88] E. Gawehn, J. A. Hiss, and G. Schneider, “Deep learning in drug Executive, vol. 19, no. 4, 2020.
discovery,” Molecular informatics, vol. 35, no. 1, pp. 3–14, 2016. [113] Y. Liu, T. Han, S. Ma, J. Zhang, Y. Yang, J. Tian, H. He, A. Li,
[89] A. Lavecchia, “Deep learning in drug discovery: opportunities, chal- M. He, Z. Liu et al., “Summary of chatgpt/gpt-4 research and per-
lenges and future prospects,” Drug discovery today, vol. 24, no. 10, spective towards the future of large language models,” arXiv preprint
pp. 2017–2032, 2019. arXiv:2304.01852, 2023.
[90] F. Urbina, F. Lentzos, C. Invernizzi, and S. Ekins, “Dual use of [114] S. Biswas, “Importance of chat gpt in agriculture: According to chat
artificial-intelligence-powered drug discovery,” Nature Machine Intel- gpt,” Available at SSRN 4405391, 2023.
ligence, vol. 4, no. 3, pp. 189–191, 2022. [115] M. S. Farooq, S. Riaz, A. Abid, K. Abid, and M. A. Naeem, “A
[91] M. H. Segler, T. Kogej, C. Tyrchan, and M. P. Waller, “Generating survey on the role of iot in agriculture for the implementation of smart
focused molecule libraries for drug discovery with recurrent neural farming,” Ieee Access, vol. 7, pp. 156 237–156 271, 2019.
networks,” ACS central science, vol. 4, no. 1, pp. 120–131, 2018. [116] H. Wang, H. Wu, H. Zhu, Y. Miao, Q. Wang, S. Qiao, H. Zhao,
[92] M. Ahsan, M. Rahaman, N. Anjum et al., “From chatgpt-3 to gpt-4: C. Chen, and J. Zhang, “A residual lstm and seq2seq neural network
A significant leap in ai-driven nlp tools,” Saidur and Anjum, Nishath, based on gpt for chinese rice-related question and answer system,”
From ChatGPT-3 to GPT-4: A Significant Leap in AI-Driven NLP Tools Agriculture, vol. 12, no. 6, p. 813, 2022.
(March 27, 2023), 2023. [117] Y. K. Dwivedi, N. Kshetri, L. Hughes, E. L. Slade, A. Jeyaraj, A. K.
[93] D. M. Levine, R. Tuwani, B. Kompa, A. Varma, S. G. Finlayson, Kar, A. M. Baabdullah, A. Koohang, V. Raghavan, M. Ahuja et al.,
A. Mehrotra, and A. Beam, “The diagnostic and triage accuracy of ““so what if chatgpt wrote it?” multidisciplinary perspectives on op-
the gpt-3 artificial intelligence model,” medRxiv, pp. 2023–01, 2023. portunities, challenges and implications of generative conversational ai
[94] I. Alghanmi, L. Espinosa-Anke, and S. Schockaert, “Probing pre- for research, practice and policy,” International Journal of Information
trained language models for disease knowledge,” arXiv preprint Management, vol. 71, p. 102642, 2023.
arXiv:2106.07285, 2021. [118] S. Biswas, “Prospective role of chat gpt in the military: According to
[95] H. Ali, “The potential of gpt-4 as a personalized virtual assistant for chatgpt,” Qeios, 2023.
bariatric surgery patients,” Obesity Surgery, pp. 1–1, 2023. [119] P. Helo and A. Shamsuzzoha, “Real-time supply chain—a blockchain
[96] D.-M. Vulturar, M. A. Neag, S. C. Vesa, A.-D. Maierean, D. Gherman, architecture for project deliveries,” Robotics and Computer-Integrated
A. D. Buzoianu, O. H. Orăsan, and D.-A. Todea, “Therapeutic efficacy Manufacturing, vol. 63, p. 101909, 2020.
and outcomes of remdesivir versus remdesivir with tocilizumab in se- [120] R. Kadel, H. Shrestha, A. Shrestha, P. Sharma, N. Shrestha, J. Bashyal,
vere sars-cov-2 infection,” International Journal of Molecular Sciences, and S. Shrestha, “Emergence of ai in cyber security,” International
vol. 23, no. 22, p. 14462, 2022. Research Journal of Modernization in Engineering Technology and
[97] B. Wang, Q. Xie, J. Pei, P. Tiwari, Z. Li et al., “Pre-trained language Science, 2022.
models in biomedical domain: A systematic survey,” arXiv preprint [121] H. Benbya, S. Pachidi, and S. Jarvenpaa, “Special issue editorial:
arXiv:2110.05006, 2021. Artificial intelligence in organizations: Implications for information
[98] S. Arslan, “Exploring the potential of chat gpt in personalized obesity systems research,” Journal of the Association for Information Systems,
treatment,” Annals of Biomedical Engineering, pp. 1–2, 2023. vol. 22, no. 2, p. 10, 2021.
[99] A. Blanchard and M. Taddeo, “Ethical challenges of using artifical [122] T. Zheng, M. Ardolino, A. Bacchetti, and M. Perona, “The applications
intelligence for intelligence analysis,” Available at SSRN 4226631, of industry 4.0 technologies in manufacturing context: a systematic lit-
2022. erature review,” International Journal of Production Research, vol. 59,
[100] B. Balsmeier and M. Woerter, “Is this time different? how digitalization no. 6, pp. 1922–1954, 2021.
influences job creation and destruction,” Research policy, vol. 48, no. 8, [123] B. Rathore, “Digital transformation 4.0: Integration of artificial in-
p. 103765, 2019. telligence & metaverse in marketing,” Eduzone: International Peer
[101] F. Pesapane, M. Codari, and F. Sardanelli, “Artificial intelligence Reviewed/Refereed Multidisciplinary Journal, vol. 12, no. 1, pp. 42–48,
in medical imaging: threat or opportunity? radiologists again at the 2023.
forefront of innovation in medicine,” European radiology experimental, [124] J. Bulchand-Gidumal, “Impact of artificial intelligence in travel,
vol. 2, pp. 1–10, 2018. tourism, and hospitality,” in Handbook of e-Tourism. Springer, 2022,
[102] I. Carvalho and S. Ivanov, “Chatgpt for tourism: applications, benefits pp. 1943–1962.
and risks,” Tourism Review, 2023. [125] N. Gillani, R. Eynon, C. Chiabaut, and K. Finkel, “Unpacking the
[103] R. S. Rathore, S. Sangwan, and O. Kaiwartya, “Towards trusted green “black box” of ai in education,” Educational Technology & Society,
computing for wireless sensor networks: Multi metric optimization vol. 26, no. 1, pp. 99–111, 2023.
approach.” Adhoc & Sensor Wireless Networks, vol. 49, 2021. [126] N. Mehrabi, F. Morstatter, N. Saxena, K. Lerman, and A. Galstyan,
[104] S. Shah, H. Ghomeshi, E. Vakaj, E. Cooper, and R. Mohammad, “An “A survey on bias and fairness in machine learning,” ACM Computing
ensemble-learning-based technique for bimodal sentiment analysis,” Surveys (CSUR), vol. 54, no. 6, pp. 1–35, 2021.
Big Data and Cognitive Computing, vol. 7, no. 2, p. 85, 2023. [127] F. T. Tschang and E. Almirall, “Artificial intelligence as augmenting
[105] J. Salminen, C. Kandpal, A. M. Kamel, S.-g. Jung, and B. J. Jansen, automation: Implications for employment,” Academy of Management
“Creating and detecting fake reviews of online products,” Journal of Perspectives, vol. 35, no. 4, pp. 642–659, 2021.
Retailing and Consumer Services, vol. 64, p. 102771, 2022. [128] V. Jain, B. Malviya, and S. Arya, “An overview of electronic commerce
[106] A. S. George and A. H. George, “A review of chatgpt ai’s impact on (e-commerce),” Journal of Contemporary Issues in Business and Gov-
several business sectors,” Partners Universal International Innovation ernment— Vol, vol. 27, no. 3, p. 666, 2021.
Journal, vol. 1, no. 1, pp. 9–23, 2023. [129] B. Feijoo and A. Garcı́a González, “Online shopping routines among
[107] A. El-Ansari and A. Beni-Hssane, “Sentiment analysis for personalized chilean children: level of expansion and main causes,” 2020.
chatbots in e-commerce applications,” Wireless Personal Communica-
tions, vol. 129, no. 3, pp. 1623–1644, 2023.
 39

[130] X. Zhang, Y. Jiang, Y. Shang, Z. Cheng, C. Zhang, X. Fan, Y. Xiao, and [152] A. J. Veal, “The concept of lifestyle: a review,” Leisure Studies, vol. 12,
B. Long, “Dsgpt: Domain-specific generative pre-training of transform- no. 4, pp. 233–252, 1993.
ers for text generation in e-commerce title and review summarization,” [153] M. Jensen, “Defining lifestyle,” Environmental sciences, vol. 4, no. 2,
in Proceedings of the 44th International ACM SIGIR Conference on pp. 63–73, 2007.
Research and Development in Information Retrieval, 2021, pp. 2146– [154] P. Contoyannis and A. M. Jones, “Socio-economic status, health and
2150. lifestyle,” Journal of health economics, vol. 23, no. 5, pp. 965–995,
[131] A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever et al., 2004.
“Language models are unsupervised multitask learners,” OpenAI blog, [155] M. J. Reeves and A. P. Rafferty, “Healthy lifestyle characteristics
vol. 1, no. 8, p. 9, 2019. among adults in the united states, 2000,” Archives of internal medicine,
[132] T. Eloundou, S. Manning, P. Mishkin, and D. Rock, “Gpts are gpts: vol. 165, no. 8, pp. 854–857, 2005.
An early look at the labor market impact potential of large language [156] Y. H. Yeo, J. S. Samaan, W. H. Ng, P.-S. Ting, H. Trivedi, A. Vipani,
models,” arXiv preprint arXiv:2303.10130, 2023. W. Ayoub, J. D. Yang, O. Liran, B. Spiegel et al., “Assessing the
[133] P. Maddigan and T. Susnjak, “Chat2vis: Generating data visualisations performance of chatgpt in answering questions regarding cirrhosis and
via natural language using chatgpt, codex and gpt-3 large language hepatocellular carcinoma,” medRxiv, pp. 2023–02, 2023.
models,” IEEE Access, 2023. [157] S. S. Biswas, “Role of chat gpt in public health,” Annals of Biomedical
[134] N. Jain, S. Vaidyanath, A. Iyer, N. Natarajan, S. Parthasarathy, S. Ra- Engineering, pp. 1–2, 2023.
jamani, and R. Sharma, “Jigsaw: Large language models meet program [158] M. Patkar, “5 Free Travel Planning AI and ChatGPT Apps to Get an
synthesis,” in Proceedings of the 44th International Conference on Instant Itinerary — makeuseof.com,” https://www.makeuseof.com/free-
Software Engineering, 2022, pp. 1219–1231. travel-planning-ai-chatgpt-apps/, [Accessed 24-Apr-2023].
[135] A. Haleem, M. Javaid, and R. P. Singh, “An era of chatgpt as a [159] “Your ai-powered personal chef.” [Online]. Available: https://www.
significant futuristic support tool: A study on features, abilities, and chefgpt.xyz/
challenges,” BenchCouncil transactions on benchmarks, standards and [160] lechjaLearnCrafts, “Learn Crafts & Hobbies w/ GPT Chat —
evaluations, vol. 2, no. 4, p. 100089, 2022. lechja.com,” https://www.lechja.com/ai/learn-crafts-hobbies-w-gpt-
[136] N. Brand, W. Odom, and S. Barnett, “Envisioning and understanding chat, 2023, [Accessed 24-Apr-2023].
orientations to introspective ai: Exploring a design space with meta. [161] “How to use chatgpt in your job search — indeed.com.” [On-
aware,” in Proceedings of the 2023 CHI Conference on Human Factors line]. Available: https://www.indeed.com/career-advice/news/chatgpt-
in Computing Systems, 2023, pp. 1–18. job-search
[137] N. Dehouche, “Plagiarism in the age of massive generative pre-trained [162] L. Floridi and M. Chiriatti, “Gpt-3: Its nature, scope, limits, and
transformers (gpt-3),” Ethics in Science and Environmental Politics, consequences,” Minds and Machines, vol. 30, pp. 681–694, 2020.
vol. 21, pp. 17–23, 2021. [163] S. Toshniwal, S. Wiseman, K. Livescu, and K. Gimpel, “Learning chess
[138] B. D. Lund, T. Wang, N. R. Mannuru, B. Nie, S. Shimray, and Z. Wang, blindfolded,” 2021.
“Chatgpt and a new academic reality: Artificial intelligence-written [164] J. Freiknecht and W. Effelsberg, “Procedural generation of interactive
research papers and the ethics of the large language models in scholarly stories using language models,” in Proceedings of the 15th Interna-
publishing,” Journal of the Association for Information Science and tional Conference on the Foundations of Digital Games, 2020, pp.
Technology, 2023. 1–8.
[139] M. Javaid, A. Haleem, and R. P. Singh, “Chatgpt for healthcare ser- [165] S. Värtinen, P. Hämäläinen, and C. Guckelsberger, “Generating role-
vices: An emerging stage for an innovative perspective,” BenchCouncil playing game quests with gpt language models,” IEEE Transactions on
Transactions on Benchmarks, Standards and Evaluations, p. 100105, Games, pp. 1–12, 2022.
2023. [166] J. van Stegeren and J. Myundefinedliwiec, “Fine-tuning gpt-2 on
[140] P. P. Ray, “Chatgpt: A comprehensive review on background, appli- annotated rpg quests for npc dialogue generation,” in Proceedings
cations, key challenges, bias, ethics, limitations and future scope,” of the 16th International Conference on the Foundations of Digital
Internet of Things and Cyber-Physical Systems, 2023. Games, ser. FDG ’21. New York, NY, USA: Association for
[141] E. G. Wilcox, J. Gauthier, J. Hu, P. Qian, and R. Levy, “On the Computing Machinery, 2021. [Online]. Available: https://doi.org/10.
predictive power of neural language models for human real-time 1145/3472538.3472595
comprehension behavior,” arXiv preprint arXiv:2006.01912, 2020. [167] P. Roetzer and M. Kaput, Marketing Artificial Intelligence: AI, Mar-
[142] R. Sridhar and D. Yang, “Explaining toxic text via knowledge enhanced keting, and the Future of Business. BenBella Books, 2022.
text generation,” in Proceedings of the 2022 Conference of the North [168] J. Thiergart, S. Huber, and T. Übellacker, “Understanding emails
American Chapter of the Association for Computational Linguistics: and drafting responses–an approach using gpt-3,” arXiv preprint
Human Language Technologies, 2022, pp. 811–826. arXiv:2102.03062, 2021.
[143] J. Bryant and P. Vorderer, Psychology of entertainment. Routledge, [169] X. Bai, L. Duan, R. Tang, G. Batra, and R. Agrawal, “Improving text-
2013. based similar product recommendation for dynamic product advertising
[144] H. H. Thorp, “Chatgpt is fun, but not an author,” pp. 313–313, 2023. at yahoo,” in Proceedings of the 31st ACM International Conference on
[145] S. Shahriar and K. Hayawi, “Let’s have a chat! a conversation with Information & Knowledge Management, ser. CIKM ’22. New York,
chatgpt: Technology, applications, and limitations,” arXiv preprint NY, USA: Association for Computing Machinery, 2022, p. 2883–2892.
arXiv:2302.13817, 2023. [Online]. Available: https://doi.org/10.1145/3511808.3557129
[146] M. U. Haque, I. Dharmadasa, Z. T. Sworna, R. N. Rajapakse, and [170] P. S. Neves, “Chat gpt ais “interview” 1, december 2022,” AIS-
H. Ahmad, “” i think this is the most disruptive technology”: Exploring Architecture Image Studies, vol. 3, no. 2, pp. 58–67, 2022.
sentiments of chatgpt early adopters using twitter data,” arXiv preprint [171] X.-R. Gong, J.-X. Jin, and T. Zhang, “Sentiment analysis using autore-
arXiv:2212.05856, 2022. gressive language modeling and broad learning system,” in 2019 IEEE
[147] O. AI. [Online]. Available: https://openai.com/product/gpt-4 International Conference on Bioinformatics and Biomedicine (BIBM).
[148] opchatsgpt, “Impact of chat gpt on the entertainment industry,” IEEE, 2019, pp. 1130–1134.
Mar 2023. [Online]. Available: https://opchatsgpt.com/chat-gpt-in- [172] A. H. Sweidan, N. El-Bendary, and H. Al-Feel, “Sentence-level aspect-
entertainment/ based sentiment analysis for classifying adverse drug reactions (adrs)
[149] “The Power of Chat GPT: Creating Personalized Marketing using hybrid ontology-xlnet transfer learning,” IEEE Access, vol. 9, pp.
and Efficient Customer Support. — linkedin.com,” https: 90 828–90 846, 2021.
//www.linkedin.com/pulse/power-chat-gpt-creating-personalized- [173] F. Wei and U. T. Nguyen, “Stock trend prediction using financial market
marketing-customer-bundhoo/, [Accessed 24-Apr-2023]. news and bert,” Wall Street Journal, 2018.
[150] “Chat GPT-4 in the Film Industry: Scriptwriting, Editing, and More; [174] T. Yue, D. Au, C. C. Au, and K. Y. Iu, “Democratizing financial knowl-
TS2 SPACE — ts2.space,” https://ts2.space/en/chat-gpt-4-in-the- edge with chatgpt by openai: Unleashing the power of technology,”
film-industry-scriptwriting-editing-and-more/#:∼:text=GPT\%2D4\ Available at SSRN 4346152, 2023.
%20has\%20the\%20potential,and\%20consistency\%20than\ [175] ”Siri”. [Accessed on 25.03.2023]. [Online]. Available: https://www.
%20ever\%20before., 2023, [Accessed 24-Apr-2023]. apple.com/in/siri/
[151] E. i. bd, “Chatgpt: The impact of chat gpt on the entertainment [176] ”Siri ChatGPT”. [Accessed on 25.03.2023]. [Online]. Available:
industry - march 24, 2023 educationsinbd,” Mar 2023. [Online]. https://support.apple.com/en-in/guide/shortcuts/apd07c25bb38/ios
Available: https://educationsinbd.com/the-impact-of-chat-gpt-on-the- [177] ”Siri ChatGPT”. [Accessed on 25.03.2023]. [Online]. Available:
entertainment-industry/ https://gpt3demo.com/apps/chatgpt-sirigpt-apple-ios
 40

[178] ”AI Dungeon: A text-based adventure-story game you direct (and 2021, Virtual Event, February 4–6, 2021, Revised Selected Papers.
star in) while the AI brings it to life.”. [Accessed on 30.03.2023]. Springer, 2022, pp. 236–249.
[Online]. Available: https://aidungeon.io/ [196] A. Chan, “Gpt-3 and instructgpt: technological dystopianism, utopi-
[179] ”It Began as an AI-Fueled Dungeon Game. It Got Much anism, and “contextual” perspectives in ai ethics and industry,” AI and
Darker”. [Accessed on 25.03.2023]. [Online]. Available: https: Ethics, pp. 1–12, 2022.
//www.wired.com/story/ai-fueled-dungeon-game-got-much-darker/ [197] M. Zhang and J. Li, “A commentary of gpt-3 in mit technology review
[180] ”Whatever you want to ask, our chat has the answers”. [Accessed on 2021,” Fundamental Research, vol. 1, no. 6, pp. 831–833, 2021.
25.03.2023]. [Online]. Available: https://www.copy.ai/ [198] H. R. Kirk, Y. Jun, F. Volpin, H. Iqbal, E. Benussi, F. Dreyer,
[181] ”Introducing, The BOND Network.”. [Accessed on 25.03.2023]. A. Shtedritski, and Y. Asano, “Bias out-of-the-box: An empirical
[Online]. Available: https://www.bond.ai/ analysis of intersectional occupational biases in popular generative
[182] ”Save hundreds of hours analyzing feedback.”. [Accessed on language models,” Advances in neural information processing systems,
29.03.2023]. [Online]. Available: https://www.askviable.com/ vol. 34, pp. 2611–2624, 2021.
[183] ”ai—channels: make contact with intelligent minds”. [Accessed on [199] S. S. Biswas, “Potential use of chat gpt in global warming,” Annals of
30.03.2023]. [Online]. Available: https://aichannels.app/ biomedical engineering, pp. 1–2, 2023.
[184] ”Automate your meeting notes”. [Accessed on 30.03.2023]. [Online]. [200] P. H. Seo, A. Nagrani, A. Arnab, and C. Schmid, “End-to-end genera-
Available: https://fireflies.ai/ tive pretraining for multimodal video captioning,” in Proceedings of the
[185] A. Papangelis, M. Namazifar, C. Khatri, Y.-C. Wang, P. Molino, and IEEE/CVF Conference on Computer Vision and Pattern Recognition,
G. Tur, “Plato dialogue system: A flexible conversational ai research 2022, pp. 17 959–17 968.
platform,” arXiv preprint arXiv:2001.06463, 2020. [201] W. Ji, Y. Wei, Z. Zheng, H. Fei, and T.-s. Chua, “Deep multimodal
[186] P. Dhariwal, H. Jun, C. Payne, J. W. Kim, A. Radford, and learning for information retrieval,” in ACM International Conference
I. Sutskever, “Jukebox: A generative model for music,” arXiv preprint on Multimedia, 2023.
arXiv:2005.00341, 2020. [202] V. Liu, H. Qiao, and L. Chilton, “Opal: Multimodal image generation
[187] D. Adiwardana, M.-T. Luong, D. R. So, J. Hall, N. Fiedel, R. Thop- for news illustration,” in Proceedings of the 35th Annual ACM Sympo-
pilan, Z. Yang, A. Kulshreshtha, G. Nemade, Y. Lu et al., “Towards sium on User Interface Software and Technology, 2022, pp. 1–17.
a human-like open-domain chatbot,” arXiv preprint arXiv:2001.09977, [203] C. Guo, A. Sablayrolles, H. Jégou, and D. Kiela, “Gradient-
2020. based adversarial attacks against text transformers,” arXiv preprint
[188] M. M. van Buchem, H. Boosman, M. P. Bauer, I. M. Kant, S. A. arXiv:2104.13733, 2021.
Cammel, and E. W. Steyerberg, “The digital scribe in clinical practice: [204] A. Shafahi, M. Najibi, A. Ghiasi, Z. Xu, J. P. Dickerson, C. Studer,
a scoping review and research agenda,” NPJ digital medicine, vol. 4, L. S. Davis, G. Taylor, and T. Goldstein, “Adversarial training
no. 1, p. 57, 2021. for free!” CoRR, vol. abs/1904.12843, 2019. [Online]. Available:
[189] K. Hao, “Facebook’s new polyglot ai can translate between 100 http://arxiv.org/abs/1904.12843
languages,” 2020. [205] T. Bai, J. Luo, J. Zhao, B. Wen, and Q. Wang, “Recent advances
[190] Y. Gan, G. Lu, Z. Su, L. Wang, J. Zhou, J. Jiang, and D. Chen, “A in adversarial training for adversarial robustness,” CoRR, vol.
joint domain-specific pre-training method based on data enhancement,” abs/2102.01356, 2021. [Online]. Available: https://arxiv.org/abs/2102.
Applied Sciences, vol. 13, no. 7, p. 4115, 2023. 01356
[191] “Chatgpt plugins,” Mar 2023. [Online]. Available: https://openai.com/ [206] M. Verma, “Integration of ai-based chatbot(chatgpt) and supply chain
blog/chatgpt-plugins management solution to enhance tracking and queries response,” 02
[192] B. Bhattarai, O.-C. Granmo, and L. Jiao, “Convtexttm: An explainable 2023.
convolutional tsetlin machine framework for text classification,” in [207] Y. Kubo and T. Trappenberg, Mitigating Overfitting Using Regulariza-
Proceedings of the Thirteenth Language Resources and Evaluation tion to Defend Networks Against Adversarial Examples, 04 2019, pp.
Conference, 2022, pp. 3761–3770. 400–405.
[193] A. Narasimhan, K. P. A. V. Rao et al., “Cgems: A metric model for au- [208] X. Liu, Y. Zheng, Z. Du, M. Ding, Y. Qian, Z. Yang, and J. Tang, “Gpt
tomatic code generation using gpt-3,” arXiv preprint arXiv:2108.10168, understands, too,” arXiv preprint arXiv:2103.10385, 2021.
2021. [209] J. Jia, H. Liu, and N. Z. Gong, “10 security and privacy problems in
[194] J. E. Zini and M. Awad, “On the explainability of natural language self-supervised learning,” arXiv preprint arXiv:2110.15444, 2021.
processing deep models,” ACM Computing Surveys, vol. 55, no. 5, pp. [210] M. Abadi, A. Chu, I. Goodfellow, H. B. McMahan, I. Mironov,
1–31, 2022. K. Talwar, and L. Zhang, “Deep learning with differential privacy,”
[195] R. K. Yadav, L. Jiao, O.-C. Granmo, and M. Goodwin, “An inter- in Proceedings of the 2016 ACM SIGSAC conference on computer and
pretable word sense classifier for human explainable chatbot,” in Agents communications security, 2016, pp. 308–318.
and Artificial Intelligence: 13th International Conference, ICAART