The importance of ethics in artificial intelligence is increasing, and this must be reflected in the contents of computer engineering curricula, since the researchers and engineers who develop artificial intelligence technologies and applications play a key part in anticipating and mitigating their harmful effects. However, there are still many open questions concerning what should be taught and how. In this paper we suggest an approach to building a syllabus for a course in ethics of artificial intelligence, make some observations concerning effective teaching methods and discuss some particular challenges that we have encountered. These are based on the pilot implementation of a new course that aimed to give engineering students a comprehensive overview of the ethical and legislative aspects of artificial intelligence, covering both knowledge of issues that the students should be aware of and skills that they will need in order to competently deal with those issues in their work. The course was well received by the students, but also criticized for its high workload. Substantial difficulties were experienced in trying to inspire the students to engage in discussions and debates among themselves, which may limit the effectiveness of the course in building the students' ethical argumentation skills unless a satisfactory solution is found. Several promising ideas for future development of our teaching practices can be found in the literature.
Applications of artificial intelligence (AI) involve many ethical challenges. The increasing autonomy of AI-based systems to make decisions that may have harmful consequences to living beings raises questions concerning the safety, transparency and accountability of such systems. Individuals’ rights to privacy and self-determination are threatened by the collection of vast amounts of personal data to be analyzed using AI algorithms for purposes such as surveillance and psychological manipulation. There are various efforts to regulate the use of AI either currently underway or already implemented, but given the rate at which AI technology is advancing, it seems doubtful that policymaking alone will be enough to curb morally questionable uses of AI effectively. At the forefront of this progress are the AI researchers and engineers, so as a complementary approach, we should aim to foster a strong sense of professional ethics among this community. This, in turn, means that ethics education should be included in AI engineering curricula at higher education institutions.
It is much less obvious, however, what exactly should be taught and how. For example, if we consider autonomous cars, there are some things that can be imparted to the students as facts, such as the safety record of these vehicles so far and the current legislation governing their development and deployment, but arguably the majority of interesting issues are subject to ambiguity and debate. These include relevant ethical principles and their interpretation and prioritization (e.g. how an autonomous vehicle should weigh the safety of passengers against that of other people), broader societal implications (e.g. effects of the proliferation of autonomous vehicles on the transportation sector), and implications of possible game-changing future developments (e.g. artificial general intelligence). A comprehensive syllabus thus needs to strike a balance between technical and philosophical topics, as well as between practical topics that the students can apply immediately and more theoretical ones that enable them to keep up with future developments.
Finding such a balance is a considerable challenge, especially when the time devoted to ethics in the curriculum is limited. Furthermore, the non-technical nature of much of the subject matter must be reflected in teaching and evaluation methods, which may require both students and teachers to adopt unfamiliar ways of thinking, since courses designed to build engineering skills typically do not expose them to philosophical concepts or methods. In this paper we examine the problem of teaching AI ethics to university-level engineering students based on the pilot implementation of a new course in the spring term of 2021. The course was offered at the University of Oulu, Finland, to any interested students as a non-mandatory part of a computer engineering curriculum, with emphasis on practical knowledge and skills that the students can use to identify and address ethical issues likely to be relevant to them in the present or the near future, but also some coverage of more theoretical topics.
In the remainder of the paper, we first present a review of related work in Section 2, focusing on surveys and reports of existing courses dealing with AI ethics. In Section 3 we give an overview of the objectives, topics and implementation of our own AI ethics course. In Section 4 we propose a framework for categorizing study topics and an approach to syllabus design based on the framework. In Section 5 we offer some reflections on teaching methods and practices, based on what we learned from the pilot implementation. In Section 6 we look at feedback received from students who took the course, and in Section 7 we present a discussion and conclusions. 2
Building on this work and another previous study
Furey and Martin (2018) have developed a module on ethical thinking about autonomous vehicles in an AI course. They suggest incorporating this type of introductory lesson about ethics into a one-semester AI course. Through a modular approach, students have an opportunity to connect specific AI topics to the related ethical implications. In this module, students become familiar with the use of thought experiment through the trolley problem, as well as learn to understand the complex nature of ethical dilemmas. Regarding the ethics module, the course assessment is carried out as follows: the final project paper of the course requires discussion of the ethical implications of the project idea, and in the final examination there is a question assessing the students’ understanding of the trolley problem.
Based on the AI ethics teaching examples described in this section, we have made the following remarks: The learning outcomes and objectives of the described courses and modules are relatively similar. They aim to increase students’ awareness of the ethical implications of AI usage, and to train the students to identify and address ethical issues related to the development and deployment of AI in practice. To achieve the learning outcomes, students are typically taught the key concepts and theories of ethics in parallel with current ethical issues related to AI usage. Furthermore, critical thinking as well as skills to discuss and appraise ethical issues are often promoted through teaching methods that support learning of this type of skills (such as article or case study analyses, or class discussions). In some courses, the students are also equipped with skills to use research methods for exploring the ethical implications of AI or trained to use practical tools to address the ethical issues. In higher education AI ethics teaching, a diverse set of methods and strategies have been reported. Standalone courses typically utilize assignments that comprise analysis of articles, stories or movies, discussions or seminars on the identified ethical issues, as well as consideration of how to solve the studied issues using suitable tools. AI ethics modules that can be integrated into other courses, for their part, aim to train students e.g. to use specific methods for investigating ethical issues, as well as to consider the ethical aspects of a specific application, project or activity. It seems that assessment methods other than traditional written exams can be employed when teaching AI ethics; this may be an approach that matches well the reported learning outcomes.
Like the majority of the AI ethics teaching examples presented above, our new
course was carried out as a standalone course. Many of the existing courses and modules
comprise topics and learning outcomes that are relatively similar to our ones,
specifically with respect to the ethical issues of AI and the aim to provide the students
with ethics skills needed in the development and deployment of AI systems in practice.
Teaching methods and strategies seem to vary a great deal among the reviewed courses
and modules, and an extensive set of these has been reported in the literature. Similarly
to what
The planning of the course, worth 5 ECTS credit points, began in 2020. The full title of the course was “AI Ethics, Privacy and Legislation”. Because of the COVID-19 pandemic, the course was designed from the start to be taught remotely using the Moodle online learning platform, with lectures implemented as Zoom meetings. The learning outcomes for the course were defined as follows: Students will be aware of the ethical and legislative aspects and conditions that need to be considered in the design and deployment of AI applications; They will understand the ethical special characteristics of AI applications, compared to information technology applications in general; They will be able to examine existing and hypothetical AI applications from an ethical viewpoint and identify potential issues; They will be able to weigh the benefits of AI applications against their drawbacks also in a wider, societal context; They will be able to apply ethical principles in AI application design.
The course was lectured over a period of 8 weeks, with each week dedicated to one of 8 course themes. The themes are explained in Table 1. For each week of the course, two Zoom sessions were planned: a main lecture of approximately 1.5 hours and a supplementary session of approximately 45 minutes, the latter featuring a short presentation or demonstration on a specialized topic followed by discussion. Exceptions to this plan were week 5, when the supplementary session had to be cancelled due to scheduling conflicts, and week 8, when the supplementary session took the form of a tutorial for a design methodology.
The course participants were evaluated based on a series of 8 smaller assignments, one for each course theme, and a larger final assignment spanning all themes. Passing all 9 assignments was required in order to pass the course. The lecturers reviewed each submitted assignment and either accepted it as such or sent it back to the student for revisions; in either case the student would be given some verbal feedback on their work. The format of the smaller assignments varied from week to week, but as a typical example, the student would be instructed to choose an AI application and write a short essay about it, addressing questions related to the theme of the week.
For the final assignment, the students could pick one of two options. Students who attended a certain number of Zoom sessions could choose to write a lecture diary, in which they would discuss the topics covered by the lectures and supplementary sessions and reflect on what they had learned. As an alternative, students could choose to complete a final exercise, in which they would come up with an AI application concept and write a report discussing the ethical aspect of the application, guided by the 8 course themes. This second option was provided so that students who were unable to attend the lectures or preferred to work independently for other reasons could complete the course; to study the course themes, they had access to lecture slides as well as items of further reading curated by the lecturers. Each student had the opportunity to send in an incomplete draft of their final assignment and receive feedback on it to help them in preparing the final submission. Among the challenges of designing a balanced syllabus for an AI ethics course, there are two in particular that we wish to draw attention to. The first one of these is striking a balance between topics related to ethical reasoning and argumentation and those related to understanding AI technology and systems. Covering both areas is essential: the latter to enhance the students’ awareness of AI applications and their ethical implications, the former to enhance their ability to analyze these implications and to make justified decisions regarding ethical issues when encountering them in their work. We refer to this as the philosophy-technology axis.
The other challenge is the rapid rate of progress in AI research and development, which raises the question of how we can ensure that at least some of the course content remains relevant to the students also in the future. For dealing with issues that are immediately relevant to today’s AI practitioners, the course can offer practical tools such as ethical design guidelines and methodologies, but the further into the future we look, the less we know about what the capabilities of AI will be, and therefore the less likely it is that any concrete guidelines given today will be useful in dealing with the ethical implications of those capabilities. To help the students gain some degree of preparedness, the course should cover technology topics representing visions of the future of AI, as well as philosophy topics that are more theoretical in nature but also more stable over time. This we refer to as the practice-theory axis.
If we consider these two axes as perpendicular dimensions, we can visualize them as shown in Fig. 1. We can thus divide possible study topics in AI ethics into four broad categories, corresponding to the four quadrants of the figure. Counterclockwise from the top left, these are as follows: Timeless Foundations: established concepts, theories and traditions in ethics; Practical Guidance: applied ethics principles and guidelines relevant to AI; Here and Now: AI applications of the present and the near future and the ethical issues associated with them; Beyond the Horizon: future potential of AI and the ethical implications of hypothetical scenarios.
Fig. 2 shows a selection of possible topics to be discussed on an AI ethics course, placed approximately where they are located on these two axes. The grey ellipse represents an approximation of the area from where we took the bulk of the subject matter for our course. Notably, the course syllabus was centered on the major ethical issues identified in present-day AI applications – safety, bias, accountability, explainability, privacy – and on the topics in the immediate vicinity of these, which provide essential context for the current issues and/or means for dealing with them. The least attention was devoted to the topics at the top of the figure, i.e. the most abstract and/or futuristic ones such as metaethics and superintelligence.
We argue that this represents a reasonable core syllabus for a practically oriented AI ethics course, and furthermore that this system for classifying study topics can be used as a practical framework for designing a syllabus for a course with a wider scope. Expanding the scope with respect to a given aspect of AI ethics can be thought of as stretching the grey scoping bubble; the visualization of affinities between topics is useful in ensuring that the course remains a coherent whole, since it suggests clusters of closely related topics that it would make sense to include together (and possibly also to discuss together, so the framework provides some hints for planning the course structure as well). The overall shape of the bubble serves as an indicator of the balancedness of the syllabus, so if the bubble becomes strongly elongated in some direction, this suggests that reviewing the planned syllabus and possibly including some additional topics for the sake of balance should be considered. A central challenge that we faced in the planning and execution of the course was presenting the subject matter to engineering students in such a way that it engages them to think about AI more philosophically than their other courses require them to. Toward this end, each week of teaching would follow the same general pattern, where the lecture would introduce the students to the theme of the week from a mostly theoretical perspective. The role of the supplementary session would then be to have the students discuss their thoughts with their peers, and the role of the assignment would be to have them apply what they had learned.
Real-world examples of AI applications were used in the Zoom sessions to illustrate theoretical concepts, and in many of the assignments to provide the students with a context where they can demonstrate their understanding of the subject matter. Perhaps predictably, when instructed to choose an application to analyze, the students tended to choose ones that are close to their own everyday lives; for example, services offered by Facebook or Google were popular choices. The evaluation of the assignments was based mainly on the students’ ability to build and defend arguments using the concepts and theories discussed in the Zoom sessions, although some of them also included a factfinding element. Furthermore, several assignments invited the students to show original thought by proposing ideas for addressing ethical issues associated with a given application or technology. Another category of AI applications that engaged the students’ interest were those involving palpable risk to core values such as health, safety or liberty. This is suggested by the popularity of such applications (e.g. medical AI systems, autonomous vehicles/weapons, predictive policing applications) as assignment subjects, and also by active audience participation during the week 2 lecture on controversial AI applications. The first recorded case of a pedestrian killed by an autonomous car was used already in the first lecture as an introductory example, since it provides an effective way to illustrate a number of key points: A malfunction in an AI system may have severe harmful consequences, including loss of life. When this happens, it may be hard to ascertain why it happened and who is morally or legally responsible. Even a hypothetical perfect AI system may encounter a situation where it has to make a value-based choice between multiple bad options. Short-term harm may be inevitable in the pursuit of long-term good, and the justification of such harm is also debatable.
Overall, the course achieved mixed success at best in inspiring discourse and debate among the students. This represents a missed opportunity for the students to learn valuable argumentation skills through peer-to-peer interaction. We can only speculate on the reasons here; presumably, the teaching methods employed were less than optimal for this purpose, but on the other hand, it may also be that many of the students (who were computer engineering majors) were not predisposed to working like this. Furthermore, it seems plausible that a physical classroom where the participants can better connect with one another would have been a more suitable environment for this. This was particularly evident in the final supplementary session, where the participants were given a tutorial for a design methodology based on a set of cards, designed to be printed out and manipulated as physical objects. Since this was not an option, the tutorial was implemented using an online whiteboard application, with the result that the students participated mainly by adding their own ideas to the whiteboard without interacting with one another on Zoom while doing it.
Online lecturing did have the unanticipated advantage that the students could use Zoom chat to ask questions and exchange thoughts even while the lecturer is speaking. We also set up a message board on Moodle and encouraged the students to use it for discussions, aiming to engage also those students who feel less comfortable expressing their views live on Zoom, but this proved unsuccessful. Another benefit of remote teaching was that this made it convenient to include lectures and presentations by visiting experts in the course program. Visitors were featured in two of the main lectures and three of the supplementary sessions; since all of these were planned from the start to be implemented as Zoom meetings, the visitors could deliver their presentations from where they are normally based and no special arrangements were required, making physical distance a non-factor in deciding which external experts to invite as visitors. Besides contributing their expertise to the course, the visitors enabled the principal lecturers to focus their efforts on course themes closer to their own respective areas of expertise, resulting in a higher overall standard of teaching quality than would otherwise have been achievable. 6
Feedback on the course was solicited from the participating students using two anonymous surveys. A standard feedback form, generated automatically for every course at the university, was filled in and submitted by 7 course participants. The form included 11 statements to be rated on a Likert scale from 1 (totally disagree) to 5 (totally agree). The results are summarized in Table 2. The students were also requested to evaluate their workload during the course, bearing in mind that 1 ECTS credit is equivalent to 27 hours of student work. The answer options and corresponding integer values were no workload (2), light workload (1), suitable workload (0), high workload (1) and very high workload (2); the answers were split between suitable workload and very high workload for an average of 1.14.
I have tried to advance my own learning outside the lectures (e.g. reading the lecture material, reading some literature connected to the topic or searching more information).
I have tried to advance my own learning during the lectures by discussing the topic with other students, by asking questions from the lecturer, by starting discussion in the whole group or questioning the teaching.
Avg
In addition to these, the standard form had open-ended questions on good practices, things to develop and other feedback. Notably, several of the answers to these questions included complaints about the workload of the course assignments. There were also individual criticisms concerning the course timetable and the supplementary session concept. Positive feedback was received on lectures, study materials, weekly assignments and guest lecturers. A supplementary feedback form was designed to capture more detailed information on the students’ expectations and learning outcomes. This form was available in the course workspace on Moodle and was filled in and submitted by 7 course participants. Concerning expectations, the form included the following questions: What were your reasons for signing up for the course? What were your expectations for the course with respect to your personal and professional development? Would you agree that the course fulfilled your aforementioned expectations? The first two of these were open-ended questions. Among the answers were some generic motivations such as finding the subject interesting and needing the credits to complete one’s studies, but several answers also indicated that the student was expecting knowledge of AI ethics to be an asset in jobs involving AI. The third question was to be answered on a Likert scale from 1 (strongly disagree) to 5 (strongly agree); the answers ranged from 3 to 5 for an average of 4.29. Optionally, the students could specify their answers to the third question; here one student indicated that they would have liked to learn more about the fundamentals of ethics, while another remarked that the exercises required a lot of effort and more writing than the student was used to, but that they were also good practice. Better visualization of lecture materials was also requested, and numeric grading instead of pass/fail was suggested as an incentive to put more effort into the assignments.
Concerning learning outcomes, the form included four Likert-scale questions corresponding to the four basic categories of study topics identified in Section 4. The theory here was that a balanced syllabus would contribute to the students’ awareness of ethical issues in AI – both those that are relevant now and those that are still in the future – as well as their confidence in their ability to deal with these issues in their work. The results are summarized in Table 3. The course increased my awareness of ethical issues that are relevant to AI applications today.
The course increased my confidence in my ability to competently address these current issues in my work, if and when I encounter them.
The course increased my awareness of ethical issues that may arise in the future, given the development potential of AI technology.
The course increased my confidence in my ability to competently address these future issues in my work, if and when they arise.
Avg
In this paper we discussed the teaching of AI ethics to engineering students, based on experience of lecturing a pilot course in the spring semester of 2021. We faced several challenges in designing a comprehensive and balanced syllabus for the course, as well as in the selection of effective teaching methods. The positive feedback received from participants suggests that the course was at least a moderate success, but since the number of students who took the course was small and only a fraction of them answered the feedback surveys, the results mut be interpreted with caution.
The aspect of the course that received the most criticism from the students was the workload. The most obvious candidate for an explanation is that the effort required to complete the assignments was underestimated when planning the course, but there are other possible contributing factors. For example, if the format of the assignments was such that it took the students out of their comfort zone – as engineering students they would be more familiar with e.g. programming assignments than essay writing – then this may have affected the students’ perception of the course workload. Another possible factor is that both the individual assignments and the course as a whole were graded on a simple pass/fail grading scale, so there were no rewards other than words of praise from the evaluators to be earned by surpassing the minimum requirements (as one student explicitly pointed out in their feedback). This is perhaps the most important issue to be addressed when planning future iterations of the course.
As another student noted in their feedback, commenting on the workload, the course format required a lot of effort from the teachers as well, which is an important point since it raises a question concerning the scalability of the course. Out of an original 30 students who signed up for the course, roughly half completed all the assignments – the high dropout rate perhaps another indicator that the students found the workload higher than expected – which kept the effort required to review the submissions manageable, but had all 30 completed the course, the allocated teaching resources would have been stretched to their limits. The issue here is that designing an alternative to the current assignment format that scales up for a significantly larger number of students while still effectively building and testing their ethical argumentation skills is far from trivial. Getting the students engaged in debates among themselves would both develop these skills and reduce the burden on the teachers, so again the crucial question is how to effectively facilitate such debates, given that there may be many students who are not comfortable with this type of work. This is another major aspect of the course that will need to be addressed in future development.
Among the practices reported in the studies reviewed in Section 2, there are some that seem particularly useful and promising for the future development of our teaching. For example, as there is still a lack of established textbooks, we should maintain and regularly update our current reading list and make sure that it includes both scientific papers and news articles (the latter to illustrate the real-world consequences of AI use). Due to the rapid technological development and changes expected in the field, the need to update the list should be reviewed regularly. As engineering is a typical field with a strong authority-based view to knowledge, it could be fruitful to think even more about how to move from this type of thinking to providing our students with skills to deal with unforeseen ethical issues in their future work tasks. Some of the proposed methods in the reviewed studies could fulfill this need well and would fit our course implementation, such as seminars with formal defender and opponent roles or case study analyses. Finally, we may also consider whether we should extend AI ethics teaching to reach a larger portion of the students, as well as how to make the connection between technical AI topics and their ethical implications clearer; in addition to our current standalone course, AI ethics modules or assignments could be integrated into technical AI courses starting from the early stages of engineering study programs. and