Teaching Engineering and Computing Ethics with Deepfakes

Duration: July 2022 – July 2023
Status: Completed
Educational Level: Tertiary Level
Topic: Artificial Intelligence AI
Keywords: Ethical Sensitivity

Initial Situation

The rise of deepfakes poses particular educational issues both for potential consumers of deepfakes and for their potential producers. Since the production and use of deepfakes typically depends on multiple inputs from a wide variety of actors, they provide an evident example of the “many hands” problem in engineering ethics (van der Poel, Royakkers, and Zwart, 2015), in which the attribution of individual responsibility becomes extremely difficult in collective settings. Furthermore, the capacity of technology to create separation between those producing deepfakes and those affected by their production increases the risk of producers feeling themselves free of traditional social obligations to others (Hoffman, 2000; 2008).

At present, the education of engineers and computer scientists seems ill-equipped to face the above challenges. Indeed, research evidence from engineering programmes (both internationally and in Switzerland) indicates that, far from developing ethical engagement during their studies, engineering students appear to become increasingly disengaged from ethical concerns (Cech, 2014; Tormey et al., 2015; Lönngren, 2020). Although there are efforts to develop ethics materials which could be used for teaching around deepfakes (see, for example, https://mediaethicsinitiative.org/, or the EPFL AMLD), like in the field of engineering ethics more widely, the design of case studies and other educational materials is typically not based upon evidence about how people learn ethics or indeed about how they learn more generally (see Hess and Fore, 2019).

Objectives

Our study aims to compare the effects of a computing education topic, specifically deepfakes, that includes both technical and ethical aspects with one that is purely technical on engineering students. Two main effects are of interest:

1. the impact of the educational content on students’ attention and emotional engagement with authentic and deepfake representations of a person, and
2. students’ moral judgment in ethically ambiguous situations.

Method

We used an experimental model to evaluate two potential educational approaches in using deepfakes to develop students’ moral sensitivity and moral reasoning (Bebeau, 2002). In both conditions students watched three videos: an authentic video, a high quality deepfake video, and a low quality deepfake video. Three types of data were collected: (i) attentional data was collected using eye-tracking (ii) emotional responses were collected using emotional facial recognition software, and (iii) moral judgement was collected using a protocol modelled on Kohlberg’s “Moral Judgement Interview”. In the control and experimental conditions, students were primed differently: In the control condition, they were primed to assess the video quality in terms of technical proficiency, and in the experimental condition, they were primed to assess the video quality in terms of both technical proficiency and ethical considerations. The research participants were engineering students who had taken at least one machine learning course.

Results

As a result of the project, a comprehensive methodological report was produced, describing an approach to studying the impact of ethical and technical education in the context of deepfakes. The report documents a multimodal research methodology designed to capture effects on three key dimensions: participants’ ability to recognize deepfakes, their attention and emotional responses toward individuals affected by deepfakes, and their moral judgment in ethically ambiguous situations.

Implemented Translation

The research findings offer numerous opportunities for practical application in ethics education within engineering. Universities can use them to redesign curricula, particularly in computing, by integrating ethics into courses on AI use in education. Educators may require training to implement this effectively; workshops can help them teach responsible AI use. The findings also support the development of educational materials such as textbooks, online resources, and case studies, enabling a blend of cognitive and emotional learning. New courses could integrate ethical discussions into existing technical programs. Collaborations with institutions and organizations in AI and ethics can broaden impact through shared practices. The results may also inform ethics guidelines in industry, helping update codes of conduct for AI-related work. Finally, public lectures, seminars, and workshops can engage wider audiences in discussions on deepfakes, AI ethics, and responsible technology use.

Participating Institutions