Available via license: CC BY-NC 4.0
Content may be subject to copyright.
Exploring AI Ethics Syllabi Through NLP Cluster Analysis
Kerrie Hooper, Stephanie Lunn
Florida International University Miami, FL
khooper@fiu.edu, sjlunn@fiu.edu
Abstract
With new technology come new responsibilities. Ex-
amining artificial intelligence (AI) and its applications
through an ethical lens has become increasingly impor-
tant. Academia can play a critical role in shaping grad-
uates who may work in both the ethical and technical
spheres. To better understand how this may be inte-
grated into higher education institutions, we assessed
the content covered on AI ethics using Natural Lan-
guage Processing (NLP) syllabi. A total of 45 AI ethics
syllabuses made publicly available online were exam-
ined. Some important features captured from each syl-
labus were the course description, topics, department,
and year. We observed overarching patterns across the
AI ethics syllabi through supervised and unsupervised
clustering and Latent Dirichlet Allocation (LDA) anal-
ysis. Some of these included information across various
academic departments and the pre-post Chat-GPT era.
This study is insightful as it offers a baseline for inves-
tigating various AI ethics topics that are described in
academic departments, as well as uncovering potential
gaps in the contents of AI ethics syllabi.
Introduction
The emergence of new technologies brings about increased
responsibilities, such as the ethical implications of artifi-
cial intelligence (AI). Numerous professional organizations,
such as IEEE, have begun to develop AI ethics guidelines. In
addition, academia can be vital for fostering innovation and
producing individuals skilled in both technical and ethical
aspects [1]. Towards this goal, educators can prepare stu-
dents to be culturally responsive, have a collaborative mind-
set with interdisciplinary skills, and be aware of equity is-
sues and other social problems that plague society. A Stan-
ford University Cyberlaw article [2] suggests several actions
universities can take to support AI ethics: identifying key is-
sues, promoting an ethical vision, and implementing ethical
practices in teaching and research.
Examining AI ethics syllabi using natural language pro-
cessing (NLP) clustering techniques can offer insights into
the current state and areas for improvement in AI ethics ed-
ucation. This research sought to explore the contents of AI
Copyright © 2024 by the authors.
This open access article is published under the Creative Commons
Attribution-NonCommercial 4.0 International License.
ethics syllabi to infer patterns and trends in higher educa-
tion. By analyzing syllabus data, academic institutions can
enhance their curricula to better address ethical consider-
ations in AI development and deployment. Therefore, this
research aims to answer the following question: What can
be inferred when exploring the contents of AI ethics syllabi
using NLP clustering techniques?
Methodology
Data Collection
Data was gathered through a strategic online search for pub-
licly available syllabi and a few from the Tech Ethics Cur-
riculum spreadsheet by [3]. Course description, topics, de-
partment, year, and public/private university were some of
the important features that were collected and organized in
a CSV file, where each row was a syllabus, and the columns
were features and contents of the syllabus. Massive open on-
line courses (MOOCs) were excluded, as well as syllabi that
had missing data like department, university name, year, or
location. A total of 45 syllabi were collected and included in
this analysis.
Data Cleaning and Preprocessing
Regarding our important features, there were two cases
where course descriptions were missing. In those cases,
we used the course topics to replace the missing course
descriptions. After the dataset was cleaned, we used the
tweet-preprocessor package in Python, which allowed for
tokenization, parsing, and removal of URLs and reserved
words. The NLTK package was used to remove stop words.
After those steps, the text was ready for embedding. We
tested our data performance on sentence-transformer em-
beddings and term frequency-inverse document frequency
(TF-IDF). We found that sentence transformers performed
better with the clusters by having higher silhouette scores
and cluster distances. We also compared the performance
between course descriptions and topics. Topics showed
greater similarities, and therefore, topics were prioritized to
uncover uniqueness in syllabi in the analysis unless other-
wise specified.
Results
K-means Clustering was used to find any patterns in terms
of polarity among the syllabus topics. We tested values for
k from 2 to 10 using the silhouette score, inter-cluster dis-
tances, intra-cluster distances, and cluster size as the evalu-
ation metrics. We found that k = 2 was the best value for k,
which was therefore used for this analysis.
As shown in Figure reffig:topicClusters, these 15 words
are unique across the clusters. This means they do not re-
side in both clusters. Using k = 2, we found that the syllabi
contents were separated into those that are more general or
philosophical and those that are technical or specific ethical
concepts.
Figure 1: Topics of the two clusters
Supervised Clustering was conducted based on features
such as department, year, and pre-post ChatGPT era. As
shown in Figure 2, there is a topical vastness of AI ethics
as well as how it is distributed across departments offering
an AI ethics-related course. In addition, topics changed in
the post-ChatGPT era (See Figure 3).
Figure 2: Topical range and distribution across departments
offering an AI ethics-related course
Latent Dirichlet Allocation (LDA) Analysis was per-
formed on k = 2, given that it was the best k in the K-means
clustering. Python LDAvis library was used. The LDA anal-
ysis helped us see where the topics converged across all the
syllabi (see Figure 4). We saw that Data,Bias,Privacy, and
Fairness were the top converging topics across the syllabi,
suggesting them as a priority for AI ethics.
Discussion and Implications
This approach utilized NLP techniques to analyze patterns
across AI ethics syllabi, revealing that topics involving lists
Figure 3: The topics covered pre-post Chat-GPT era, where
the post-chat-GPT era is defined as 2023 and onwards
Figure 4: Topics across the syllabi
of words or phrases exhibited more distinctness than course
descriptions, which may contain more noise. K-means clus-
tering with k = 2 indicated polarity in concepts based on
sentence-transformer embeddings. Supervised clustering re-
vealed differences and patterns across departments and the
pre-post Chat-GPT era. LDA analysis showed privacy, bias,
fairness, and data as key issues across AI ethics syllabi,
aligning with global literature on AI ethics principles [4].
Other researchers can replicate this methodology to analyze
syllabi, aiding university professors and curriculum devel-
opers in identifying potential gaps and critical topics. The
study’s broader impact encompasses enhancing understand-
ing of the sociotechnical system of AI ethics and the inter-
disciplinary nature of the field, as well as bridging gaps be-
tween academia and industry through student learning.
Limitations
K-means clustering requires specifying K, making it vulner-
able to outliers and potentially skewing results. Analyzing
clusters can be challenging due to this skew. Similarly, LDA
analysis, while showing topic distribution across clusters,
can be difficult to interpret.
Conclusion and Future Direction
The approach highlights the utility of NLP clustering anal-
ysis in this domain, revealing distinct patterns in AI ethics
syllabi. K-means clustering demonstrates a polarization be-
tween general/philosophical and specific/technical topics.
Supervised clustering uncovers additional patterns, suggest-
ing implications for interdisciplinary collaboration and tech-
nological advancement. LDA topic modeling findings align
with globally converging critical areas of AI ethics.
For future work, other clustering techniques, such as ag-
glomerate clustering, can be applied to find more patterns
in the dataset. More insights can be derived through using
other NLP techniques, such as word similarities, to assess to
what extent syllabi cover various topics.
References
[1] S. S. Bhullar, V. K. Nangia, and A. Batish, “The impact
of academia-industry collaboration on core academic
activities: Assessing the latent dimensions,” Technolog-
ical Forecasting and Social Change, vol. 145, pp. 1–11,
2019.
[2] B. W. Smith, “An Academic Vision for AI Ethics —
cyberlaw.stanford.edu,” https://cyberlaw.stanford.edu/
blog/2023/04/academic-vision-ai-ethics, April 2023.
[3] C. Fiesler, “Tech ethics curricula: A
collection of syllabi,” Jul 2022. [On-
line]. Available: https://cfiesler.medium.com/
tech-ethics-curricula-a-collection-of-syllabi-3eedfb76be18
[4] A. Jobin, M. Ienca, and E. Vayena, “The global land-
scape of ai ethics guidelines,” Nature machine intelli-
gence, vol. 1, no. 9, pp. 389–399, 2019.
