Artificial Intelligence (AI) has the potential to enhance the traditional educational approach through the use of E-Learning and Massive Open Online Courses (MOOCs). Indeed, the application of AI-based techniques to MOOCs opens to providing a wide range of high-quality courses to a large global audience, increasing the accessibility of education and improving the learning process's efectiveness. At the same time, the growing popularity of MOOCs highlights the need to carefully consider AI use and potential negative consequences, prioritizing ethical considerations when developing and implementing this kind of technology. In this paper, we describe the use of AI techniques, specifically Deep Learning (DL), discussing the advantages, problems, and ethical concerns associated with using generative models in education. We present a case study implemented at the University of Naples Federico II that utilizes Deep Fakes to generate MOOC lessons. Furthermore, we report details about the Human-Centred AI Master's Programme (HCAIM), which supports the legal, regulatory-compliant, and ethical adoption of AI. We also introduce the use of ChatGPT, an AI-based teaching support tool, and discuss its benefits and potential risks. Overall, this paper emphasizes the importance of considering ethical concerns when implementing AI in education and highlights the potential benefits and challenges associated with classical as well as AI-based MOOCs.
when developing and implementing generative technoloArtificial Intelligence (AI) has the potential to revolu- gies in the context of learning. Indeed, while tools such tionize education, particularly through online learning as ChatGPT may represent a new advance, they can also platforms like Massive Open Online Courses (MOOCs), pose ethical dificulties, such as perpetuating bias and which provide unlimited access to educational materials. inequality, privacy concerns, and the risk of fake news AI can enhance the quality, accessibility, and efective- or manipulated opinions. The cybersecurity industry ness of education by personalizing learning experiences, should also consider the consequences of a cyber attack facilitating automatic grading, and enabling intelligent on these models. It is important for designers and society tutoring systems. In particular, one of the significant as a whole to reflect on the ethical implications of these advantages of using AI in education is the ability to tai- tools and to develop ethically responsible and sustainable lor learning experiences to diferent user typologies. A models that take into account issues of equity, privacy, notable example is LessonAble, a new pipelined method- and accountability. This requires a multidisciplinary apology introduced in a recent study that leverages the proach and a thorough assessment of social and ethical concept of deep fakes to generate MOOC visual content impacts. directly from a lesson script. The goal is to create a tool The Human-Centred AI Master’s Programme (HCAIM) that relieves lecturers of the burden of writing and record- is a project that aims to cope with this wide range of needs ing lectures while generating high-quality metadata to and issues by supporting the legal, regulatory-compliant, support impaired students. and ethical adoption of AI in education and other fields. It started in 2021 and is being conducted by four univerItal-IA 2023: 3rd National Conference on Artificial Intelligence, orga- sities, three research institutes, and three SMEs from five *nCizoedrrebsypCoInNdIi,nMg aayut2h9o–r3.1, 2023, Pisa, Italy European countries, with the support of the European $ flora.amato@unina.it (F. Amato); antonio.galli@unina.it Platform for Digital Skills and Jobs. The project seeks (A. Galli); michela.gravina@unina.it (M. Gravina); to promote the development of AI technologies that can lidia.marassi@unina.it (L. Marassi); stefano.marrone@unina.it enhance the quality, accessibility, and efectiveness of (S. 0M0a0r0r-o0n00e2);-5ca1r2l8o-s5a5n5@8(uFn.iAnam.iatt(oC);. 0S0a0n0s-o0n00e)1-9911-1517 (A. Galli); education by personalizing learning experiences and fa0000-0001-5033-9617 (M. Gravina); 0000-0001-6852-0377 cilitating automatic grading. However, the project also (S. Marrone); 0000-0002-8176-6950 (C. Sansone) acknowledges the ethical concerns that must be consid© 2023 Copyright for this paper by its authors. Use permitted under Creative Commons License ered when developing and implementing generative techCPWrEooUrckReshdoinpgs IhStpN:/c1e6u1r3-w-0s.o7r3g ACttEribUutRion W4.0oInrtekrnsahtioonpal (PCCroBYce4.0e).dings (CEUR-WS.org) nologies in education, such as the risk of perpetuating bias and inequality, privacy concerns, and the risk of fake news or manipulated opinions. The HCAIM project emphasizes the importance of a multidisciplinary approach and a thorough assessment of social and ethical impacts to ensure the development of ethically responsible and sustainable AI models that take into account issues of equity, privacy, and accountability.
In this paper, we report and deepen all these topics, highlighting for each the central role that the Universities can, should and must have.
three main modules: the audio generation module, the video generation module and a lip-syncing step.
The proposed methodology is a pipelined architecture summarized in Figure 1 and designed to create, by sequential steps, the video lesson generating the voice, the video, and lip-syncing the latter on the former. In particular, the implemented work is mainly focused on making the modules independent from each other, giving the possibility to replace or add new modules without the need for rewriting the whole architecture.
The Voice Generation module aims to generate an
audio file (e.g., in .wav format) reproducing the user’s voice
from a written text. To create a module able to generate
audio in a target subject’s voice, it is needed to collect
a dataset consisting of audio samples belonging to the
target subject associated with their transcription.
Besides the particular technology used to generate the voice,
during the design of LessonAble we observed that the
dataset collection stage can strongly afect the final
quality. Indeed, although the formatting step often depends
on the used text-to-speech technique, the dataset should
always respect some characteristics, including both short
and long voice clips having tone and pitch diferences,
avoiding wrong or broken files or background noise. We
leveraged the NVIDIA Tacotron 2 [
In the Video Generation module we used the First
Order Motion Model (FOMM) [
Despite other approaches being available, such as X2Face
[
To generate videos with a desired emphasis, LessonAble
is able to choose the most suited target image and
driving sequence, by extracting the metadata from the audio
In recent years, Massive Open Online Courses (MOOCs)
have spread exponentially, with their global market size
expected to grow from USD 3.9 billion in 2018 to USD 20.8
billion by 20231, at a Compound Annual Growth Rate
(CAGR) of 40.1% during the forecast period. This success
is mostly due to the wide range of benefits they ofer,
such as the prospect to rethink course content based
on analytics and the opportunity to provide diferent
course experiences through A/B testing to know which
educational experience is more efective. MOOCs [
An educational institution creates the learning content of a MOOC to teach and train students and experts.
However, creating the MOOC content, i.e. a video
lesson, often implies following a script (a text) already
deifned, with the author required to interpret it following
every line of the defined text instead of recording a
lesson on a wimp (as during a classical frontal lecture) to
generate high-quality metadata (e.g. dubs) to support
impaired students. Thus, despite this need tends to be
extremely time-consuming, it is often a mandatory fair
requirement. To take the best from this need, in [
As a case study, we generated a deep fake MOOC les- harmful. However, the recognised capabilities of this tool son for professor Carlo Sansone, from the University of make it necessary to reflect on the possible negative conNaples, Federico II, testing the generation process in two sequences that ChatGPT might entail. From an ethical languages, English and Italian. An example of Lesson- point of view, while misuse of this language model should able output is available at this link. We believe our eforts certainly be discouraged, it might be good to understand and our ideas in this approach can lead to new direc- how instead this tool might help if it were included in a tions, such as the possibility of creating real fake MOOC controlled way within educational institutions. ChatGPT content hubs. Furthermore, updating courses will be may in fact prove to be a valuable resource for both stumuch easier as authors will only need to edit the text, dents and faculty. In Naples (Italy) a collaboration with while the video will be generated automatically. Future the Department of Electrical Engineering and Computer research should focus on the application of expressions Science (DIETI) at the Department of Social Sciences at to the generated audio, providing the possibility to add the Federico II University has been initiated, with the slides during the lesson. The result will be an extremely main objective of implementing digital solutions to supintuitive tool to support MOOC content generation lever- port teaching in the humanities area. Specifically, the aging for a good purpose Deep Fakes, a technology often University of Naples Federico II intends to use ChatGPT associated with AI misuses, such as fake news, hoaxes in a controlled manner within its degree programs. In and scams. Thus, a side efect of this work is also to particular, the University is using this model contextually further highlight that AI is only a (very powerful) tool, in the course Social Epidemiology, Algorithms and Big which misuse is to be associated with the user and not Data (included in the curriculum of the bachelor’s degree with the tool itself. in public, Social and Political Communication). ChatGPT is currently being used as a teaching support for faculty members, with a special interest in the activities of:
Nowadays, several companies have released free or opensource demos of their language models, which are attracting a great deal of interest. ChatGPT is one of the first examples of these chat-based generative models, whose original responses to each user input and ability to adapt to a wide range of topics make it suitable for performing diferent tasks. As a result, ChatGPT has quickly become popular among students as a valuable tool to improve their academic performance and save time in • Learning support, to help students understand concepts covered in lectures or clarify any doubts they may have about specific topics; • Teacher training, to train teachers on best teaching practices, pedagogy, and how to develop effective teaching materials; • Student feedback, to collect anonymous and rapid
feedback from students.
Controlled use of this digital tool can help teachers improve their approach and create a better learning environment for students. ChatGPT should also be used for models, but also that society itself reflects on the changes providing information and guidance to students on study that these tools might produce. It cannot be overlooked programs, course requirements, available services, and that these systems deal with human language processing, other useful information as if it were an extension of and thus directly afect the ethical dimension as well. It the teaching secretariat. And if these tools can be use- is important to develop ethically responsible and sustainful to students, they in turn can also help improve and able generative models that take into account issues of analyse the performance of AI models. Currently, the Uni- equity, privacy, and accountability. It is equally essential versity of Naples is also considering including Federico that there be more transparency about the use of these II students in a social experiment to assess ChatGPT’s models and their implications. This requires a multidislearning behaviour/ability with respect to constant inter- ciplinary approach to the design and implementation of action with a specific group of users. The general idea generative models, involving a variety of stakeholders would be to proceed, after a constant period of interac- and a thorough assessment of social and ethical impacts. tion with the algorithm, to take in targeted information Technological development also implies social responsito evaluate its consequences. The dialogue between the bility, so that the potential of this technology does not areas of computer science and the humanities will thus become a disadvantage instead. be handled both from a practical point of view, through the implementation of digital solutions to support the Department of Social Sciences, and by providing DIETI 4. HCAIM with a concrete opportunity to analyse the possible ethical consequences that the use of these solutions entails.
As the technology continues to develop, it is likely that we will see more universities adopting ChatGPT for a variety of applications. The role of academic institutions, in the development and dissemination of models, should not be underestimated. Indeed, the main reason for the rise in popularity of chat-themed generative models is their ease of use and accessibility, which is a concern with respect to the risk of people adapting to this AI without thinking through the possible negative consequences. What seems necessary, then, is for society to be pushed toward the prudent use of AI through increased awareness of new technologies.
is an interdisciplinary program that provides students
with a comprehensive understanding of designing,
developing, and evaluating human-centred AI systems [
While this tool may represent a new advance, the worst risk is that it is used without thinking through the possible ethical dificulties that this AI, like all generative models, might pose. Generative models are trained on data, which means they can easily perpetuate bias and inequality if present in the training set. The use of generative models can also pose concerns about intellectual property, privacy, and the risk of fake news or manipulated opinions. The cybersecurity industry should also consider the possible consequences of a cyber attack on these models. Indeed, ChatGPT itself could quickly generate targeted phishing emails or malicious code for malware attacks. The potential impact of generative models on society also raises several questions of accountability, as careless use of these models can have unforeseen efects. As is often the case with technologies, ethical responsibility seems to be shared. It is therefore important that not only designers of language models work responsibly and consider ethical issues in the development and use of
the necessary skills and knowledge to design, develop, and evaluate human-centred AI systems. It is an interdisciplinary program that fosters collaboration between computer science, social sciences, and humanities.
Students are required to complete a master’s thesis as part of the program, providing them with an opportunity to apply the skills and knowledge acquired during the program to a real-world problem. The thesis is an important part of the program, allowing students to demonstrate their ability to apply critical thinking, problem-solving, and research skills. The HCAIM program is supported by a network of industry partners and stakeholders who provide real-world case studies and research opportunities for students. This collaboration between academia and industry ensures that students are exposed to the latest developments and trends in the field of AI.
AI experts who prioritize the human perspective when designing, developing, and implementing AI systems.
The program aims to impart students with the expertise and knowledge necessary to create AI solutions that are socially responsible, technically sound, and ethically acceptable. HCAIM is also committed to producing innovative educational materials that will be used consistently at partner institutions, fostering the exchange of project ideas and students among participating countries.
Furthermore, the program ofers short-term Erasmus actions that facilitate the exchange of knowledge and ideas among students.
The program is designed to promote interdisciplinary collaboration among computer science, social sciences, and humanities, equipping students with a comprehensive skill set that can be applied across various domains 5. Conclusion and contexts. It is this collaborative and interdisciplinary approach that makes HCAIM stand out in the field of AI In conclusion, the use of AI in education has the potential education. Although the precise programs vary to some to revolutionize traditional educational approaches and extent across participating universities based on their improve accessibility, efectiveness, and personalization respective traditions, resources, and legal and cultural of learning experiences. In this paper we highlighted environments, the diverse array of elective courses avail- the importance of developing ethically responsible and able enables students to build their own personalized sustainable AI models that take into account issues of eqHCAIM portfolio, based on their interests in diferent uity, privacy, and accountability, and the central role that domains. Nonetheless, the program maintains a consis- universities can, should, and must have in this process. tent and unified curriculum across all partner institutes.
This is achieved through the collaboration of the HCAIM consortium, which has developed all courses and lec- References tures in a collective efort to provide students with a comprehensive and well-rounded education. Through this collective approach, HCAIM students will gain a shared understanding of the key concepts, principles, and techniques related to the development and implementation of human-centred AI systems. This ensures that students will acquire a consistent and cohesive foundation of knowledge, regardless of the university they attend.
degree ofered by four European universities, each with their own unique approach to delivering the program. The duration of the program and the number of credits required to complete it depend on the delivering university, ranging from one year (60 ECTS) to two years (120 ECTS). Some universities are ofering HCAIM as a stand-alone program, while others are integrating it into their existing programs. The HCAIM program is delivered entirely in English and comprises a combination of lectures, seminars, hands-on projects, and research activities. The program is designed to equip students with