5.1 Introduction1
Extended Reality (XR) is usually considered the umbrella term for Virtual Reality (VR), Augmented Reality (AR), and Mixed Reality (MR) (Vasarainen et al., 2021). Extended reality could be defined as ‘the tool that sits between the digital twin and real-world information’ (Casini, 2022). Virtual Reality encompasses virtually produced worlds, replacing the real world, where AR supplements reality through digital products. Mixed Reality is a term that is used for AR and forms of VR that are complemented with real world elements (also referred to as Augmented Virtuality, AV). Extended Reality is increasingly applied in many different fields, e.g. in education (Alnagrat et al., 2022; Sharma, 2021), the business world (Marr, 2021), health care (Andrews et al., 2019; Morimoto et al., 2022; Zhang et al., 2023), urban planning (Çöltekin et al., 2020), design (Schneider; 2021; Kharvari and Kaiser, 2022), military training (Boyce et al., 2022; Garcia Estrada et al., 2024), tourism (Santoso et al., 2022) and the Internet of Things (IoT) (Andrade and Bastos, 2019). The broad application scope touches many professions and individuals in society, including the major theme of this book (green technology), and the overall impact is considered to be significant (see all previous sources). At the same time, serious concerns around ethics issues are identified, which has led to an extensive academic debate (Tromp et al., 2018; Carter and Egliston, 2020; Burns, et al., 2022; Ramirez et al., 2023; Indradevi et al., 2024). Amongst others, the question is raised on how to regulate XR developments in the near future, and how to solve moral dilemmas that come with the wide use of XR applications.
Regulation is done by law and assessment of conformity to that law supported by standardisation documents. It usually does not focus on how to resolve moral dilemmas calculating harm/risk/benefits but on principles how to avoid harm. The language of ‘weighing harm’ is far away from the principled language of ‘fundamental rights’.
regulation
The EU AI act, that also applies to XR systems that make use of AI components, was recently adopted and breaks with this principles approach. This law has a risk-based product safety approach for the design and development of AI systems where risk is understood as the probability of harm and the severity of that harm where severity can be both qualitative and quantitative. Harm is further detailed in relation to safety, health and fundamental rights. Therefore for high-risk AI systems we have to shift from the principle of avoiding all harm to the concept of acceptable level of risk since for high-risk AI systems not all possible harm can be excluded. The consequence of that is that AI systems engineers need a more explicit formulation of harm. In addition, to the list of AI systems to which this law applies, special attention is given to ‘General Purpose AI’ (GPAI) models, also known as foundation models or Generative AI models. Special ‘systemic risk’ management requirements are formulated where systemic risk relates to the whole of (EU) society.
EU AI act
System applications are classified into categories of low, medium, high, and unacceptable risk. Most attention is on high-risk AI system applications since there are potential high risks, but the benefits are considered higher that the (potential) costs in terms of harm to citizens or society at large.
In the development of high-risk AI systems, developers typically encounter ethical and moral dilemmas pertaining to safety, health, and fundamental rights. It is imperative for them to mitigate these risks to an acceptable level.
Dealing with moral dilemmas in system design is not specific for AI but also holds for XR systems. In addition, most XR systems make extensive use of AI in these systems and will therefore also have to comply to the AI act. Also, a wide range of specific AIXR system risks and subsequent ethical/moral dilemmas might arise particularly in relation to GPAI where multimodal generative AI can be harnessed to mimic or forge more and more aspects of human behavior. What is unique about introducing AI in XR systems is that these systems can adapt during operation in order to avoid or reduce risk which necessitates a more formal (mathematical) specification of how to deal with moral dilemmas.
AIXR system risks
5.2 Moral problems
To discuss moral problems that come with AI or XR (or both), we first need to be clear of the nature of moral problems. This has been a matter of (Western) academic debate since the ancient Greek (the so-called Platonic conflict) and continued to be discussed ever since (e.g. Sartre, 1957). A moral problem would encompass a situation in which a moral actor feels the moral obligation or has moral reason to two or more things (moral actions) that rule out (but not override) the other moral actions (Sinnott-Armstrong, 1987) and/or both do something and refrain from doing so at the same time (Statman, 1990).
Whether or not a moral problem can exist in the first place depends on how this is approached in theories on normative ethics. One crucial element is how we consider the existence of moral truth (Dubbink, 2018). In absolutist normative theories, there are moral truths. A moral problem then exists when there is a conflict between moral truths, and one needs to prioritize one over the other. In some absolute normative theories, it is assumed that a normative theory should not allow the existence of (genuine) moral problems e.g. Immanual Kant’s categorical imperative (Timmermann, 2013), and some approaches in utilitarianism (Gowans, 1987), where moral dilemmas cannot exist within the normative theory in itself, since there is always one objectively justifiable and reasonable way to act in situations with moral challenges. In contrast, there are relativist theories that assume that moral truth does not exist, and moral righteousness is context and culture dependent (De Haan, 2001). In relativist theories therefore, moral problems are widely accepted phenomena, and there is a focus on handling differences between and prioritizing multiple moral solutions of whom each is morally justified/desirable according to the dominant moral viewpoints in one moral context, but not (per se) the other.
moral truth
In the second place, we need to establish that if moral problems exist, its possible moral solutions are comparable. When we accept that moral solutions are incomparable, the existence of moral problems becomes problematic since they cannot be solved (Chang, 2012). This implies that we need a method that allows the comparability of multiple moral solutions to a moral problem. In so called ‘moral hard cases’ (borrowed from the legal domain, e.g. (Statman, 1990) there is more than one solution to a moral problem that can be defended by moral theory, but these moral solutions contradict one another (Messerli et al., 2017). Therefore, it is necessary to find ways to meaningfully compare these moral solutions in order to choose one. Existing approaches in meta-ethics to compare moral solutions are comparative methods, such as the Agents, Deeds and Consequences (ADC)-model (Dubljević, and Racine; Dubljević, 2020), or the Expected Moral Value (EMV)-approach (Lockart, 2000) as we will discuss further in section 5.4
moral solutions are comparable
However, in the reality of technological innovation, tech-companies and tech regulators, it would be rather far-fetched to assume there is an absolute moral truth or universal moral procedure that drives their techno moral decision-making, or that the involved moral actors are capable of finding/using such a moral truth or procedure. We therefore have to assume that moral problems exist, regardless whether or not this is something that ought to be. Also, scenarios of human-crafted algorithms in unsupervised settings without human interference complicates matters in the application of meta-models, that are initially designed for human moral action (Wernaart, 2021). We therefore have to conclude there is a need for a moral procedure that does justice to and helps solve the practical challenges of techno-moral problems on a meta level.
5.3 Moral authority
When we accept that there are techno-moral problems, and there is a need to compare multiple solutions to a techno-moral problem, we need to address the matter of moral authority: who decides on the techno-moral problem (Wernaart, 2022)? In general, we could say that those who are affected by the techno-moral problem ought to have the moral authority to decide on these problems. With the design of technology, it is mostly the designers of the new technology that are able to make the most important design decisions, and strongly influence the way techno-moral problems are decided on. Considering that it is mostly not the designers who bear the moral consequences of their design choices, they usually do not have the moral authority to take such decisions (Millar et al., 2017). Ideally, designers are capable of sensing and incubating the values of those who bear the moral consequences of their design. This is however not as straightforward as it may seem. We need to find ways to establish who should have moral authority, and then map what type of moral solutions that are supported by those who have moral authority.
There are multiple ways to establish who has moral authority. Millar (2017) suggests making a distinction between high and low stake moral problems, and individual or societal consequences of these moral problems. When moral problems represent low stakes, it is more practical to assume that the designer of technology can take the moral decision and assume moral authority. When they are high stake, it would be irresponsible if the designer would decide on behalf of the moral authority, and therefore would have an obligation to incorporate the moral views of the moral authority in their design. If the moral consequences are centralized around a specific group of people or individuals (e.g. users) then they should decide. Sometimes, moral consequences are felt society-wide, and therefore would require the involvement of society. This is a rough sketch of how to distribute moral authority. A more detailed approach is offered by Wernaart (2021) through a moral dashboard for engineers, using the issue-contingent model that was introduced by Jones (1991). The moral intensity is measured according to six dimensions: magnitude of consequences, level of social consensus, probability of the moral effect, temporal immediacy, proximity and the concentration of the effects. These dimensions can help to establish into more detail whether moral issues are to be considered high or low stake, and to what degree moral consequences are individual or societal. This dashboard can help individual engineers or organizations to better understand the stake of their moral programming, and reflect on the allocation of moral authority.
high and low stake moral problems
So, considering that with most technological innovations the moral authority is not the same as the moral programmer, it is important to be able to separate those, and find useful manners to enable a moral programmer to involve the moral authority when designing/engineering new technologies (Wernaart, 2022). It is also important to find useful approaches to decide on how to establish the moral preferences of the moral authority, and in case of conflicting preferences, how to settle these. To this end, theories in moral programming may be of help, involving meta-ethical calculation methods.
5.4 Moral programming
As discussed earlier, the EU AI act has risk-based product safety approach for the design and development of AI systems where risk is understood as the probability of harm and the severity of that harm. It is therefore required that during product design and development there should be a risk management procedure where someone in the company is responsible for risk management. Some confusion has come up with recent developments in AI regarding the misconception that an AI system would be capable of fully autonomous risk management. However, an AI system cannot understand morality and its underlying explanations and therefore cannot construct a new run-time-adaptive moral theory itself so it is up to humans to craft machine processable (and therefore mathematical) heuristic moral models (Aliman and Kester 2022a) which needs continuous human supervision and feedback loops to keep evaluating if the moral model still functions as intended.
It is also for this reason that the EU AI act states that the objectives of the AI system should be human specified. For high-risk AI systems these objectives should be updatable by-design and machine readable (i.e. a mathematical objective function) and contain this moral model. Thereby, it is possible to select a strategy wherein the responsibilities of designer/developer and moral authority are clearly separable (Aliman et al., 2019) – a procedure termed “orthogonality-based disentanglement” – where the moral authority is engaged in the crafting of the objective function (also known as ethical goal function) which encodes the moral model while the designer is responsible for designing a system that optimizes on this updatable objective function/ethical goal function.
But how are we to heuristically represent morality in a mathematical model? For this we first need to address the meta-ethical question: “what would a proper mathematical representation of the moral model be”? In academia, various meta-ethical models h ave been proposed that can help make a decision between various normative ethical frameworks. One example is a comparative approach (Chang, 2012), in which the extent of which a value is realized per normative ethical theory is compared and expressed in percentages. Another related approach is the Expected Moral Value Theory (Lockhart, 2000) which adds the subjective assessment of the credibility of normative theories to the equation. Also here, multiple actions are reviewed against multiple moral theories, comparing to what extent the moral action that is proposed fulfils the values that are supposed to be fulfilled by each moral theory. The result is multiplied by the credibility of that moral theory according to the moral authority in a scale of 0.0 to 1.0. Yet another example of a meta-ethics model is the Agent-Deed-Consequence (ADC) Model (Dubljević, and Racine; Dubljević, 2020), interlacing consequentialism, deontology and virtue ethics through a score system.
meta-ethical models
However, these methods are still based on predefined (mostly Western-oriented) normative theories that are both abstract and not translatable to a fine-grained moral programming code. Instead, a science based non-normative approach is required that allows for broader cultural perspectives, that involves the moral authority and other stakeholders to give feedback as they perceive the harm done and a method to continuously provide updates as society evolves.
So, instead of putting forward a normative theory, and instead of purely basing it on empirical ethical enquiries, Aliman and Kester (2022a) have developed a descriptive and explanatory framework which follows the scientific method to resolve the issues discussed above which is coined Augmented Utilitarianism (AU).
Augmented Utilitarianism (AU) is a generic non-normative framework insofar as it does not prescribe any values that should be set as ethical goals. Instead, it leaves it open to the moral authority to fill in the blanks. In short, it is a mathematical model that does not specify what should be optimized (as the conventional utilitarianism does). It does not apply one abstract, rigid theory, nor does it deliberately mimic cognitive dissonances or mirror moral contradictions. This also allows the framework to be applicable cross-culturally as the attributes and values can be adjusted according to the relevant society’s worldview and values. Insofar as it prescribes a method, the idea of optimization inevitably imposes its own ontological and epistemological assumptions, that is, respectively, its own theory of being and theory of knowledge. To ensure that the values of a pluralistic society can be encoded into moral models for AI systems, “AU targets what one could conceive of as a possible smallest heuristic moral superset (SHMS) capturing the plurality of candidate ethical frameworks available in practice for moral programming – with the pre-condition that AU must flag known formal inconsistencies to forestall predictable practical safety problems in AI deployment”. Such a SHMS aims at accommodating for the breadth and variety of widespread moral frameworks by covering the following parameters: the agent (as e.g. in the focus of virtue ethics), the action (as e.g. key to deontology), the user (as e.g. in the focus of consequentialism) and the experiencer of the moral situation (as e.g. foregrounded in care ethics and many non-Western ethical frameworks).
Augmented Utilitarianism
On the whole, AU has its scientific basis in affective and cognitive neuroscience, and in moral psychology literature. Regarding the latter, AU is compatible with Gray and Schein’s theory of dyadic morality or dyadic completion. They argue that moral judgement and the nature of harm are not an objective matter of reason and should instead ‘be redefined as an intuitively perceived continuum’ (Gray and Schein, 2018). Dyadic Morality postulates that an act is assigned moral ‘value’ according to ‘norm violation’, ‘negative affect’ and crucially, ‘perceived harm’. In sum, AU allows the mathematical combination of consequentialism, deontology, virtue ethics and other frameworks and in addition perceived moral intuitions. It positions itself as a descriptive and explanatory framework that aims to capture the nuances in human morality’s functioning and moral pluralism rather than as a normative framework. Indeed, it recognizes that values can be attributed differently cross-culturally, and that harm is often perceived differently.
As such, AU is a ‘mathematical scaffold’ of how humans view scenarios and attribute value and is mathematically operationalizable in terms of ethical goal functions. This methodological framework follows the tenets of the scientific method. The functions are composed of attributes and values, which can be adjusted and determined by the relevant society. What matters is that this specification matches the world view of this society. Once specified, these functions can be tested in simulation environments and adjusted based on feedback from the relevant society. As such, AU adopts a dynamic process which is referred to as the ‘socio-technological feedback (SOTEF) loop’ (Aliman et al., 2019). If a function did not capture what was originally intended by programmers, it can be amended to better match the relevant conception. The ethical goal functions are the requirements we have on the system – with such settings, a system may be able to deduce how to adapt at run-time. To the extent that these functions match with the world view of (the) relevant society, such functions should be able to resolve impossibility theorem scenarios, as described in more depth elsewhere (Gros et al., 2024) and the engagement enables that ethical considerations are grounded in the lived experiences of individuals and adapt to shifts in societal attitudes over time.
mathematical scaffold
socio- technological feedback (SOTEF) loop
5.5 Implications for moral programming in XR contexts
A first concrete step in constructing the ethical goal function for a specific application is to identify the relevant attributes in the ‘mathematical scaffold’.
The identification of relevant attributes is just a first step. The next step is how the different attributes are mathematically related to each other. Do they just add up like in a cost benefit analysis or are the different mathematical components related to the attributes correlated in a more complicated way? A final step is how the different attribute specific utility functions are shaped (e.g. concave, convex or linear) within the mathematical scaffold of AU. For all these different steps specific designed elicitation process will be needed.
The SOTEF loop describes how the different actors are engaged in the various processes. In figure 5.1 an illustration, presented in Aliman and Kester 2022a, of such a loop is shown.
Illustration of the Socio-Technological Feedback Loop from a human-machine interaction perspective (Aliman and Kester 2022a)
Another angle (Heijnen et al. 2024) on interpreting the SOTEF loop is the life cycle of the AIXR system as depicted in figure 2. In this view four layers can be distinguished, the governance loop, the design loop, the development loop and the operational loop. The idea is that the actors identified in figure 1 are grouped in governance, design, development and operational teams. It is clear from both these pictures that already at an early stage of the AIXR system life cycle the process of moral programming should be centre stage. After all, the ethical goal function can also be regarded as a requirement specification that is needed not only to make run-time decisions but also to make design choices for the AIXR system. It can also be regarded as the function on which the system is optimising run-time, through which the system can be controlled by the operational team, by which the functioning of the system can be explained, and against which the system can be tested. It is therefore crucial that the relevant actors in the SOTEF loop are familiar with what the ethical goal function stands for and how it affects the functioning of an AIXR system. Relevant actors in the early stages of the AIXR system development are, referring to figure 1, the manufacturer and the stakeholders. Being familiar with AU encoding, they will identify the relevant attributes and, after consulting the law, come up with a first version of the ethical goal function. Thereafter, manufacturer, stakeholders and legislative power can evaluate the functioning of the AIVR system. In figure 1, we can also identify the inner feedback loop between stakeholder and intelligent system as a constant interaction where the AU function can be adapted and updated during operation that relates to the control loop in figure 5.2. Also, a wider societal feedback loop with environment, legislative power and judicial power can be identified that acts as moral authority at the longer timescale of the governance loop in figure 2.
Illustration of the Socio-Technological Feedback Loop from a system life-cycle perspective
For high-risk AIXR systems it may often be too dangerous to test the system in real world scenarios. In that case, at early stages of the development, the test environment could e.g be a VR environment where certain aspects of the real world are simulated, also known as sandbox approach. The moral programming procedure with AU and the SOTEF loop also suggests that in such a simulated environment all the different people involved in construction of the ethical goal function could already have rich experiences able to augment their moral understanding and judgement. There is already experience with moral decision making by people in VR environments (Aliman and Kester, 2019), however, in many of these experiments the question was what people would decide themselves or how to augment their own moral decision making, sometimes even under time constraints. In the case of moral programming the question is more: “what would the relevant actors in particular the moral authority as experiencer want the AIXR system to do?”. For this question it is important to realize that the time pressure on decisions does not play a role since whatever the moral authority decides to value, the AIXR system could execute that instantly. As Lisa Barrett explains (Barrett, 2017) human core affect and rationality are inseparable. Interoception and core affect are part of the construction of moral judgments. Moral authorities should take as much time as needed for a more detailed construction of particularly the moral model of societal-level ethical goal functions.
high-risk AIXR systems
5.6 Looking ahead
We have outlined the concept of a new approach of moral programming for AIXR systems dealing with many scientific and philosophical issues. One might wonder why such an approach is not already a well-established procedure? Many issues may play a role:
-
The procedure of moral programming is a transdisciplinary approach and involves scientists from many different domains. This can easily lead to misunderstanding, inconsistent concepts from the different disciplines, incompatible doctrines and lack of trans-disciplinary oversight.
-
Some people feel uncomfortable talking or analysing themselves about their values particularly if it concerns severe harm. They rather talk about what choice they would make in particular circumstances, but this is much harder to generalize.
-
Developers of systems do not always want to be transparent about the moral choices they make during system design, particularly if the objective of the system is not in the interest of users/society or it would unveil malfunctioning of the system, or when transparency would potentially damage Intellectual Property benefits or stands in the way of the overall business model.
-
The moral authority is not always keen to exercise that authority particularly if it is publicly transparent, also it is not always clear to the moral authority where he/she takes responsibility for.
-
Policy makers who are key players in regulating procedures that enable the moral authority to put forth its preferred moral programming are hesitant in regulating this. Sometimes this is due to a lack of technological expertise, sometimes this is due to a powerful tech-lobby who opposes such regulation.
-
A new type of problem arises with Large Language Models where model providers could use their model to spread misinformation and disinformation about AI that is in the favour of that company. Particularly misinformation and disinformation about misconceived existential risks (see e.g. (Duarte et al., 2024)) can complicate value agreements and regulation activities.
It is for these reasons that for moral programming an integral approach is needed that also takes the issues into account that are not directly related to moral theory, value elicitation and moral programming.
In addition, the possibility of so-called AI agents in XR (Wan et al., 2024) is a new challenge that needs to be addressed. Advances in Generative AI facilitate the development of these agents both for beneficial and malicious use. Unintentional substantial risks can arise in the domain of privacy and concerning the overestimation of AI agent capabilities. Defences against malicious use are even more challenging. Detection of artificial agents deployed by malicious actors will be very difficult because every detection mechanism can in principle be countered unless the AI agent is tested on something AI agents cannot do: the construction of new scientific or moral theories. In Aliman and Kester 2022b and 2023 possible defences in XR environments are discussed. Moral programming can also help here to determine possible risk mitigation strategies.
References
Aliman, N.M., and Kester, L., 2019. Extending socio-technological reality for ethics in artificial intelligent systems. In: IEEE international conference on artificial intelligence and virtual reality (AIVR), pp. 275–2757.
Aliman, N.M., Kester, L., Werkhoven, P. and Yampolskiy, R., 2019. Orthogonality-based disentanglement of responsibilities for ethical intelligent systems. In: Hammer, P., Agrawal, P., Goertzel, B., Iklé, M. (Eds.) Artificial General Intelligence. AGI 2019. Lecture Notes in Computer Science, pp. 22–31. Springer Verlag. https://doi.org/10.1007/978-3-030-27005-6_3.
Aliman, N.M. and Kester, L., 2021. Epistemic defenses against scientific and empirical adversarial AI attacks, CEUR Workshop Proceedings, vol. 2916.
Aliman, N.M. and Kester, L., 2022. Moral programming: crafting a flexible heuristic moral meta-model for meaningful AI control in pluralistic societies, in: B. Wernaart (Ed.), Moral design and technology. Wageningen Academic Publishers, Wageningen, the Netherlands, pp. 63–80 https://doi.org/https://doi.org/10.3920/978-90-8686-922-0_4.
Aliman, N.M. and Kester, L., 2022. VR, deepfakes and epistemic security. In: IEEE international conference on artificial intelligence and virtual reality (AIVR). IEEE, pp. 93–98.
Aliman, N.M., Kester, L., 2023. AI-related misdirection awareness in AIVR. PhilPapers.
Alnagrat, A., Ismail, R.C., Idrus, S.Z.S., and Alfaqi, R.M.A., 2022. A review of extended reality (XR) technologies in the future of human education: current trend and future opportunity. Journal of human centered technology, 1:81–96.
Andrade, T., and Bastos, D., 2019. Extended reality in iot scenarios: concepts, applications and future trends. In: 5th Experiment International Conference (exp. at’19). IEEE, pp. 107–112.
Andrews, C., Southworth, M.K., Silva, J.N., and Silva, J.R., 2019. Extended reality in medical practice. Current treatment options in cardiovascular medicine, 21:1–12.
Barrett, L.F., 2017. The theory of constructed emotion: an active inference account of interoception and categorization. Social cognitive and affective neuroscience, 12. https://doi.org/10.1093/scan/nsw154.
Boyce, M.W., Thomson, R.H., Cartwright, J.K., Feltner, D.T., Stainrod, C.R., Flynn, J., … and Rovira, E., 2022. Enhancing military training using extended reality: a study of military tactics comprehension. Frontiers in virtual reality, 3:754627.
Burns, M.B., Lebkuecher, G., Rahman, S., Roytman, M., Samoska, S., and Vukov, J., 2022. Extended frameworks for extended reality: ethical considerations. AJOB neuroscience, 13:171–173.
Carter, M., and Egliston, B., 2020. Ethical implications of emerging mixed reality technologies, the University of Sydney.
Casini M., 2022. Extended reality for smart building operation and maintenance: a review. Energies, 15:3785. https://doi.org/10.3390/en15103785.
Chang, R., 2012. Are hard choices cases of incomparability? Philosophical issues, 22:106–126.
Çöltekin, A., Lochhead, I., Madden, M., Christophe, S., Devaux, A., Pettit, C., … and Hedley, N., 2020. Extended reality in spatial sciences: a review of research challenges and future directions. ISPRS international journal of geo-information, 9:439.
De Haan, J., 2001. The definition of moral dilemmas: a logical problem. Ethical theory and moral practice 4:267–284. https://doi.org/10.1023/A:1011895415846.
Duarte, T., Barrow, N., Bakayeva, M., and Smith, P., 2024. The ethical implications of AI hype. AI and ethics, 1–3.
Dubbink, W., 2018. A typology of ethical problems. Ethical perspectives, 25: 683–714.
Dubljević, V., and Racine, E., 2014. The ADC of moral judgment: opening the black box of moral intuitions with heuristics about agents, deeds, and consequences. AJOB neuroscience, 5:3–20. https://doi.org/10.1080/21507740.2014.939381.
Dubljević, V., 2020. Toward implementing the ADC model of moral judgment in autonomous vehicles. Science and engineering ethics 26:2461–2472. https://doi.org/10.1007/s11948-020-00242-0.
Garcia Estrada, J., Prasolova-Førland, E., Kjeksrud, S., Themelis, C., Lindqvist, P., Kvam, K., and Ricci, S., 2024. Military education in extended reality (XR): learning troublesome knowledge through immersive experiential application. The visual computer, 1–30.
Gros, C., Kester, L., Martens, M., and Werkhoven, P., 2024. Addressing ethical challenges in automated vehicles: bridging the gap with hybrid AI and augmented utilitarianism. AI and ethics. https://doi.org/10.1007/s43681-024-00592-6.
Gowans, C.W. (Ed.)., 1987. Moral dilemmas. Oxford University Press, USA.
Heijnen, M., Schoonderwoerd, T., Neerincx, M., van der Waa, J., Kester, L., van Diggelen, J., and Elands, P.A., 2024. Socio-technical feedback loop for responsible military AI life- cycles from governance to operation. In: Responsible use of AI in military systems. Chapman and Hall/CRC, Oxford, UK, pp. 17–36. https://doi.org/10.1201/9781003410379-3.
Indradevi, R., Venkataswamy, K.P., Saravanan, S.A., and Mittal, A., 2024. Ethical considerations as limitations of extended reality: an empirical study through the lens of retail consumers. In: Multidisciplinary applications of extended reality for human experience. IGI Global, Hershey, PA, USA, pp. 44–63.
Jones, T.M., 1991. Ethical decision making by individuals in organizations: an issue- contingent model. Academy of management review, 16:366–395.
Kharvari, F., and Kaiser, L.E., 2022. Impact of extended reality on architectural education and the design process. Automation in construction, 141:104393.
Lockhart, T., 2000. Moral uncertainty and its consequences. Oxford University Press, Oxford, UK.
Marr, B., 2021. Extended reality in practice: 100+ amazing ways virtual, augmented and mixed reality are changing business and society. John Wiley & Sons, Hoboken, NJ, USA.
Messerli, M. and Reuter, K., 2017. Hard cases of comparison. Philosophical studies, 174:2227–2250. https://doi.org/10.1007/s11098-016-0796-y.
Millar, J., 2017. Ethics settings for autonomous vehicles. In: Patrick Lin, P., Abney, K. and Jenkin, R. (editors). Robot ethics 2.0: from autonomous cars to artificial intelligence, Oxford Academic, Oxford, UK, pp. 20–34.
Morimoto, T., Kobayashi, T., Hirata, H., Otani, K., Sugimoto, M., Tsukamoto, M., … and Mawatari, M., 2022. XR (extended reality: virtual reality, augmented reality, mixed reality) technology in spine medicine: status quo and quo vadis. Journal of clinical medicine, 11:470.
Ramirez, E.J., Jennett, S., Clay, D., and Gandhi, M., 2023. Extended reality, control, and problems of the self: an ethical analysis. In: Exploring extended realities. Routledge, Oxfordshire, UK, pp. 104–129.
Santoso, H.B., Wang, J.C., and Windasari, N.A., 2022. Impact of multisensory extended reality on tourism experience journey. Journal of hospitality and tourism technology, 13:356–385.
Sartre, Jean-Paul, 1957. Existentialism is a humanism, Trans, Philip Mairet, in Walter Kaufmann (ed.), Existentialism from Dostoevsky to Sartre. Meridian, New York, NY, USA, pp. 287–311.
Schein, C., and Gray, K., 2018. The theory of dyadic morality: reinventing moral judgment by redefining harm. Personality and social psychology review, 22:32–70. https://doi.org/10.1177/1088868317698288.
Schneider, R.O., 2021. Extended reality: the next frontier of design. The international journal of design management and professional practice, 15:43.
Sharma, R., 2021. Extended reality: It’s impact on education. International journal of scientific & engineering research, 12:247–251.
Sinnott-Armstrong, W., 1987. Moral dilemmas and ‘ought and ought not’. Canadian journal of philosophy. 17:127–139. https://doi.org/10.1080/00455091.1987.10715904.
Statman, D., 1990. The debate over the so-called reality of moral dilemmas. Philosophical papers, 19:191–211. https://doi.org/10.1080/05568649009506337.
Timmermann, J., 2013. Kantian dilemmas? Moral conflict in Kant’s ethical theory. Archiv für geschichte der philosophie, 95:36–64. https://doi.org/10.1515/agph-2013-0002.
Tromp, J., Le, C., Le, B., and Le, D.N., 2018. Massively multi-user online social virtual reality systems: ethical issues and risks for long-term use. Social networks science: design, implementation, security, and challenges: from social networks analysis to social networks intelligence, pp. 131–149.
Van de Poel, I. and Kroes, P., 2014. Can technology embody values? In: Kroes, P. and Verbeek, P.P. (eds.). Moral agency and technical artefacts. Springer, Dordrecht, the Netherlands, pp. 103–124.
Vasarainen, M., Paavola, S., and Vetoshkina, L., 2021. A systematic literature review on extended reality: virtual, augmented and mixed reality in working life. International journal of virtual reality, 21:1–28.
Wan, H., Zhang, J., Suria, A.A., Yao, B., Wang, D., Coady, Y., and Prpa, M., 2024. Building LLM-based AI agents in social virtual reality. In: Extended abstracts of the CHI Conference on Human Factors in Computing Systems, pp. 1–7.
Wernaart, B., 2021. Developing a roadmap for the moral programming of smart technology. Technology in society, 64:101466.
Wernaart, B., 2022. Moral authority. In: Moral design and technology. Wageningen Academic Publishers, Wageningen, the Netherlands, pp. 29–46.
Zhang, J., Lu, V., and Khanduja, V., 2023. The impact of extended reality on surgery: a scoping review. International orthopaedics, 47:611–621.
This chapter was written as a part of XR4HUMAN: a three-year EU-funded project. Its mission is to co-create living guidance documents on ethical and related policy, regulatory, governance, and interoperability issues of XR technologies whilst building public trust and acceptance and a strong and competitive European XR ecosystem.