Macro Ethics Principles for Responsible AI Systems: Taxonomy and Directions

In the context of AI and ethics, there are two types of principles referred to: (1) those inferred from normative ethics such as deontology and consequentialism, as found in the work of Leben [85], and (2) those adapted from other disciplines like medicine and bioethics, such as those suggested by Cheng et al. [28], Fjeld et al. [50], Floridi and Cowls [54], Jobin et al. [76], Khan et al. [80], and Whittlestone et al. [137], including beneficence, non-maleficence, autonomy, justice, fairness, non-discrimination, transparency, responsibility, accountability, safety and security, explainability, human control of technology, and promotion of human values.

To ensure clarity of terminology, we refer to principles from normative ethics as ethical principles, and those highlighted by Floridi and Cowls [54] and Jobin et al. [76] as AI principles. We define ethical principles and AI principles as follows:

Existing taxonomies and surveys are present in the relevant but distinct domain of AI principles, such as those of Floridi and Cowls [54], Jobin et al. [76], and Khan et al. [80]. However, these works do not consider ethical principles. The rest of this article therefore surveys ethical principles rather than AI principles. Dignum [40], Leben [85], and Robbins and Wallace [110] provide summaries of normative ethics. Tolmeijer et al. [130] give an overview of implementations of machine ethics, providing useful guidance as to the technical and non-technical aspects of implementing ethics and evaluating systems. Similarly, Yu et al. [141] provide a concise guide to ethical dilemmas in AI and identify a high-level overview of ethical principles. From a philosophical perspective, Boddington [18] presents a comprehensive exploration of the application of three major normative ethics theories (deontology, virtue ethics, and consequentialism) to AI, and issues that might arise.

We expand upon previous literature to address gaps concerning principles which were not included in previous reviews, and provide further detail about how ethical principles have been operationalised in AI and computer science.

We build upon previous research, especially that of Tolmeijer et al. [130], to collate a broader range of ethical principles discussed in the AI and computer science literature, summarising operationalisation principle by principle. There are three key aspects of novelty contributed by this article:

Section 2 explains our methodology in brief. This will be useful for future research seeking to expand the taxonomy of ethical principles by reproducing the methods used here. Section 3 explores our findings for which ethical principles have been proposed in AI and computer science literature. Section 4 examines how ethical principles have previously been operationalised, as well as steps practitioners seeking to operationalise principles should take. Section 5 identifies gaps and future directions for operationalising ethical principles in AI and computer science. Section 6 concludes the article.

We investigate the current understanding of ethical principles in AI and computer science, and how these principles are operationalised. Specifically, we address the following questions:

We conducted an initial search on 2022-May-23, a second search on 2023-January-14, and a third search on 2024-February-01. The first search produced 3.74 million results on Google Scholar and 998,613 results on the University of Bristol Online Library. Looking at the first five pages of results, we applied the inclusion and exclusion criteria, which led to around 10 to 20 studies from each resource. Closer examination of these works resulted in the identification of relevant citations which we incorporated into our review. The selection of these works was critiqued by a secondary researcher which helped to identify further relevant research. This resulted in 57 papers being included in the review. The second search resulted 10 more papers being included in the review. The third search identified a further 14 papers to include in the review.

We categorise papers identified in our review based on contributions of the paper, type of evaluation, and ethical principles explicitly mentioned. Expanding on previous work, we adapt the taxonomies of Yu et al. [141] and Tolmeijer et al. [130] to categorise papers by contribution and evaluation type. We categorise papers by which principle(s) they explicitly mention with the exception of three papers. Jiang et al. [75] and Shi et al. [119] do not explicitly state ethical principles, but utilise the dataset of Hendrycks et al. [68] which consists of scenarios based on ethical principles. We include the work of Jiang et al. [75] and Shi et al. [119], as they provide valuable demonstrations of how ethical principles can be implemented. In addition, Noothigattu et al. [102] do not explicitly state ethical principles, but they demonstrate how inverse reinforcement learning, a valuable technique for learning ethics through observation, can be used to implement ethical rules (where rules are integral to deontological approaches) in machines. Tables 1 and 2 display the categorisation of papers.

Based on principles explicitly mentioned, works are broadly categorised into 11 key principles (deontology, egalitarianism, proportionalism, Kantian, virtue, consequentialism, utilitarianism, maximin, envy-freeness, doctrine of double effect, and do no harm). We categorise six types of contribution: descriptive, model representation, individual ethical decision making, centralised collective ethical decision making, decentralised collective ethical decision making, and ethics in human-AI interaction. Descriptive papers abstractly evaluate how normative ethics relates to AI. Model representation examines how to appropriately represent ethical knowledge in a system, or what features an ethical system should include. Individual decision making examines how individual agents may judge or select their own actions or the actions of others. Centralised collective decision making involves a central mechanism which makes decisions concerning multiple agents. Decentralised collective decision making involves multiple agents making distributed ethical decisions, such as in multi-agent systems (MAS). Ethics in human-AI interaction investigates ethical considerations of agents designed to influence or work in conjunction with humans.

We categorise four evaluation types adapted from Tolmeijer et al. [130]: test, proof, informal, and none. Test involves empirical analysis, comparing system outcomes against some ground truth. Proof examines if the system behaves according to some known specifications, typically using logic. Informal compares the system to example scenarios or application domains. When none of these evaluation types pertain, papers are categorised as ‘none’. Appendix A.2 provides further details of the methodology for paper classification.

We find that certain principles, such as utilitarianism, are more commonly discussed than other principles such as do no harm, as can be seen in Table 2. We also find that there is a large amount of research referencing ‘deontology’ and ‘consequentialism’ as broad terms, but not specifying what types of deontology or consequentialism they are referring to—for example, the work of Anderson and Anderson [6], Cointe et al. [30], and Greene et al. [61]. Precisely stating the ethical principles used (e.g., type of deontology) would allow for more exact operationalisation.

In terms of contribution, we find that a large majority of works utilise ethical principles in descriptive and model representation papers. Descriptive papers include overviews of ethics, such as the work of Boddington [18], or ethical critiques with reference to ethical principles, such as the work of Heilinger [66]. Model representation harnesses ethical principles to examine which ethical features should be considered in models, such as the work of Anderson and Anderson [5], Dignum [40], and Lee et al. [87]. For individual decision making, we find that more works implement consequentialist principles, especially utilitarianism, than deontology or virtue ethics. The majority of these works—for example, those of Dehghani et al. [35], Svegliato et al. [129], and Ajmeri et al. [2]—also provide tests. In centralised collective decision making approaches, we find more works implementing consequentialist principles rather than deontology or virtue ethics, such as those of Bakker et al. [13], Lera-Leri et al. [88], and Patel et al. [105]. For decentralised collective decision making, we find that most works implement deontology, as seen in the work of Greene et al. [61], or utilitarianism, as seen in the work of Governatori et al. [58]. In human-AI interaction, we find more works implementing deontological principles or virtue ethics rather than consequentialism, such as those of Pflanzer et al. [107], Hagendorff [63], and Anderson and Anderson [6]. We find decentralised collective decision making and human-AI interaction involved in the least number of papers, suggesting avenues for future work in these areas.

Deontology entails conforming to rules, laws, and norms, and respecting relevant obligations and permissions that stem from duties and rights [30, 62, 112]. For deontological theories, the permissibility of action lies within the intrinsic character of the act itself [18]. An action is permissible if and only if the act itself is intrinsically morally good, independent of the outcome [90].

To implement deontological theories, a rules-based approach may be used to provide moral orientation in identifying appropriate actions [66]. An example of a rules-based approach is found in the work of Limarga et al. [89], who use predicates to encode rules and then reason about different types of actions. Berreby et al. [15] implement different deontological specifications as rules in a ‘model of the right’. The model of the right is used to generate a ‘rightness assessment’ of available actions, considering context. The model contains deontological principles in conjunction with consequentialist principles. Similarly, Pflanzer et al. [107] suggest implementing deontology as part of a model which utilises consequentialism and virtue ethics, in which the role of deontology is to analyse the intention of actions. Tolmeijer et al. [130] argue that deontology could be implemented by inputting the action, using rules and duties as the decision criteria, and then mechanising actions via the extent to which they fit with the rule.

Deontology has been applied to different contexts. Binns [16] uses deontology to choose between incompatible fairness metrics, whereas Leben [85] applies it to evaluate distributions of binary classification algorithms. Hendrycks et al. [68] implement deontological principles in contextualised scenarios to measure the ethical knowledge of natural language processing models. Jiang et al. [75] use this dataset to test a model trained on people’s moral judgements. Some works suggest using deontology only in specific circumstances: Dehghani et al. [35] choose to implement deontology in situations with ‘sacred values’, selecting the action that does not violate a sacred value.

However, issues may arise when applying deontology. One common concern is that because deontological approaches focus on the intrinsic nature of an action, they fail to take the most likely consequences into account. Focusing solely on the intrinsic nature of action makes it challenging for deontology to adequately capture complex ethical insights [1, 113]. The complexity of ethical insight entails that any system of rules requires some interpretation and understanding of background assumptions. This means that the same rules might be interpreted differently in different contexts or by different people [18]. In addition, rights-based ethics revolve around decisions based on the rights of those who are affected by the decision. Focusing on rights is less helpful in situations where rights are not impinged, yet some sort of ethical dilemma is still occurring. For example, spreading hate speech does not necessarily infringe the rights of others, and there are arguments that preventing it infringes the right to free speech. However, there is an intuition that hate speech is wrong. Issues may arise with implementation when exceptions to rules emerge. Rules are expected to be strictly followed, implying that for every exception they must be amended, which could result in very long rules. Determining the right level of detail is thus important to ensure interpretability for the machine [130]. Lastly, there may be conflicts between rules. Conflicts may be addressed by ordering or weighing the rules, but this gives rise to difficulties in determining the order of importance.

According to virtue ethics, ethicality stems from the inherent character of an individual and not the rightness or wrongness of individual acts [141]. Right action is performed by someone with virtuous character. In following this theory, one should not be asking what one ought to do but rather what sort of person one should be [113]. The qualities one possesses should be of primary importance and actions secondary. Moral virtues can be learnt and developed through habit and practice. The stability of virtues (if one has a virtue, one cannot behave as if one does not have it) entails that virtue ethics may be a useful way of imbuing machines with ethics [134].

Virtue ethics can be used to formulate ideals for the use of AI, or to create AI which is virtuous itself [66]. To improve ethical use of AI, Robbins and Wallace [110] advocate for applying virtuous characteristics to resolve problems. Vanhée and Borit [131] propose that this may be aided by using education to help designers of systems develop virtues. Hagendorff and Danks [64] advance this by delineating that teaching virtues involves imparting tacit knowledge, social perception skills, and emotion, leading to the automatic ‘feeling’ of the right thing to do.

Other works focus on implementing virtues directly into machines; according to Tolmeijer et al. [130], inputs for implementing virtue ethics in machines would be properties of the agent, the decision criteria would be based on virtues, and this would be mechanised through the instantiation of virtues. Instantiating virtue ethics in automated decision making is exemplified by Govindarajulu and Bringsjord [59], who define virtues as learnt by experiencing the emotion of admiration when observing virtuous people, then copying the traits of those people. Computational formal logic is used to formalise emotions (in particular, the emotion of admiration), represent (virtuous) traits, and establish a process of learning traits. Greene et al. [61] argue that a virtue-based system would have to appreciate the entire variety of features which call for one action rather than another in a given situation.

Virtue ethics can also be used alongside other approaches. Hagendorff [62] argue that deontological approaches should be combined with virtue ethics, using virtues to examine values and character dispositions. Pflanzer et al. [107] suggest implementing virtue ethics to assess the character of an agent in a model which also utilises deontology and consequentialism. Hendrycks et al. [68] implement virtue ethics as part of an assessment criteria, integrating scenarios demonstrating virtue ethics into a dataset used to measure the ethical ability of natural language processing models. Jiang et al. [75] use this dataset to test their model trained on people’s moral judgements of various situations.

A problem with virtue ethics highlighted by Saltz et al. [113] is that the holistic view it takes makes it more difficult to apply to individual situations. Tolmeijer et al. [130] identify further challenges relating to the concretion of virtues and conflicting virtues. To judge whether a machine or human is virtuous is not possible by just observing one action or a series of actions that seem to imply virtue—reasons behind actions need to be clear. Requiring understandings of reasons behind actions makes it difficult to build virtues into machines, as there is a high level of abstraction in defining virtues. Additionally, conceptions of virtues can change greatly across time and culture. Virtues instantiated in machines today may lead to unfair outcomes in the future as virtues change, or certain virtues may conflict with other virtues.

In consequentialist approaches, right actions are identified through their effects [23]. The moral validity of an action can thus be judged only by considering its consequences [112]. Consequentialist principles can be used to weigh risks and opportunities [66]. A strength of this is that it can be used to evaluate decisions with complex outcomes where some benefit and some are harmed, by examining how benefits and harms are distributed. It can thus explain many moral intuitions that trouble deontological theories, as consequentialists can say that the best outcome is the one in which benefits outweigh costs [121]. In addition, Boddington [18] asserts that the goal-based nature of consequentialist theories suits computing well.

Consequentialist principles can be operationalised by analysing the consequences of different actions. Assessing ethics through consequences denotes a different approach to deontology, which regards mental states as very important for determining the ethicality of an action. For consequentialism, Tolmeijer et al. [130] denotes that mental states can be largely disregarded. Pflanzer et al. [107] propose a multi-theory model in which consequentialism analyses the consequences brought about by a situation. Deontology and virtue ethics are used in conjunction with consequentialism to make ethical judgements.

Consequentialism has been implemented by weighing actions. Limarga et al. [89] assign each action a weight according to its worst consequence. Actions are part of a sequence to reach a goal, and their weights accumulate to a total amount. This total amount is optimised to select the sequence with the best overall consequence. Suikkanen [126] similarly suggests ranking agents’ options in terms of how much aggregate value their consequences have. An option is right if and only if there are no other options with higher evaluative ranking. Tolmeijer et al. [130] argue that input for consequentialist principles would be the action (and its consequences), and the decision criteria would be the comparative well-being. This would be mechanised by selecting the consequence with maximum utility. For binary classification algorithms, Leben [85] suggests implementing consequentialism by examining how weights are assigned to each group outcome based on relative social cost.

However, in practice, assigning weights to each outcome may be unrealistic to do for all outcomes [85]. There might be high computational costs if machines attempt to represent all possible outcomes available [61]. A related issue is that estimating long-term or uncertain consequences and determining which consequences should be taken into account is difficult [18, 48]. There may be moral constraints outside consequentialism which prohibit certain actions even when they have the best outcomes, therefore rendering consequentialist theories incomplete [126]. Another common criticism of consequentialism concerns deciding what is valuable or intrinsically good: whether it is pleasure, preference satisfaction, the perfection of one’s essential capacities, or some list of disparate objective goods (knowledge, beauty, etc.) [18, 23, 130].

In addition to the principles mapped out here, there are other principles mentioned in the literature which we now describe. For reasons that shall be stated, we did not include these in the taxonomy.

A variety of technical implementations have been used to encode ethical principles. Expanding upon the categorisation of Tolmeijer et al. [130], approaches to encode principles into a format computers can understand include logical reasoning, probabilistic reasoning, learning, optimisation, and case-based reasoning. In Table 6 and Table 7, we map each ethical principle found in the literature to their technical implementations.

To engineer morally sensitive systems, Wallach et al. [134] argue that practitioners must decide on the architecture for integrating ethical principles. These fall within three broad approaches: (1) top-down imposition of ethical theories, (2) bottom-up building of systems with goals that may or may not be explicitly specified, and (3) hybrid approaches which combine top-down and bottom-up features. We discuss examples of each architecture and issues which may arise. Table 8 summarises our findings of the technical implementation of ethical principles according to the various architectures.

Practitioners should specify which ethical principle(s) will be operationalised. This could be aided by referring to the taxonomy we have suggested, which contains a broad array of ethical principles found in AI and computer science literature (see Figure 1). Leben [85] emphasises that being clear about which principle is being used will help designers further clarify what inputs are necessary for their application, which in turn will improve ethical reasoning capabilities and explainability of how decisions have been made.

We have found that abstract implementation of principles falls into three main categories: rules, consequences, or virtues. We discuss examples of each implementation category and potential difficulties that may arise. Deontological principles have been operationalised by applying rules and choosing an action based on how it accords with those rules. Virtue ethics has been operationalised by developing virtuous characteristics. Consequentialist principles have been operationalised by evaluating consequences and choosing an action based on the consequences it produces.

Understanding strengths and weaknesses of various approaches improves critical understanding and constructive engagement [18]. Therefore, key gaps include research on lesser-utilised principles. We suggest that future directions consider less commonly seen principles, or incorporate a wider array of principles. This includes researching principles from other cultures outside of the Western doctrine, which is important as ethics is culturally sensitive [69]. Implementing ethical principles from various cultures will aid the accessibility and fairness of AI, as it can better apply to stakeholders from diverse backgrounds. Hongladarom and Bandasak [71] survey non-Western guidelines for AI principles, finding unique cultural presuppositions in some areas and global consensus in others. Expanding this research to examine AI and non-Western principles from normative ethics is a future research direction.

We identify various difficulties with the implementation of ethical principles that may result in ethical dilemmas, from which various gaps arise. Anderson and Anderson [5] define ethical dilemmas as situations where either there is not a good choice between different outcomes or where the choice between different outcomes is not obvious (e.g., the distinction of how good one outcome is compared to another is not obvious). Future research could address gaps that arise in resolving these dilemmas.

In certain situations, dilemmas arise when the application of one principle cannot support one action over another. Azad-Manjiri [11] suggest that one way to resolve this could be by examining how similar decisions were made previously. If no similar decisions have been made previously, an action is selected at random. However, this approach may run into the naturalistic fallacy, looking at what is the case rather than what ought to be the case. In addition, relying on random choice may not result in the most ethically appropriate action. A gap exists in further examining how to resolve dilemmas where principles cannot support one action of another.

For each principle, dilemmas arise when its application leads to an unfair outcome, as all moral theories have some counterintuitive implications [111]. This implies that no single theory can denote how to program ethical AI [104]. Pluralist approaches, in which different principles can be weighed against one another to find the most appropriate answer, could help mitigate these issues. Works such as those of Governatori et al. [58], Lera-Leri et al. [88], and Pflanzer et al. [107] provide methodologies accommodating multiple principles. A gap exists in applying such methodologies to compare the application of multiple principles in scenarios where particular principles lead to unfair outcomes. However, weighing alternatives may not always be possible. Future research should investigate the feasibility of applying different principles in diverse scenarios.

Dilemmas may arise with the application of multiple principles, as different principles can give different answers which may conflict [106]. This is exemplified in the work of Nashed et al. [100], who find that agents implementing different principles favour different policies. In addition, it is difficult to apply abstract theories to concrete situations. To aid this, Tolmeijer et al. [130] suggest that particularism (which incorporates relevant contextual factors in ethical reasoning to identify if a certain feature is morally relevant or not) could help identify which principle is the most appropriate in that setting. A gap exists in exploring if aspects of particularism can be used to resolve dilemmas where different principles promote conflicting outcomes. However, there are also issues that arise with the application of particularism. Whilst ethics examines principles that socially impose what is right or wrong, morality deals with social values of right or wrong [75]. Moral disagreement can arise when stakeholders have different beliefs about which facts are morally relevant, and which ethical principle is true [10, 111].

There are thus various gaps and difficulties which arise in regard to resolving ethical dilemmas. Drawing these ideas together, there are gaps in finding reliable methodologies for AI practitioners to decide which principle is most appropriate for a particular case, considering the dilemmas which may arise. Robinson [111] explores three solutions which may help to resolve dilemmas: moral solutions, compromise solutions, and epistemic solutions. Moral solutions select a moral theory either by what we think is true, some general theory which we can agree on, or what we could hypothetically agree on under certain disagreements. Compromise solutions choose principles based on a social choice approach, or treat principle selection as a multi-objective optimisation problem, optimising based on inferred moral values or goals. Epistemic solutions harness information about the disagreement as evidence of moral facts, then appeal to a rule for decision making under moral uncertainty. Alternatively, epistemic solutions could attempt to achieve an overarching moral view which accommodates as many relevant ethical judgements as possible. Each approach has various strengths and limitations, as discussed in the paper. Robinson [111] concludes that problems of moral disagreement should be treated as problems of managing moral risk, where moral risk is the chance of getting things wrong and what you thereby risk. A gap exists in implementing such solutions to evaluate how they address ethical dilemmas in practice.

An application of implementing ethical principles in STS is to support governance capacities. Governance of STS involves establishing standards for the proper use and development of technology, as defined by Floridi [52], and administration of systems by stakeholders themselves, as defined by Singh [120]. Under the perspective of macro ethics, responsible governance should incorporate norms and value preferences of different stakeholders. However, dilemmas arise when norms or values conflict. Operationalising ethical principles in reasoning helps support governance capacities to resolve these dilemmas in equitable ways that support the needs of different stakeholders [138]. Gaps exist relating to how ethical principles can be operationalised in STS to promote equitable governance capacities.

Previous work provides guidance for applying ethical principles to reason about values and norms in computational decision making. For example, Ajmeri et al. [2] broadly reference the principles of egalitarianism and utilitarianism within the context of utilising values and norms in MAS for ethical reasoning in individual decision making. This research may benefit from the consideration of other ethical principles to enable broader applicability. Lera-Leri et al. [88] present a method for applying multiple ethical principles to aggregate different value preferences, but they do not consider the influence of norms. Serramia et al. [117] demonstrate how to select norms that best align with a known value system. Combining these approaches to aggregate different value systems using ethical principles, and then using the aggregated value systems to select value-aligned norms, is a promising direction for future research. In addition, a gap exists in examining how such approaches can be incorporated in decentralised collective decision making.

However, challenges arise when implementing ethical principles in STS considering value systems. Norms and values are interdependent with context and decision makers, and research should consider how value systems change according to context, for example, over time [103, 122]. Gaps exist related to implementing principles in STS in ways that account for the relationship between context and changing value systems.

Properly incorporating broad social context requires careful consideration to avoid entrenching dominant relations of power [136]. This includes accounting for the ways in which existing dynamics shape how technology is developed and deployed. Applying ethical principles must therefore integrate broad social dynamics, as well as appreciate how social dynamics affect governance capacities [97]. Gaps emerge with respect to accommodating broad social dynamics and avoiding perpetuating unjust power dynamics in the application of ethical principles to governance capacities.

To understand how ethical principles can accommodate for broad social dynamics, participatory approaches may be useful to incorporate human input throughout the design process. For example, Dubljević et al. [43] combine participatory approaches with multi-criteria decision making to capture the importance of different harms and make clear the perspectives of different stakeholders. Weinberg [136] emphasises that collaborating with those affected by the technology improves the propensity to leverage knowledge from marginalised groups, understand how the technology is situated in its social context, and address what is most ethically concerning, rather than what is most convenient to measure. Participatory approaches present opportunities to investigate questions related to what extent ethical principles are generalisable across different groups of people, what people morally disagree on, what preferences people have over ethical principles, and if and how people follow ethical principles in their daily lives. Gaps exist in further examining these questions.