Addressing the Ethical Implications of AI Models Developed:

A Case Study of Master’s Degree Dissertations in Data Science for

Industry and Society

Alina Delia C

alin

Faculty of Mathematics and Computer Science, Babes

-Bolyai University, Kog

alniceanu Street, Cluj-Napoca, Romania

Keywords:

Ethical AI, Data Science, Responsible AI, Bias, Transparency, Accountability, Fairness, Explainable AI,

Safety, Data Privacy, Machine Learning, Artiﬁcial Intelligence.

Abstract:

The increase in the development and use of AI models has generated many ethical and societal concerns. In

this paper, we examine the ethical element in several dissertations presented in July and September 2024 by

students enroled in the Data Science for Industry and Society Master’s Degree Programme. We assess the

level of awareness of ethical principles by analysing in these case studies the ethical concerns addressed by

most students, the ethical principles that are mostly neglected, and possible implications for the society. The

ﬁndings reveal that data bias is the most addressed concern, while accountability is the most neglected ethical

principle. Some recommendations for possible improvements include the use of ethical AI tools for the design

and assessment of AI models and applications.

1 INTRODUCTION

The past decade has known an exponential growth of

Artiﬁcial Intelligence (AI) based applications. Many

of these, for example public Large Language Models

(LLM), introduce disruptive technologies that chal-

lenge and modify many aspects of life. The activi-

ties involved include teaching and learning, produc-

ing written content such as essays, reports, or news,

and constitute a huge knowledge and decision base

for many users. Despite many disclaimers from the

manufacturers of these tools explaining that erroneous

content might be generated, many users are trustingly

using them in professional and learning contexts. De-

spite being scrutinised, the use of these models in ed-

ucation to produce writing that is creative, cohesive,

and pertinent is deeply explored, emphasising their

role in improving the quality and efﬁciency of student

work (Kenwright, 2024).

More often than not, for many such applications,

the level of accountability demonstrated by the devel-

oper is often neglected, with the result that the user

has the primary responsibility to use the AI (Vakkuri,

2022). Caution and reluctance are often advised be-

cause the ethical and safety implications are not made

known to the user, who is at high risk. Rather than

spending too much controversial time on whether we

should use and trust these systems or not we might fo-

https://orcid.org/0000-0001-7363-4934

cus more productively on developing ethical models.

Promoting transparency, agreeing on safety and inclu-

sive guidelines to be enforced, could ﬁnally make this

technology serve actual societal needs and improve

life quality (Remian, 2019).

The correct identiﬁcation of ethical concerns is the

ﬁrst step in addressing them. For example, some re-

searchers apparently argue that unemployment as a

result of the automation power of AI models is one

of the main ethical implications (Wiesenthal, 2022).

This calls for a shift in mindset and culture, to em-

ploy trustworthy mitigation solutions. Even if these

issues are identiﬁed and discussed, few studies offer

recommendations on how to deal with them when de-

signing and developing AI systems (Huriye, 2023).

Moreover, there isn’t always agreement among schol-

ars as to what constitutes best practices for ethical AI

models (Jobin et al., 2019).

The quality of AI is determined by the availabil-

ity and quality of data, and it is where the majority of

problems originate. It is critically important to ensure

that AI systems are trained on a variety of representa-

tive datasets and that bias is rigorously tested in them

(Konidena et al., 2024). Moreover, it has been demon-

strated that AI attacks that utilise gradient-based en-

hancements are possible (Rosenblatt et al., 2023) re-

sulting in falsely highly accurate AI models.

Real-world case studies offer valuable insight into

the successful integration of ethical considerations in

alin, A. D.

Addressing the Ethical Implications of AI Models Developed: A Case Study of Master’s Degree Dissertations in Data Science for Industry and Society.

DOI: 10.5220/0013112000003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 1, pages 279-286

ISBN: 978-989-758-737-5; ISSN: 2184-433X

279

AI and Data Science (Tatineni, 2019). Examples in-

clude proactive bias mitigation, transparent decision

making, and community participation. These strate-

gies ensure fair outcomes for diverse user groups, fos-

ter trust, and align ethical considerations with soci-

etal values, thereby fostering a more inclusive and

ethical AI landscape. Some researchers study how

engineering requirements can help businesses han-

dle ethical concerns with AI technologies (Balasub-

ramaniam, 2019), by examining three Finnish ﬁrms’

AI ethics guidelines using a multiple-case research

methodology. Accountability, justice, privacy, safety,

security, transparency, and trust were the main topics

of the recommendations. Speciﬁc suggestions are cre-

ating multidisciplinary development teams to address

divergent ethical perspectives and prioritising quality

requirements by using AI ethical norms.

The current literature has recognised the follow-

ing major categories of ethical issues: privacy, auton-

omy, anonymity, transparency, security, safety, jus-

tice, and dignity (Tatineni, 2019). These should nor-

mally come on top of software engineering quality

requirements of usability, performance, reliability, se-

curity, safety, maintainability, accuracy, interoperabil-

ity, and reusability (Balasubramaniam, 2019).

The ethical considerations in AI and Data Science

development are crucial for responsible use. Strate-

gies include conducting ethical impact assessments,

developing agile ethical frameworks, and fostering

collaboration between researchers, ethicists, policy-

makers, and industry experts. By understanding the

implications of advanced AI, anticipating future chal-

lenges, and implementing proactive measures, we

can contribute to responsible and beneﬁcial technol-

ogy development. Ethical considerations in AI and

Data Science development are crucial for responsible

use, requiring impact assessments, agile frameworks,

and collaboration among researchers, ethicists, poli-

cymakers, and industry experts.

The aim of this paper is to analyse the existing

level of knowledge and awareness of ethical AI im-

plications among master’s degree students from a spe-

cialised Data Science domain. Several thesis are be-

ing assessed on 4 selected ethical criteria: (1) fair-

ness and bias, (2) safety and security, (3) account-

ability and liability, and (4) transparency and explain-

ability. The analysis will focus on the following re-

search questions (RQ): RQ1: Which ethical principles

for developing AI-based applications are most often

addressed by students?; RQ2: Which ethical princi-

ples for developing AI-based applications are gener-

ally neglected by students?; RQ3: What solutions for

the implementation of ethical AI principles from the

literature could be applied in these case studies?.

2 MATERIALS AND METHODS

2.1 Methodology

In this study, we evaluate the ethical element in sev-

eral dissertation theses presented in July and Septem-

ber 2024 by students enroled in the Data Science

for Industry and Society Master’s Degree Programme

of the Faculty of Mathematics and Computer Sci-

ence from Babes

-Bolyai University (UBB) in Cluj-

Napoca. The criteria used in the analysis of these

case studies are: (1) fairness and bias, (2) safety and

security, (3) accountability and liability, and (4) trans-

parency and explainability. All identiﬁed concerns are

then assessed as to how they are addressed in the mas-

ter thesis, if at all, by means of: discussion, evaluation

tools utilised, ethical design, and other approaches.

From a number of 12 student papers, we selected

those that involve either the development of an AI

module or the development of an application that uses

pre-trained AI models. Papers that do not have as

scope a speciﬁc application addressed to a large pool

or users, but are focused more on research methods,

comparisons, or assessment of AI models have been

excluded. The focus of this study is to assess the

level of awareness and mitigation of ethical issues re-

lated to developing consumer-oriented applications,

and their impact in society.

The selected dissertations are being assessed con-

sidering the four criteria that are derived from both the

related literature, and from ethical AI norms and rec-

ommendations, by international organisations . Sev-

eral concerns speciﬁc to each AI model or application

are identiﬁed and evaluated if they are addressed in

the paper, in order to examine the ethical AI literacy

among these students and propose several recommen-

dations based on literature.

2.2 Ethical AI Principles

The criteria chosen for this study are based on related

literature (Horv

ath, 2022), (Konidena et al., 2024),

(Nguyen et al., 2023) and in line with the recommen-

dations provided by UNESCO’s ”Ethics of Artiﬁcial

Intelligence” (UNESCO, 2024). The 4 chosen crite-

ria are described in Figure 1, reﬂecting what we are

focusing on in the analysis of the thesis and identiﬁ-

cation of possible concerns.

2.3 Summaries of the Selected Papers

Case Study 1 (CS1): Speed Bump Detection.

(Complete Title: Robust Speed Bump Detection

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

280

Figure 1: The Ethical AI criteria used for assessment.

Based on Data Collected from GPS-Enabled Dash-

cams) (Muntean, 2024). Using predetermined GPS

locations, this dissertation proposes a novel way

for Advanced Driver Assistance Systems (ADAS) to

gather speed bump data. A labelled dataset is con-

structed with 2351 photos of plastic speed bumps that

are black and yellow, used to train a speed bump de-

tection AI model. The research shows that, while

maintaining a lower False Positive Rate (of 3.4%),

YOLOv8, a cutting-edge object identiﬁcation model,

can achieve performance levels comparable to the

state of the art.

Case Study 2 (CS2): Dental Issues Identiﬁca-

tion. (Complete Title: Advancements in Dental Care

through Deep Learning) (Moldovan, 2024). In or-

der to detect frequent dental problems (cavities, im-

plants, impacted teeth, and ﬁllings), this study exam-

ines pre-trained ResNet models and custom models

trained with TensorFlow Keras. YOLOv9 and other

YOLO models are tested for real-time processing on

mobile devices. The study showcases the superior

performance of the YOLOv9 model in dental X-ray

processing, demonstrating its potential to expedite pa-

tient sessions and enhance oral health insights with a

mobile application.

Case Study 3 (CS3): News Analysis. (Complete Ti-

tle: Beyond the Headlines: Leveraging AI to Stream-

line News Analysis) (Miclea, 2024). The project’s

goal is to create an intelligent news aggregation and

analysis system that uses artiﬁcial intelligence and

natural language processing to automate the task. The

system consists of a sentiment analysis module, data

processing engine, news retrieval module, and user in-

terface.

Case Study 4 (CS4): Anime Recommendation.

(Complete Title: Anime recommendation system)

arescu, 2024). This thesis compares the re-

sults of experiments conducted on a small real-world

dataset with the complete MyAnimeList database for

the purpose of designing an application to generate

anime recommendation. Three models were evalu-

ated: NCF,Wide & Deep, and SVD, with the best re-

sult for R

= 0.5473. The authors’ research is a basis

for creating an app for anime recommendations.

Case Study 5 (CS5): Stock Trend Prediction.

(Complete Title: Predicting Romanian stock move-

ment trends: a complex network approach combined

with machine learning) (Holgyes, 2024). This study

concentrates on predicting the Bucharest Stock Ex-

change (BSE) movements. It examines stock price

data using network science principles, deriving spe-

ciﬁc features from the network model (centrality mea-

sures and connectivity properties), in order to train

several regression algorithms. The top performance

was achieved using decision tree, which predicts the

following day’s price movements (rise, fall, or stag-

nate) with accuracy of 70% and precision of 50%.

Case Study 6 (CS6): Blood Cell Classiﬁcation.

(Complete Title: Blood cells classiﬁcation for Acute

Lymphoblastic Leukemia detection using federated

learning) (Chiorean, 2024). This research investi-

gates the application of federated learning for classi-

fying white blood cells to detect Acute Lymphoblas-

tic Leukemia (ALL). The paper uses Convolutional

Neural Networks (CNNs) and the advancements pro-

vided by MobileNet and ResNet architectures. The

model was able to improve performance up to 90%

for accuracy and 80% for precision. ResNet demon-

strates superior precision and reliability in diagnosing

ALL, while MobileNet excels in performance with a

smaller number of parameters.

Case Study 7 (CS7): Travel Recommendation.

(Complete Title: A Travel Recommendation System

Based on Weather Data and Traveller Proﬁle Using

Machine Learning Algorithms) (Zb

arcea, 2024). This

dissertation aims to assist users in choosing the ideal

Addressing the Ethical Implications of AI Models Developed: A Case Study of Master’s Degree Dissertations in Data Science for Industry

and Society

281

vacation spot tailored to their individual preferences

and requirements. It introduces a web-based platform

that provides customised recommendations, valuable

insights, and weather-related functionalities. The sys-

tem employs Long Short-Term Memory models for

predicting weather conditions in conjunction with a

deep learning model trained on a dataset of virtual

users with preferences and past travel experiences.

3 RESULTS AND DISCUSSION

A total of 83 ethical AI concerns have been identiﬁed

from these papers. One paper of the seven, namely

CS1, is distinguished as having addressed most of the

safety and security and some of the fairness and bias

related issues (9 issues out of 21), while the others

fail to address ethical concerns in most aspects, as de-

scribed in Figures 2 and 3. Also, CS1 has the most

ethical concerns addressed in total perhaps because

the application domain (automotive) is a popular ﬁeld

studied worldwide. The percentage of issues identi-

ﬁed per case study and criteria are presented in Fig-

ures 4 and 5, showing that the majority of concerns

are related to fairness and bias, and the fewest are re-

lated to accountability and liability. This is perhaps

because data are the ﬁrst source of bias in a model.

Although accountability is not less important, it is a

more straightforward issue of responsibility for the

technology and its use.

Figure 2: Number of addressed concerns per criteria for

each case study.

In Table 1 we present in detail several concerns

related to fairness and bias, as they were identiﬁed in

Figure 3: Number of concerns for each case study (yes -

addressed, no -not addressed).

Figure 4: Distribution of concerns per criteria .

Figure 5: Distribution of concerns per case study.

each of the 7 dissertations, and if they are addressed

in any way (either just identiﬁed, perhaps discussed,

or even mitigated) by the student. A similar analysis

for AI safety and security is performed in Table 2.

Table 3 presents the issues related to accountability

and liability of AI systems, while Table 4 is concerned

with transparency and explainability of the models.

3.1 Discussion

RQ1: Which Ethical Principles for Developing

AI-Based Applications Are Most Often Addressed

by Students?

The master’s degree study programme of the students

whose dissertations we analyse has only one course

dedicated to Ethics and Academic Integrity in Data

Science. The course teaches general legal and ethi-

cal issues of computer science, IP, software licensing

(incl. open source), risks and liabilities, privacy, in-

ternet and cyberspace, computer security (hacking).

There is also some content on ethical AI, such as data

access, use, and collection and ethical aspects of re-

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

282

Table 1: Fairness and bias principle related concerns.

Concerns Addr.

CS1: Speed bumps are less common,

datasets lack images with them

YES

Object detection uncommon aspect ratio. YES

Data collection limited to: Black-Yellow

Speed Bumps, one small area, one camera.

One single annotator bias. YES

The negative class unclearly deﬁned. No

”Superior dataset” unbased claim. No

CS2: Existing annotated dataset split on

train/validate/test with no description to

ensure relevant evaluation.

The classes might be imbalanced or unrep-

resentative for the context.

Model classiﬁes all input as ﬁllings be-

cause it is a predominant class.

YES

Dataset not representative for the entire

population (gender, age etc.)

CS3: Pre-trained tool may contain biases. No

News content might be unfairly interpreted

by sentiment analysis and clustering.

Unsupported statement ”impressive efﬁ-

cacy of AI-generated summaries”.

CS4: Dataset from MyAnimeList website,

with possible user biases (age, region).

Classes of anime genres are not balanced. YES

Some content is never recommended, lead-

ing to some artists never being promoted.

CS5: Training data is not representative

for trading ﬂuctuations (too short).

As a Regulated Market for stock trading,

this tool should be equally available to all.

CS6 : Data distribution on multiple de-

vices can introduce biases, if unbalanced.

Dataset has no patient demographics. No

One class presents fewer samples. YES

Images generated by other miscroscopic

devices might not be used in the model.

CS7: Model trained on synthetic data

might not be representative.

YES

Access to the app by introducing personal. No

Marketing user proﬁles are biased. No

Popular destinations might become the

only recommendations.

search. These aspects of ethical AI from this course

seem to be well addressed by the students in their

work. In terms of fairness and bias, most students ad-

dress the issues related to unrepresentative data due to

class unbalance, synthetic data, the presence of a pre-

dominant class or an under-represented class in the

dataset. This is speciﬁcally acknowledged in 5 of the

7 case studies and is the most predominant, if not the

only aspect related to the principle of fairness and bias

that is popular. Data annotation issues are also ob-

served, as they may introduce a signiﬁcant level of

subjective interpretation in the data labelling process

if not handled properly. From transparency and ex-

Table 2: Safety and security principle related concerns.

Concerns Addr.

CS1: Lack of speed bump standardisation

(system won’t work in a new region).

YES

Speed bump detection could be used to

softnes suspention rather than lower speed.

YES

A false positive can make the car unneces-

sarily brake, confusing other drivers.

YES

A false negative can lead to dangerous situ-

ations if driver is overreliant on the system.

YES

At what distance are speed bumps detected

(is there enough time to lower speed)?

Most false positives are observed on

drainage ditches, followed by crosswalks,

but also pole shadows or unpaved roads.

YES

Signiﬁcant false negatives and low recall

suggest speed bump is not detected in time,

leading to dangerous situations.

CS2: Misdiagnose exacerbates illness. No

System unable to diagnose less frequent is-

sues, as it’s trained only on 4 classes.

Diagnose does not replace consultation and

specialised treatment.

CS3: May increase anxiety, agression or

form unbalanced opinions on events.

May introduce false elements in summariz-

ing thus spreading misinformation.

CS4: Might promote aggressive content. No

Might promote inappropriate age-content. No

CS5: Economic decisions are still unpre-

dictable, even with a high model accuracy.

User might have ﬁnancial losses. No

Artiﬁcially create market destabilisation

due to uniform predictions and decisions.

CS6: When can a doctor make safe deci-

sions based on the system input?

A misdiagnosys can be dangerous on the

life of a pacient.

System requires data security for image

upload, diagnoses information.

CS7: System might recommend places

with sudden political conﬂict (dangerous ).

System might recommend places with sud-

den bad weather or natural calamities.

plainability, the commonly identiﬁed concern is over-

ﬁtting (in 2 of the 7 papers). In terms of safety and

security, only CS1 identiﬁes several cases of possible

misuse that can lead to dangerous situations.

As mentioned previously, one of the works (CS1)

has identiﬁed a higher number of concerns than oth-

ers. A possible explanation could be the domain of

the paper: automotive, a popular subject of interest

in industry, that has already attracted many ethical

discussions. Given the advanced technology in this

area and the high interest from both industry and re-

search, there are many norms and regulations already

in place, generated by ethical implications of self-

driving car. The effect is that it increases awareness

Addressing the Ethical Implications of AI Models Developed: A Case Study of Master’s Degree Dissertations in Data Science for Industry

and Society

283

Table 3: Accountability and liability related concerns.

Concerns Addr.

CS1: Software system has quality stan-

dards to be used in automotive indepen-

dently, and there are people accountable to

fulﬁl them (maintenance, updates etc.).

Responsible party for using the break. YES

Responsible party for a misclassiﬁcation is

named (software system quality).

CS2: Software system quality standards

assigned to speciﬁc people. The person re-

sponsible for a misdiagnosis is speciﬁed.

CS3: Software system has quality stan-

dards assigned to speciﬁc people. The per-

son responsible for app malfunction (fake

news, discriminating content) is clear.

CS4: The user is responsible for the con-

tent they to watch.

Responsibility for the promotion of biased

or aggressive or racist content and its reper-

cussions is assigned to a party.

CS5: The user is responsible for their own

trading activity and ﬁnancial losses.

Companies may be affected by the soft-

ware predictions and be sabotaged. The

responsibility in this case is clear and soft-

ware quality ensures no misconduct.

CS6: The system is informing the doc-

tor, there are people responsible when the

system is at fault (medical device).

Software system has quality standards and

people accountable to fulﬁl them, includ-

ing updates, model maintenance.

CS7: User is responsible for choosing a

place to travel.

The system usage limitations are clear. No

among the users, developers and researchers alike.

This is also perhaps because the safety implications

are more direct and more severe than for other top-

ics. Even so, a long-term approach and analysis of

ethical implications is necessary in all areas where AI

technology is to be frequently used.

In conclusion, most of the students are very famil-

iar with dataset concerns related to dataset containing

unbalanced classes. This is probably because they

study this aspect speciﬁcally in the master’s study

programme. Also, they are able to identify more

ethical concerns in the domains which are highly re-

searched and popular in industry.

RQ2: Which Ethical Principles for Developing

AI-Based Applications Are Generally Neglected

by Students?

Many ethical concerns are neglected by students in

their dissertations. In the area of fairness and bias,

there is no concern as to how representative is data

to the population (concerning age, ethnicity, gender,

etc.), what is the dataset used for the pre-trained mod-

Table 4: Transparency and explainability related concerns.

Concerns Addr.

CS1: Model is black box. No

Model is overﬁtting. No

The user understands the beneﬁts and lim-

itations of the system, how it works, and

how to make decisions based on it.

Human agency is deﬁned at design level. No

CS2: Trains a CNN model - black box. No

Model is overﬁtting. YES

Human agency is speciﬁed from the design

and clear to the user of the app.

Results for each class are presented. No

CS3: The accuracy of tools used is not

mentioned (and it often falls under 50%).

Uses black box models. No

It is clear to the user how the system selects

balanced and comprehensive set of news.

CS4: Black box unexplained. No

User is clear on what they need to do to get

an accurate recommendation.

CS5: The model is a combination of

network analysis/modelling and decision

trees, some decision making explained.

YES

Overﬁtting is suspected based on results. No

It is clear how the input features are used

in the prediction model.

CS6: Black box unexplained. No

Possible overﬁtting results. YES

User understands what images can be used. No

CS7: Black box unexplained. No

User is informed on proper app usage. No

els integrated and their biases, or if the data distri-

bution ensures sound evaluation. The resulting soci-

etal implications are drastic: some populations will

not get a correct dental diagnosis, summarised news

content might add more bias and confusion to the user

and inﬂuence political developments of a region, sev-

eral young anime artists will never get a chance to get

promoted to the public, the stock market can get even

more destabilised by speculating AI models, the use

of a different microscope for blood cell imaging might

get the wrong diagnosis, and several tourist destina-

tions might get overcrowded while others leave peo-

ple living on tourism without an income. Fairness is

hard to achieve even with the best interests at hand,

but if not considered at all it may play havoc on peo-

ples lives, their income, safety, or well-being.

The safety and security risks, as a consequence of

system misuse, are also neglected. This is not only

related to what an AI model fails to do, but also how

that can be exploited by malevolent parties, or simply

if the system is used in other ways than intended, and

what harm it can do. For example, a false positive

diagnosis for a leukemia patient can be equivalent to

a sentence to death, while a false diagnose of dental

cavity is unlikely to cause a person’s death, even if it

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

284

may damage a tooth. In this context, identifying the

severity of the risk of a 0.1% error is very important.

The most neglected ethical principle is account-

ability and liability, with only one concern addressed

of the 13 identiﬁed. In this regard, the tendency is

to leave the whole responsibility to the user of the AI

model. While in many cases the user has an impor-

tant responsibility of the use of a software system,

totally neglecting the developer’s responsibility is a

huge issue at the moment for the whole software en-

gineering world. If software systems and AI models

are supposed to be of help and support to humans, it

is crucial that they are reliable tools. Thus, develop-

ing systems that we can trust is a responsibility of the

developer to ensure the data used, training process,

and evaluation are properly performed according to

rigorous quality standards. Our AI models can gen-

erate very high accuracy, but they might be no good

if we can not rely on them, or at least guaranteed a

level of quality that makes them a reliable support in

certain scenarios. Otherwise, the risk is that our AI

models are being used as just toys, and not as a trust-

worthy decision support technology, always with high

caution, and as a result they could become more a hin-

drance than a human support.

Responsibility in ethical AI development usually

derives from transparency and explainability. Al-

though some transparency is addressed in the se-

lected case studies, all students completely neglect

the explainability principle in their work. All of the

projects involve black box models, meaning we have

no knowledge whatsoever on how the model makes

its decisions. In consequence, the model might go

wrong at any time, and we would not know why. Not

knowing the reasoning of an AI models means we can

not rely on its ’guesses’. As many explainability tools

have been developed lately, we should encourage their

use among computer science students of master’s de-

gree level, future software developers, and software

engineers designing the AI tools of the future.

RQ3: What Solutions for the Implementation of

Ethical AI Principles from the Literature Could

Be Applied in These Case Studies?

While many of the ethical AI concerns identiﬁed

deserve a thorough discussion and consideration

in a multidisciplinary AI, ethical and law context

(Konidena et al., 2024), there are several solutions

presented in the literature.

In Figure 6 we present a collection of some rep-

resentative tools that can be used for ethical AI de-

velopment. Some are inspired from literature (Cum-

ming et al., 2024), but also online resources such as

the Responsible AI Knowledge-base, Git repository

(Alexandra, 2023). However, many researchers con-

sider that true accountability come from transparency

(Jobin et al., 2019) and that these tools are not suf-

ﬁcient to identify subtle and detailed ethical AI con-

cerns (Bubinger and Dinneen, 2024), and we cannot

rely on libraries alone. Ongoing research can help

inform developers about which tools to use best for

their speciﬁc contexts (Jeyakumar et al., 2020), while

others militate for developing interpretable AI models

rather then trying (rather unsuccessfully) to explain

black-box models (Rudin, 2019).

Figure 6: Ethical AI tools.

AI tool development often prioritizes building

complex models without considering their industrial

ﬁt. Thus, design should start with the intended user

and use cases, before implementation, and involve

a multi-disciplinary team. Intelligent models should

focus on employing interpretable models where pos-

sible, specially for high risks decision making. Im-

plementing clear AI Ethics principles in curricula for

software quality study is of utmost importance.

3.2 Study Limitations

The number of case studies analysed in this paper is

only seven, and can not be sufﬁciently representative

for the general pool of computer science students. Re-

searcher bias could also be present in the interpre-

tations and analysis of the ethical aspects of student

dissertations. The study does not aim to perform an

exhaustive analysis of ethical AI knowledge, but an

assessment used as a starting point for improving the

curricula of the master’s degree programme. Future

work would involve a proper at scale study, perhaps

using NLP or LLM - based analysis for automation.

4 CONCLUSIONS

AI applications are crucial for everyday life, but their

frequent use raises ethical concerns about data usage

and potential misuse. It is crucial for developers to

Addressing the Ethical Implications of AI Models Developed: A Case Study of Master’s Degree Dissertations in Data Science for Industry

and Society

285

build responsibility and understand the signiﬁcant im-

pact that the AI technology has on society.

This paper analysed 7 master’s degree disserta-

tions in order to assess the level of ethical AI princi-

ples addressed by the students, based on four criteria:

fairness and bias, safety and security, accountability

and liability, and transparency and explainability.

The analysis reveals that the most addressed eth-

ical AI concerns are those related to unbalanced

dataset. Explainability is not addressed at all, most

works presenting black-box models. The most ne-

glected ethical AI principles are those related to ac-

countability and liability, in which it is expected that

the user takes the whole responsibility. Only one of

the 7 papers in the study addresses the safety and se-

curity concerns of the system developed. Several ex-

isting tools for fairness, bias and explainability iden-

tiﬁed in literature and online resources are recom-

mended both as a support for identifying ethical con-

cerns as well as for mitigation. Other strong recom-

mendations are developing interpretable models and

the introduction of ethical AI principles in the curric-

ula of computer science master’s degree programme.

ACKNOWLEDGEMENTS

The publication of this article was supported by the

2024 Development Fund of the UBB.

REFERENCES

Alexandra, I. (2023). Responsible ai knowledge-base.

Balasubramaniam, N. (2019). Using ethical guidelines for

deﬁning critical quality requirements of ai solutions.

Master’s thesis, Aalto University.

Bubinger, H. and Dinneen, J. D. (2024). “what could go

wrong?”: An evaluation of ethical foresight analysis

as a tool to identify problems of ai in libraries. The

Journal of Academic Librarianship, 50(5):102943.

Chiorean, M.-A. (2024). Blood cells classiﬁcation for

acute lymphoblastic leukemia detection using feder-

ated learning. Master’s thesis, UBB.

Cumming, D., Saurabh, K., Rani, N., and Upadhyay, P.

(2024). Towards ai ethics-led sustainability frame-

works and toolkits: Review and research agenda.

Journal of Sustainable Finance and Accounting,

1:100003.

Holgyes, O.-D. (2024). Predicting romanian stock move-

ment trends: a complex network approach combined

with machine learning. Master’s thesis, UBB.

Horv

ath, I. (2022). Ai in interpreting: Ethical considera-

tions. Across Languages and Cultures, 23(1):1–13.

Huriye, A. Z. (2023). The ethics of artiﬁcial intelligence:

examining the ethical considerations surrounding the

development and use of ai. American Journal of Tech-

nology, 2(1):37–44.

Jeyakumar, J. V., Noor, J., Cheng, Y.-H., Garcia, L., and

Srivastava, M. (2020). How can i explain this to you?

an empirical study of deep neural network explanation

methods. Advances in neural information processing

systems, 33:4211–4222.

Jobin, A., Ienca, M., and Vayena, E. (2019). The global

landscape of ai ethics guidelines. Nature machine in-

telligence, 1(9):389–399.

Kenwright, B. (2024). Is it the end of undergraduate disser-

tations?: Exploring the advantages and challenges of

generative ai models in education. In Generative AI in

teaching and learning, pages 46–65. IGI Global.

Konidena, B. K., Malaiyappan, J. N. A., and Tadimarri, A.

(2024). Ethical considerations in the development and

deployment of ai systems. European Journal of Tech-

nology, 8(2):41–53.

arescu, A.-D. (2024). Anime recommendation system.

Master’s thesis, UBB.

Miclea, D. (2024). Beyond the headlines: Leveraging ai to

streamline news analysis. Master’s thesis, UBB.

Moldovan, D.-M. (2024). Advancements in dental care

through deep learning. Master’s thesis, UBB.

Muntean, A.-O. (2024). Robust speed bump detection based

on data collected from gps-enabled dashcams. Mas-

ter’s thesis, UBB.

Nguyen, A., Ngo, H. N., Hong, Y., Dang, B., and Nguyen,

B.-P. T. (2023). Ethical principles for artiﬁcial intelli-

gence in education. Education and Information Tech-

nologies, 28(4):4221–4241.

Remian, D. (2019). Augmenting education: ethical consid-

erations for incorporating artiﬁcial intelligence in ed-

ucation. Master’s thesis, University of Massachusetts

at Boston.

Rosenblatt, M., Dadashkarimi, J., and Scheinost, D. (2023).

Gradient-based enhancement attacks in biomedical

machine learning. In Workshop on Clinical Image-

Based Procedures, pages 301–312. Springer.

Rudin, C. (2019). Stop explaining black box machine learn-

ing models for high stakes decisions and use inter-

pretable models instead. Nature machine intelligence,

1(5):206–215.

Tatineni, S. (2019). Ethical considerations in ai and data sci-

ence: Bias, fairness, and accountability. International

Journal of Information Technology and Management

Information Systems (IJITMIS), 10(1):11–21.

UNESCO (2024). Ethics of artiﬁcial intelligence.

Vakkuri, V. (2022). Implementing ai ethics in software de-

velopment. JYU dissertations.

Wiesenthal, M. (2022). The ethical implications of ai-based

mass surveillance tools. Master’s thesis, University of

Applied Sciences.

arcea, S.-O. (2024). A travel recommandation system

based on weather data and traveller proﬁle using ma-

chine learning algorithms. Master’s thesis, UBB.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

286