Generative AI: Challenges to Higher Education

Sencer Yeralan, PE, PhD

Professor of Philosophy
ADA University Baku, Azerbaijan

Laura Ancona Lee, PhD

Dean of Academic Internationalization
UNIP, São Paulo, Brazil

May 15, 2023

Synopsis

Generative Artificial Intelligence has rapidly expanded its footprint of use in educa-
tional institutions. It has been embraced by students, faculty, and staff alike. The
technology is capable of carrying out a sustained sequence of interactive dialogs and
creating reasonably meaningful text. Not surprisingly it seems to be routinely used
by faculty to generate questions and assignments, by students to submit assignments
and aid in self-learning, and administration to create manuals, memoranda, and pol-
icy documents. With its potential to lead to significant social innovation, tethering
on the verge of becoming a disruptive technology, it seems most unlikely that it
will fade away without being fully enfolded into almost all aspects of academic and
pedagogical activity. While it is early to predict the exact place of this technology
in education, we present thoughts to aid deliberations and give a brief review of the
opportunities and challenges.

1 Introduction

Artificial neural networks, by and large, form the basis of current artificial intelligence
models. They were first conceptualized in the mid-20th century with the idea of mim-
icking the structure of the human brain. The brain consists of many neurons connected
through synapses in a complex network. These neurons carry electrical charges, which
led to the hypothesis that a similar network could produce electrical activity that may re-
sult in some form of intelligence, including perception, thought processes, consciousness,
self-consciousness, and intelligence.
The early circuits consisted of multiple inputs and one output, similar to an analog com-
puter where input and output values were determined by voltages at each terminal. An
input signal representing its strength by its voltage is modified by an input function and
then brought to a summer 1 ([1]). The summer then adds up the voltages to produce
a single output. These functions could amplify or attenuate, in a linear or nonlinear
fashion, the input voltages. Once a single neuron unit was built, these neurons could be
connected to form a collection of inputs and outputs of more complex networks. Typi-
cally, there would be a collection of input neurons, and a collection of output neurons,
as given in Figure 1.
1
A simple summer circuit may be constructed with a a single operational amplifier
 Figure 1: A simple artificial neural networks.

To use the neural network, the specific inputs of the application would have to be
written as a vector of values. These values encode the features of the input. The output
could also be a collection of values. Take for example, an application to identify songs.
The signal from the microphone could be examined at 100 millisecond intervals. The
frequency at each epoch could the be taken as the inputs. For a 20 second song segment
there would be 200 inputs. The output would be one of the songs that the application
could identify.
In most applications, the input is a feature set rather than raw data. In the example
given, instead of frequencies, we could first extract the chords, and then sample the input
as a chord progression. Then, the sampling could be done using longer time intervals,
say seconds. A 200-input neural network could then take as an input a song segment of
200 seconds. This application could then identify the song being played, since each song
would be expected to have a unique cord progression over time.
In this example, input vector then gives the specific frequency (or chord) at each time
interval. The output is one of many known songs. The circuit identifies which song is
being played.
The prerequisite for the neural network to properly function is the determination of the
input functions. The process of determining these functions is similar to curve fitting
using the least-squares method ([2]. Some parameters have to be estimated so that
they best relate the output to the input. Interest in neural networks dwindled when it
was discovered that a single neural circuit, called the perceptron, could not mimic the
behavior of an XOR gate, while the implementation of AND gates and OR gates was
rather straightforward. However, interest was rekindled when it was recognized that
XOR behavior or other types of behavior that were not possible with simple networks

Yeralan and Lee, July 2023 Page 2 of 9

could be implemented provided that additional hidden layers were inserted between the
input layers and output layers. Neural networks that employ many hidden layers is
referred to as deep neural networks.

Figure 2: An artificial neural network with a hidden layer.

Determining the specific values and forms of the input activation functions is a daunting
task, but it can be done through an iterative process where inputs and outputs are
presented to the new network. A mathematical search process is then implemented,
where the values of activation function parameters that best relate the given outputs to
the inputs are sought. This process is called training the neural network, and as said,
conceptually it is a curve-fitting exercise.
The logic of neural networks is straightforward. No neuron has any intelligence. Instead,
it is assumed that intelligence is an emergent behavior that occurs when a large number of
neurons interact. This approach acknowledges that each perceptron is not a component
of intelligence but rather intelligence is emergent. There are good thought experiments
such as the Game from the 1960s ([3]) and the Chinese Room from the 1980s ([4]).

2 Natural Language Processing

Natural Language Processing (NLP) is implemented in a relatively large neural network

with a sufficient number of hidden layers. Until recently, it was not known whether such
a network could mimic a person speaking a natural language, or if one could convincingly
interact with the system using natural language inputs and outputs ([5]).
Training the new network for NLP, especially the Large Language Models (LLM) that
are now commonly used, requires a lot of computing power, quite a large neural network,
and many hidden layers. However, the general rather simplistic logic remains the same.
OpenAI’s product, ChatGPT2 , was an ambitious project that required building a LLM
and training it with the help of many human trainers. GPT stands for Generative Pre-
trained Transformer. Essentially the neural network looks at the words and picks the
best words to follow a given incomplete line. This is an oversimplification, as there are
many different aspects to this process. For example, there may be alternative endings
to a given initial sub-sentence, so a determination is needed as to how to generate the
remainder of the sentence. Similarly, the system checks for each word in a sentence
2
See openai.com

Yeralan and Lee, July 2023 Page 3 of 9

to see if it is consistent with the statement taken as a whole. Nonetheless, the system
picks words and phrases and creates a text output. The neural network possesses the
rules, just like other grammar checkers, and possesses the capability, just like other spell
checkers, to make the output grammatically consistent with proper punctuation.
When the system is run, it was noticed that the speed is sufficient for it to be perceived
as an almost-human companion. One can tell interacting with ChatGPT that it is not
a human, thereby not passing the Turing test fully. This, however does not prevent it
from proving to be a useful tool. As such, it behaves like a co-pilot or a companion that
can take on some of the more tedious tasks in creating text.
Another interesting and useful feature that makes ChatGPT preferable to search engines
is that it uses the so-called Open Session and incorporates past conversations into its
generation of new text. Therefore, one should not use ChatGPT just to submit a query
and look at the output, as one would normally do with a search engine. Rather, one
could use ChatGPT in an iterative manner, asking for something, submitting a query,
looking at the output, and refining the query. For example, one may ask ChatGPT to
write a chord progression for a 12-bar blues song and later refine it by asking to make the
chords jazzier or to transpose the progression to another key. Such interactive episodic
search is perceived by humans as communicating with another human being, enhancing
the appearance that ChatGPT is a co-worker or a co-pilot.
The fact that ChatGPT is trained using material on the web makes it somewhat of
an alternative to a keyword-based search engine. ChatGPT should be considered as a
versatile engine. There are many possible ways that it could be incorporated into an
online experience. For example, one may have three open tabs while searching for a
given topic. Then, one may ask a ChatGPT plugin of their browser to summarize the
information on these three tabs and present it in text form. This is a task that could be
done by a co-worker, so clearly, it improves the search experience and simplifies going
back and forth between tabs to manually collect information.

3 Uses of AI in Higher Education

The use of AI in higher education is not a new phenomena ([6], [7], [8]). It has raised
some concerns, predominantly as a fear that students may take advantage of AI-based
tools to complete their academic assignments with little or no genuine effort. Professors
may also use such tools to generate multiple-choice quizzes or presentations for their
classes.
While AI-based tools can be useful for brainstorming and generating initial drafts, there
is concern about the potential for academic dishonesty. Plagiarism checkers have be-
come more sophisticated in detecting content that has been generated by AI systems.
However, it is still possible to submit AI generated the output as genuine work. As
usual, technology to create and to detect advance concurrently, each feeding into the
development of the other.
Despite these concerns, Generative AI (GAI) can be a valuable tool to enhance the ef-
ficiency of the writing process. One can ask GAI to provide alternative responses and
several versions of an initial draft. This may enhance and simplify the initial brain-
storming phase and enrich the thinking process. Therefore, the idea that rather than
banning GAI outright, universities should consider ways to regulate their use has merit.
The use of any tool may be deemed “good” or “bad”, “honest” or “dishonest” relative
to the circumstances and individuals. Could it be that the first cutting objects seen in

Yeralan and Lee, July 2023 Page 4 of 9

archaeological museums were used only for hunting and never for wars or murders? Does
dishonesty imply ignorance? The ethical use of learning tools benefits those who want
to learn. The discussion about the holes in the systems that allow deceit and dishonesty
separates the brightest students and teachers from others. Those who understand the
value of education and use it as a tool for learning, and those who attempt to deceive
themselves and others.

4 Deliberations and Thought Experiments

How GAI may be incorporated into higher education is not a simple proposition. It
requires substantial debate. It also requires a systems view that re-evaluates the purpose
and role of higher education. The task is hindered by the fact that most university
administrators are ill-equipped to fully comprehend and appreciate the significance of
these new technologies. This is especially worrisome at neo-liberal corporate universities
([9]).
We now present a few thoughts and point out historical parallels from a wide range of
domains in hopes of promoting the much-needed debate on this topic.

4.1 The Sand Heap Paradox

The Sand Heap Paradox raises concerns about the use of GAI language models in
academia and higher education. The paradox asks, ”When does a sand heap become
a non-heap as you remove grains of sand one by one?” Clearly, when there is only one
grain of sand left, it is no longer a heap. However, in the beginning, it is a heap. So,
when exactly does that transition occur from a heap to something else?
Similarly, when does the use of GAI in a word processor become academically unaccept-
able? Word processors already have spell-checking tools.
Spell-checking was once done manually requiring some mental effort. There is very little
concern about the use of built-in spelling tools, dictionaries, thesauruses, etc. Nowadays,
word processors also look at grammar and suggest improvements, and there does not
seem to be much academic concern about using grammar checking word processors. On
the contrary, most professors would encourage the use of such tools, as would industry.
Word processors are also good at checking for punctuation mistakes and suggesting
corrections or improvements. But when a word processor suggests rephrasing using a
GAI, it often becomes subject to banning.
However, it is important to note that the capabilities of language processors vary widely,
from no help at all (like a typewriter) to a capable GAI-driven word processor. Therefore,
it becomes difficult, almost paradoxical, to draw a line beyond which the use of GAI
becomes academically unacceptable.

4.2 The Arts and Crafts Movement

The Arts and Crafts movement of the mid to late 1800s emerged after the Industrial
Revolution ([10]). With the advent of industrialization, it became possible to manufac-
ture goods at a much lower cost than before. However, these products lacked some of
the artistic elements that were present in their handcrafted predecessors. As a result,
artisans and crafts people, who were adversely affected by industrialization, initiated

Yeralan and Lee, July 2023 Page 5 of 9

a movement that claimed that industrialization was harming the artistic and aesthetic
aspects of products. This movement later gave way to Art Nouveau and Art Deco.
In the past, having an industrial product rather than a handcrafted one may have been
frowned upon. However, today, this concern seems to have all but disappeared. People
are happy to wear clothing with the insignia of the manufacturer, and all such clothing,
whether it’s footwear, jackets, or trousers, comes from an industrial process. This does
not seem to bother consumers, rather, people with brand loyalty proudly display such
products. Society has gone from considering machine-crafted products inferior to finding
them not only acceptable but also preferable.
Could a parallel be drawn here to the possibility of people in the future considering
handwritten memoranda or essays as inferior to those that are artificially created by
GAI? Just as we no longer know how to sew our clothes, might we reach a point where
not knowing how to write our own essays will not bother us? Here, we are not questioning
whether this would be a good thing or a bad thing. We simply pose the possibility of
such a future à la the Brave New World ([11]). Will there be a time when the use of GAI
is so widespread that human-written essays are no longer preferred? It’s an interesting
question that raises concerns about the role of technology in our lives.

4.3 Legitimate Uses of GAI

Given that there are legitimate uses for a companion, co-writer, or co-pilot, outright
banning GAI does not seem to be a good idea. Nor is it likely to be accepted by
those who would like to use its beneficial contributions to improve their efficiency and
effectiveness. There are several examples of how GAI could be used in a positive and
productive manner, such as quizzing oneself on a given subject to prepare for an exam
or asking GAI how best to prepare for a given assignment (see the Appendix).
GAI can be used in multiple ways. For example, if you are studying for an exam, you
can ask GAI to give you a quiz at your level. You can specify if it is an undergraduate or
graduate course and ask for ten multiple choice questions. You then input your answers
and GAI will then tell you which questions you answered correctly. From there, you
have a couple of options. You can ask for the answers to the questions you answered
incorrectly and study those subjects, or you may ask for suggestions on what subjects
to study to make up for those incorrect answers. After a few iterations of this process,
you will have a personalized and targeted study program, with GAI as a smart tutor
providing custom-tailored support.
This indirect second-order request demonstrates another leap in perceived intelligence for
GAI. Rather than asking it to do something directly, one can ask it to tell you how best
to ask GAI to do something. This removes the burden of understanding, articulating,
and comprehending a strategy to use the technology. Ultimately, the human will make
the decision as to how best to use the technology, but there is a difference between
starting from scratch and starting with a set of options or ideas.
In this case, one could argue that the benefit of GAI is to overcome the initial hurdle of
organizing initial thoughts and quickly arrive at a set of options that could be used as
a launching pad. While the option picked to go forward with may not be among those
initial options, starting with options simplifies the burden and allows one to quickly
focus on the task at hand.

Yeralan and Lee, July 2023 Page 6 of 9

4.4 A Thought Experiment

As the field of GAI continues to advance, it raises questions about the role of human
input in the text generation process. Just as we have progressed from programming
computers in machine code to writing code in high-level languages, could we also move
towards a similar evolution in text generation?
Consider the idea of a student preparing for an essay. Instead of spending time on
the actual text, could the student simply provide the key ideas and concepts to a GAI
platform and let the software compose the text itself? This approach would be similar to
writing code in a high-level language and allowing the compiler to generate the machine
code. Just as we assume the receiving party can compile and execute our code, we could
also assume that the GAI can compose the text based on our input.
While this approach may seem unconventional, it could be a way to streamline the
writing process and minimize the energy required to convey the message. Companies
now store standard “ChatGPT prompts” rather than email messages. Here, ChatGPT
acts as a macro to generate the desired text. The focus shifts from the actual text
to the underlying ideas and concepts. Ultimately, the text format could be seen as
an intermediate step, similar to machine code, that is not necessary for conveying the
unique ideas. This approach could be a promising avenue for exploring the capabilities
of GAI in education and beyond.
We now have computer science students who have never been exposed to machine code.
Is it possible that in the future, people without a drivers license ride alone in self-driving
cars? If so, how then could one criticize the generation of text by those who only know
how to seed or prompt GAI but do not know how to write in the traditional manner?

5 The Road Forward in Academia

The use of GAI in academia has become a hotly debated topic. Should we ban GAI or
should we embrace it and ask that students do nothing but generate their essays using
such tools? The answer is not a one-size-fits-all solution, and it is too early to make any
definitive rules regarding GAI in academia.
After all, after graduation, students will undoubtedly use GAI throughout their careers,
where GAI may be seamlessly integrated into many products. It may also come to pass
that society views GAI as something that is not only acceptable but preferred, as was
the case with the Arts and Crafts Movement. We need to be cautious in making any
specific decisions and procedures regarding GAI.
What remains is that academia should understand that it has a unique role in the
road forward. We academicians are responsible for producing the next generation of
citizenry and workforce. We need to ask people to familiarize themselves with AI and
truly understand the technology. It may be a good starting point for us to compose a
list of properties, concerns, benefits, and threats of GAI in an inclusive and transparent
manner. We need to have a comprehensive dialogue so that all different aspects of this
complicated and far-reaching issue could be addressed in the fullness of time.
Once all such aspects are understood and evaluated, we should attempt to come up
with fixed processes that determine how students and faculty may use GAI. Any early
attempt to ban the use or any quick solutions, such as relying on plagiarism checkers to
disallow GAI, is probably premature at this point. On the other hand, we may consider
teaching the best way to use GAI as a copilot, a co-worker, a companion, or a digital

Yeralan and Lee, July 2023 Page 7 of 9

twin. Such a practice may better prepare the student for the latter part of the 21st
century, which promises to be much different from the professional practices of today.
Ultimately, the road forward in academia regarding GAI should be guided by caution,
understanding, and a comprehensive dialogue. We need to prepare ourselves for the
potential benefits and threats of GAI. We should not rush to any premature decisions
or solutions. With proper preparation and understanding, we can ensure that the next
generation of citizenry and workforce is well-equipped to navigate the complex world
that will most likely include GAI.

References

[1] S. Yeralan, Operational Amplifiers through Experimentation and Reflection: An

Engineer’s Journey. IUS Press, 2023.
[2] J. H. Holland, Adaptation in Natural and Artificial Systems: An Introductory
Analysis with Applications to Biology, Control, and Artificial Intelligence. Bradford
Books, 1992, isbn: 978-0262581110.
[3] A. Dneprov, “Игра [The Game]”, Знание—Сила [Knowledge—Power], с. 39—41,
1961. url: http://q-bits.org/images/Dneprov.pdf.
[4] J. R. Searle, “Minds, brains, and programs”, Behavioral and brain sciences, т. 3,
с. 415—457, 1980.
[5] I. Solaiman, M. Brundage, J. Clark и др., Release Strategies and the Social Impacts
of Language Models, 2019. arXiv: 1908.09203 [cs.CL].
[6] S. Amershi, M. Cakmak, W. B. Knox и T. Kulesza, “Power to the People: The Role
of Humans in Interactive Machine Learning”, AI Magazine, т. 35, № 4, с. 105—120,
дек. 2014. doi: 10 . 1609 / aimag . v35i4 . 2513. url: https : / / ojs . aaai . org /
aimagazine/index.php/aimagazine/article/view/2513.
[7] J. Belda-Medina и J. R. Calvo-Ferrer, “Using Chatbots as AI Conversational Partners
in Language Learning”, Applied Sciences, т. 12, № 17, 2022, issn: 2076-3417. doi:
10.3390/app12178427. url: https://www.mdpi.com/2076-3417/12/17/8427.
[8] M. A. Kuhail, N. Alturki, S. Alramlawi и K. Alhejori, “Interacting with Educational
Chatbots: A systematic Review”, Applied Sciences, т. 28, № 17, 2023, issn: 1573-
7608. doi: 10.1007/s10639- 022- 11177- 3. url: https://doi.org/10.1007/
s10639-022-11177-3.
[9] A. Taylor, “Perspectives on the University as a Business: The Corporate Management
Structure, Neoliberalism and Higher Education”, Journal for Critical Education
Policy Studies, т. 15, № 1, с. 108—135, 2017.
[10] R. P. Blakesley, The Arts and Crafts Movement. Phaidon Press, 2006, isbn:
978-0-7148-3849-6.
[11] A. Huxley, Brave New World : A Novel. Chatto и Windus, 1932.

Yeralan and Lee, July 2023 Page 8 of 9

Appendix

This appendix illustrates a session with ChatGPT as a tutor.

Input query: How can you help me to prepare for my college course final exam?
ChatGPT:
ChatGPT can assist you in preparing for your college course final exam in various ways.
Here are a few examples:

1. Quiz Preparation: You can ask ChatGPT to generate practice quizzes for your
exam. You can specify the level of difficulty, the number of questions, and the
topic of the questions. You can take the quiz and get instant feedback on your
performance.

2. Study Recommendations: Based on your performance in the practice quizzes, you

can ask ChatGPT to recommend study materials and resources to help you improve
in the areas where you struggled. ChatGPT can suggest books, articles, videos, and
other resources that are relevant to your exam.

3. Concept Clarification: If there are specific concepts that you are struggling with,
you can ask ChatGPT to explain them to you in a clear and concise manner.
ChatGPT can provide you with definitions, examples, and analogies to help you
understand the concepts better.

4. Essay Writing: If your exam requires you to write an essay, you can ask ChatGPT
to help you brainstorm ideas, structure your essay, and provide feedback on your
writing. ChatGPT can suggest relevant sources, help you develop a thesis statement,
and provide tips on how to write a compelling essay.

Overall, ChatGPT can be a useful tool in your exam preparation process by providing
personalized and targeted assistance to help you achieve your academic goals.

Yeralan and Lee, July 2023 Page 9 of 9