1 Introduction

The current ‘buzz’ around AI is slowly turning away from a ‘rush’ for ethical frameworks to discussions surrounding how governments and intergovernmental organisations should legislate the development and use of AI [40, 70]. Whilst there are differences in the approach to what exactly AI regulation should look like, the core understanding is that we need to ensure that AI should be used for good. For example the EU AI HLEG articulates that AI should enhance “individual and societal well-being and the common good” [36, p. 40]. The UK in the Bletchley Declaration denote that the design and use of AI needs to be “for the good of all” [18] and President Biden’s executive order on secure and trustworthy AI stresses the importance of “harnessing AI for good” [23]. But what do we actually mean when we say that AI should be used for good and what are the ethical implications of labelling AI as its facilitator? Many AI researchers have argued for AI for good [78] or AI for social good (AI4SG) [25,26,27] as an approach to developing AI, yet it is often unclear what good means in such projects [16, 32].

The purpose of this article is to unpack the philosophical meaning behind AI for good, arguing that it functions as a powerful normative concept within AI ethics scholarship, governance and policy. As a concept, AI for good expresses a value judgement not only on what value AI has, but also the functional role it plays in society and it is in this sense we can think of it as a normative concept itself [22, p. 64]. Furthermore, as a moral statement AI for good makes two things implicit: (i) we know what a good outcome is and (ii) we know the process by which to achieve it. The article will unpack these two claims by posing them as problems, arguing that AI for good belies an understanding of the value and process of good that is difficult to achieve and hard to articulate. Statement (i) will be addressed as the normative problem and statement (ii) will be addressed as the metaethical problem. Overall this approach will highlight, firstly how AI for good functions as a normative concept in AI ethics, secondly the problem this presents and finally how AI ethics needs to examine more clearly the relationship between the normative concepts it uses and their metaethical implications. In addressing both statements, keen attention will be paid to how AI for good can be critiqued using foundational ethical theories, questions and arguments from normative ethics and metaethics.

The layout of the article is as follows: Sect. 2 will address AI for good as a normative concept by looking specifically at how organisations such as AI for social good (AI4SG) have developed its structure alongside some of its criticisms. Section 3 will examine claim (i) as the ‘normative’ problem of AI for good. Using AI4SG, this section will question the development and justification of both the intrinsic value of good used as well as the instrumental value of AI in achieving it. Section 4 will examine claim (ii) as the ‘metaethical’ problem of AI for good, articulating how current approaches to the function of value represent metaethical stances that are not only hard to maintain, but that are problematic for how the function of ethics is seen to operate. The final section will argue that more critical reflection needs to be given to normative and metaethical assumptions in AI ethics, due to their importance in justifying moral claims and the nature of AI ethics as a practice in general.

2 AI for good: a normative concept or an organisation?

How we frame the concept of good within a narrative of ethical AI has important ramifications for the role it plays in literature, legislation and governance of AI [9]. This section will explain the relationship between AI for good as a normative concept to its various uses in scholarship and organisations. The purpose is to contextualise where the concept has most been developed and some of the criticisms it has faced.

As highlighted in the introduction, AI for good is a concept that is prevalent not only in government policies, but also as a global research focus through what is often referred to as AI for social good (AI4SG). In the past 14 years AI4SG literature and research projects have seen a massive rise [68] and provide the best example of how AI for good as a concept functions within AI ethics discourse. Yet, often it can be difficult to delineate when AI for good/social good is a research topic or an organisation. For example, Google has its own AI for social good lab that describes itself as “a group of researchers, engineers, volunteers, and other people across Google with a shared focus on positive social impact” [31]. Furthermore, there is also an AI4SG organisation that encourages students to “develop AI powered solutions that address social issues in their communities” [1]. In both instances AI4SG is a project for developing AI that promotes social good but also an organised structure that advances its cause. One of the main criticisms by scholars is that these projects and organisations benefit large technology companies as a moral stamp of approach for their AI development, without being clear what exactly they mean by social good [49, 79]. Indeed, governing bodies, international organisations and technologies companies have differing agendas in the development of ethical AI and in a very real sense undermine the meaning of good altogether [37].

For the purpose of clarity in the next section AI4SG will be examined in its capacity as a research movement that focuses on developing AI that furthers socially good outcomes [26]. Aligning social goods with the United Nations sustainable development goals (SDGs), AI4SG uses SDGs as a way of benchmarking how to know if a social good outcome is achieved [17, p. 113]. As an example of AI for good as a normative concept in use, AI4SG also highlights the argument that AI for good as a normative concept makes implicit that (i) we know what a good outcome is. What is important to understand in the context of this article is that AI for good as a normative concept runs through all of these projects, organisations and policies, promoting the notion that we can make AI ethical if the outcome of its development and use furthers good.

3 The ‘normative problem’ of AI for good

AI for good expresses a value judgement on AI’s capabilities and implicitly assumes a certain understanding of what constitutes a ‘good’ outcome. This section argues that the normative problem of AI for good is that it makes implicit that moral statement that i) we know what a good outcome is. As shall be shown the normative foundations of this claim can be interrogated through their commitment to the intrinsic value that AI for good has as well as the instrumental value AI has in achieving it.

3.1 The intrinsic value of AI for good

Within AI ethics literature, a popular view on the concept of technology and its ethical role is the consequentialist notion that AI will have societal benefits [6, p. 228]. In normative ethics, consequentialist theories judge our moral decisions based on if they increase (directly or indirectly) a good outcome [2]. The normative force of consequentialism is underlined with a theory of the value of good that aids the ethical judgement [21, p. 32]. For example, in utilitarianism, wellbeing is seen as the definition of what a good outcome will be and all actions that achieve the most amount of wellbeing are optimific (e.g. an action that is morally required because it produces the best results) [67, p. 122]. In utilitarianism, wellbeing is understood as an intrinsic good i.e. that which is good for the sake of good in its own right and defining an intrinsic good is vital for a consequentialist theory.Footnote 1

We can think of the claim that (i) we know what a good outcome is as a ‘normative problem’ of the AI for good. This is because there are two distinct aspects relating to the judgement of the moral value of a good outcome in this claim that are hard to reconcile. First is that the approach relies heavily on the outcome of AI design and development being for good, focusing more on the consequence rather than any set principle behind it. Whilst a common critique of consequentialism is that it focuses on outcomes [46], in the context of AI for good it becomes more concretely problematised by the fact that the intrinsic value of good by which we benchmark AI for good is never fully identified or loosely justified. To exemplify this let us turn to how Cowels et al. [17] define an AI for social good project;

“… as the design, development and deployment of AI systems in ways that help to (i) prevent, mitigate and/or resolve problems adversely affecting human life and/or the wellbeing of the natural world, and/or (ii) enable socially preferable or environmentally sustainable developments, while (iii) not introducing new forms of harm and/or amplifying existing disparities and inequities” [17, p. 113]

In this understanding it is the outcome of design and development of AI that is a key factor in its ability to promote ‘social good’ and therefore consequentialist in nature. Implicit in this understanding is that to some extent the intrinsic value of social goods, i.e. the well-being of the natural world or socially preferable outcomes are self-evident. On a theoretical level, AI4SG scholars have attempted to align social goods with the UN SDGs as a way of benchmarking something that is to some extent agreed upon internationally [17, p. 113]. SDGs themselves are not explicitly social goods, but a broad range of areas that have been agreed on as social challenges that humanity faces overall [34]. As an intrinsic value the definition of social goods, even if they are equated to SDGs, presents a normative problem for statement (i). This is because, even though we have a consequentialist understanding of how moral judgments are made, the intrinsic value we seek to benchmark this judgement has two flaws. First is that basing social goods on SDGs does not provide a clear definition of the intrinsic value of social goods overall. SDGs themselves have been critiqued for their oversimplification of the challenges that face humanity and for their loose commitment to social welfare [4]. Using such a vague and unclear benchmark for intrinsic value puts the consequentialist approach to AI ethics in a perilous place. Without a clear definition or dialogue on the intrinsic value that AI furthers, a good outcome becomes more of a “know it when you see it” approach [32, p. 1].

Furthermore, if we think about the practical implications of this approach, the normative problem becomes tangled in the current definition that is being upheld. In the case of AI4SG, the evaluation of a social good that is both environmentally friendly and does not introduce more harm is almost impossible. We know, for example, that AI databases need a huge amount of energy to be maintained and leave a large carbon footprint—so how can any project be truly for social good? [14]. Indeed, both the practical and theoretical issues with the normative problem put the onus on knowing what a good outcome is, on the individuals carrying out the projects. As Floridi et al. confess [27], it is not clear when you are “obliged or obliged not to, design, develop, and deploy a specific AI4SG project” [27, p. 1791).

Whilst the focus of this section has been on AI4SG, the normative problem of AI for good as a concept in general is applicable to many instances in which AI is argued to be used for good. Fundamentally, to enact AI for good there is an imperative on the ethical judgement being the outcome. In turn, this means that the intrinsic value of the good we seek to further needs to be addressed. Whilst this is a noble task, G.E Moore famously said that defining good was the most important task of ethics [48, p. 3], the normative problem of AI for good encapsulates how difficult this is to achieve. The risk is that we set up the work of AI ethics to achieve something that over 2000 years of philosophical debate has not been able to achieve; agreeing upon a universal intrinsic good. Indeed, part of the problem of not knowing when you are obliged or not to create an AI4SG project, is chiefly because the intrinsic value in which the outcome is to be defined is lacking in structure.

There is, however, another dimension to the ‘normative problem’ with AI for good that is specific in this case to the AI systems themselves. In the next section the instrumental value of AI as a means to achieve good will be shown to impact the implicitness of knowing what a good outcome is through our understanding of the AI systems themselves.

3.2 The instrumental value of AI for good

Understanding what AI is good for helps frame any normative sentiment on the value judgement of AI as achieving good. That is to say, how an AI system is presented drastically affects the context in which its ethicality is discussed and the efficacy of it producing a good outcome. In value theory, instrumental valueFootnote 2 is broadly defined as something that is a means of achieving something that is good intrinsically [19]. Certainly, intrinsic value will guide the instrumental value of something; this is because, in a consequentialist way, we are directed towards the things that maximise the good. However, AI systems complicate this relationship due to the fact that the ethicality of their use is bound up in the context and definition.

Crucially, AI systems, from a designer point of view, tend to view the normative force of AI for good through a scientific approach that ‘good’ is a functional/technical solution to a problem, not a social or ethical one [7]. The ‘AI’ in AI for good is assessed through its efficacy at providing a service that suits its function and the more congenial AI is in ‘solving’ that problem, the more value it has in furthering ‘good’. For example, in areas such as cancer research, AI has been heralded for its breakthrough in breast cancer screening accuracy [13]. As an object that maximises a function that has value to us, we could say that in this case, AI has instrumental value. We are directing our normative assessment of instrumental value, by judging if AI increases accuracy at detecting cancer and therefore it increases a good outcome. What is a problem for AI for good, is that we take this understanding of an individual case and then we attempt to universalise this notion for the normative concept as a whole.

In such a way, the consequentialist notion that AI should promote good leads to an ethical argument that, if we have the correct principles, then we can create AI systems that are a benefit to humanity. Take, for example, the argument that beneficence should be a core ethical value in guiding the use and design of AI [25, 26, 33]. As a bioethics principle, beneficence represents a commitment to promoting wellbeing and as an ethical principle in general,it represents a core aspect of AI for good literature [26, p. 679]. In the intrinsic sense, this represents an oversimplification of the promotion of well-being [50]. In the instrumental sense, the notion that AI’s value in achieving such well-being can ‘overhype’ the prospects of what AI is able to achieve [55, 75]. AI for good as a normative concept relies on the fact that AI is a means to achieving good and therefore supports the claim that we know what a good outcome will be. Whilst beneficence may be a key principle in bioethics, where research is primarily concerned with human wellbeing, AI research has no fundamental end goal as such, due to the varied nature of what the technology can achieve [74, p. 70]. In an attempt to overcome this, London and Heidari [44] proposed a mathematical criteria for when an AI system is aligned to benefitting society. However, while they identify important moral challenges like paternalism and exploitation, their model still faces difficulty in prioritising the interests of different stakeholders. The difficulty is that instrumental value in AI is not only about what principles or values help aid the instrumental value of AI itself, but how external influences also affect the notion of what AI systems are good for.

The obstacles that we face in operationalising principles such as beneficence point to the difficulty not just in abstract terms, but what counts for beneficial or harmful AI at all. The concept of AI itself is a highly contested and political term [76, p. 243]. Who or what is to determine the instrumental value of an AI system in achieving a particular notion of good, when the very concept of AI is so contested? Certainly a large factor in determining the instrumental value of AI systems are the technology companies themselves. If part of the instrumental value of AI is locked in the definition of the system itself, a factor that significantly influences that decision are the profit margins that a company designing the system takes into account when they create it [15, p. 743]. Technology companies have a significant influence in how the legislature designed to regulate AI defines not only the AI systems, but if they are a threat. For example, OpenAI CEO Sam Altman, an advocator for more AI legislation, recently lobbied EU officials against foundational models (including chat GPT) as being labelled dangerous (the highest threat level) in the EU AI act [54]. Weeks before the act was to be submitted to the European Parliament, France, Italy and Germany argued that foundational models should be stricken off as dangerous, resulting in a compromise in which the EU AI act would only regulate foundational models trained on a incredibly vast amount of data would be regulated, keeping the interests of current foundational models, safe from the act [5]. This example shows us that if a key factor to the instrumental value is not whether it furthers any intrinsic value, but the definition of the instrument itself, then whose good is AI for?

The normative problem of AI for good, then, is not only the vagueness of what good actually is, but the instrumental value AI has in achieving it. Certainly, as some scholars suggest, it would be easier instead to argue that AI ethics should not be for good, just not for bad [32, 49]. Significant to the overall problem of AI for good are the ways in which its normative features represent moral judgements of what a good outcome is. It seems that the nexus of AI for good as a normative concept lies not only with understanding what a good outcome will be, but the process in which it arises. As such, it is the way in which the role of value itself becomes a direction for the ethics of AI that makes it a metaethical problem, deeply intertwined with its normative assertion.

4 The metaethical problem of AI for good

By questioning the normative structure of AI for good as a concept, it allows for a more detailed inquiry into the moral knowledge this concept aims to represent and the process by which it makes value judgments. In moral philosophy, questions on the role of moral knowledge in normative claims have often been the foci for metaethical analysis because they reveal assertions to how we should conceive of the work of ethics [35, 39, 64, 81]. It is in this sense that claim (ii) we know the process by which to achieve it represents the metaethical problem to AI for good. The metaethical issues that centre this claim are moral knowledge and process, both of which have important implications for not only how AI ethics envisions the role of ethics, but also how metaethical inquiry is important to AI ethics as a discipline.

Recently, AI ethics researchers have become concerned with the theory–practice gap, which questions how we put AI ethics into practice [8]. In the case of AI for good, how good is put into practice is represented through an understanding of the function of value in representing moral knowledge. The following sections will look at how values have been understood in AI ethics, paying attention to the value alignment problem and the distinction between the normative and technical aspects of the problem. Furthermore, it will contextualise the metaethical problem of AI for good by bringing in debates from metaethics about the representation of moral knowledge and its relationship with value.

4.1 What do values do?

In this article, the analysis of value has been mainly foregrounded by an examination on the normative, consequentialist understanding of intrinsic and instrumental value that fund AI for good as a normative concept. In order to substantiate the claim (ii) is a metaethical problem, it is important to first explore more clearly how moral philosophy tackles the function of value and its relation to AI ethics discourse.

In value theory there are two main ways in which the process of values come to an ethical judgement, evaluative and deontic. Evaluative value refers to concepts that relate to goodness and badness, as opposed to deontic values which relate to the rightness or wrongness of actionsFootnote 3 [85, p. 13]. These functions of value commonly relate to normative ethical theories, evaluative values are mainly consequentialist and deontic values are deontological [85]. When we make a decision on what the process of values are, be it evaluative or deontic, we make a metaethical commitment to how moral knowledge is represented as a functional part of ethical evaluation. We will unpack this in more detail in the next section, however, what is important is the notion that; what we think that values do matters. This is true not only for AI for good as a normative concept but for AI ethics as a whole.

Take, for example, the value alignment problem and how it is conceptualised. As a problem, value alignment is understood as figuring out how to teach/train autonomous AI systems to make decisions that are aligned with or beneficial to human values or preferences [24, 61, 62, 71, 72]. Commonly, value alignment is partitioned into two research focuses: the technical and normative problem, the former concerns how to encode certain values within an AI system and the latter looks at which or whose AI should be aligned with [29]. Discussions on the ‘technical’ side of alignment tend to focus on machine learning and assessment techniques [38]. Debates on the normative side tend to focus on which or who’s value systems to use [29]. How we view the process of values matters considerably in this research field. This is exemplified by the fact that more recently, those on the ‘normative’ side of the debate are increasingly arguing for a ‘pluralist’ approach to values where there is no overriding ‘super-value’ to which all values are reduced [73, p. 19938]. As those in the pluralist camp see it, the problem with value alignment is that for an AI product to be fair and ethical we need to have a value system that suits the multiple context in which AI is used and that it is capable of making value judgments that are not absolutist, but still morally preferable [29, 73, 77]. The ‘technical’ side of value alignment is becoming more influenced by pluralism and there is a growing shift in the literature by those attempting to design an AI value system that can be pluralist as well as evaluative [43, 58]. In this sense, what values do matters to the ethics of AI in a fundamental metaethical way, because it represents how we are to conceive of AI systems making moral judgements. It affects both our understanding of the normative framework of values as well as the technical challenges we face in creating AI systems.

4.2 The process of value in AI for good

It is not simply that the process of values matters for how we understand value judgements, but how the process of value represents moral knowledge of those judgments. If we consider claim (ii) that AI for good as a moral statement makes implicit: we know the process by which to achieve it (good), the process of value and the knowledge it represents becomes a question of how we are to understand the relationship between normativity and moral knowledge. Because AI for good features heavily in AI ethics scholarship and is prevalent in AI policy, the concept does not simply imply that we know what a good outcome will be, but designates what the process of ethics should look like to achieve it.

As a normative concept, AI for good understands values as evaluative. This is because our attention is focused on deciding what a good outcome will be. AI for good is concerned with attributing value as the process of deciding what is good and what is bad and this approach to values as evaluative represents a growing trend in AI scholarship [11, 20, 28, 58, 66, 80]. The metaethical problem with this is how moral knowledge is represented in evaluative values and the technical challenges this represents in designing AI systems that reflect this process. The difficulty lies in the fact that moral knowledge has to be represented in a way that a machine, whose function is effectively normative in nature, can process [82]. In the metaethical sense, this is what has been argued in machine ethics as a problem of understanding and defining both what teaching means as well as the practice of ethics [30]. Thinking back to the normative problem of intrinsic value we can see the relationship between a weak conception of what is of intrinsic value to us and what exactly that moral judgement represents as ‘ethical’. The crux of the metaethical problem is centred around designing and using an AI system that understands correctly not only what a good ethical outcome will be, but the interplay between the process of value and the moral knowledge it represents.

Furthermore, within an evaluative conception of value, the moral knowledge that is represented can be produced in two ways that impact the veracity of the concept of AI for good as a whole. This is because it is important that when we argue what moral knowledge represents (this good or that bad) we must also ask how we can come to make such a judgement? Does the evaluative process of AI for good as a normative concept represent an objective standpoint of morality or is it bound up in a cultural/subjective understanding? To explore this in more detail, the following sections will highlight the metaethical debates of how a subjectivist and objectivist account of moral knowledge represents a difficulty for AI for good.

4.3 What moral knowledge do values represent?

As a normative concept itself, AI for good represents a way of doing ethics and approaching moral knowledge.There are substantial debates in metaethics that problematise the relationship between values and moral knowledge, the most famous is the fact-value distinction, which asks if we can derive a moral fact from a value [57]. To lay out all the metaethical arguments of moral knowledge would take up far more space than is intended for the length of this article and often positions in metaethics are complex and overlap making their arguments hard to follow [56, p. 2]. Whilst there is scope for future work to discuss the fact-value distinction in AI ethics, the section will lay some groundwork for these debates by examining what the objectivist/subjectivist debate on moral knowledge can do for the metaethical problem of AI for good on knowledge representation. As such, the aim will be to explicate what these positions say about moral knowledge and its representation in order for the next section to take these debates and question what moral knowledge is representable in AI for good as a concept.

Central to the objectivist/subjectivist debate is whether the source of our moral knowledge and values is culturally and subjectively determined, or whether it represents objective truths independent of human understanding. This distinction on moral truths clearly matters for AI for good, because it gives credence to the normative evaluative judgement that it has intrinsic worth and as such AI has the capacity to further validate that judgement. A subjectivist account of values would assert that the moral knowledge they represent are preferential, they are mind-dependent knowledge only [12, p. 221, 83, p. 185]. Crucially these judgements on morality are that of taste, not moral facts about the world. We can differentiate this view of subjectivism with ‘second order’ subjectivism. This states that moral statements can be true or false, but depend on the attitudes or opinions of groups and as such they are held to people’s belief systems about morality [3]. Opposed to this view is objectivism, which states that values are not subjective, meaning they are either mind-independent, subject-independent or both. In essence all values are objective or some values are, the first would be a stronger position than the last [53]. As Brink notes, it is often tempting to assume that objectivism is a form of moral realism [12, p. 223], however this debate in itself is significantly nuanced, so, for the purpose of clarity objectivism simply stands in opposition to the notion that all values can represent cultural knowledge.

So, to exemplify briefly how these two concepts interact, a subjectivist understanding of evaluative value would argue that AI being for good is simply a cultural attitude and a preference. An objectivist would argue that AI really is for good, independent of any subjective stance. From this account of objectivism and subjectivism we can see that there is a precedent for how moral knowledge is represented in values. The relevance this has to claim (ii) is that if AI for good claims to know the process of a good outcome, then it also lays some claim to what moral knowledge this represents. This is indeed a metaethical problem of AI for good, although it is closely linked to the validity of its normative evaluative judgements. In the following section, AI for good will be revisited through a metaethical lens, highlighting some of the implications of subjective and objective moral knowledge representation in AI for good.

4.4 AI for good, subjective or objective?

In this section, we will pay more attention to how the metaethical problem of AI for good interacts with claim (ii) on knowing the process and through both a subjectivist and objectivist account. Some in AI ethics contend that values are culturally bound [11, 20, 28] and some have argued that values broadly represent a consensus on what good values are [17]. Because all theories of value are inherently about intrinsic value [12, p. 217], this section will deal with the notion that value in AI intrinsically represents moral knowledge that is either subjectivist or objectivist. The accounts here are meant to outline in more detail how moral knowledge within AI for good can be problematised through metaethical discussion on the foundation of knowledge itself.

4.4.1 A subjectivist account

A subjectivist account of values would mean that intrinsic value of good in AI for good as a normative concept consists in, or depends on, an individual’s (or cultural/social groups) contingent psychological states. Here, we can return to the important issue of whose preference has the final say, is it a technology company, the laws or regulations, human rights or government? What at first seems like a question of normativity, is ultimately a question of metaethical understanding of moral knowledge. Such an understanding undermines the authority of the intrinsic value ascribed to AI for good. This concept is further complicated because it forces a metaphysical distinction between who or what a cultural group is. Such distinctions pose serious issues for integrating information into AI systems that will be used globally. Given the inherent critique of power dynamics in AI ethics [59], a subjectivist account of intrinsic value would only worsen the abuse of those seeking to use AI to force potentially oppressive cultural opinions through the guise of ethics.

Crucially, on the technical side, we are no closer to understanding which value system is even best to design in an AI system. Aspects of the ‘normative’ problem of value alignment have struggled with competing ideals on what value systems AI should be encoded with. Although value pluralism is said to overcome this problem of subjectivity [77], it still remains to be seen how we are to design a valid hierarchy of moral knowledge, based on a system of preferential values. We could contend that on a more basic level, different societies could create different AI systems aligned with their subjectivist understanding of values. Yet, it does not account for what will happen when AI systems built on different cultural values interact with each other, what happens when we have AI systems that have moral disagreements?

4.4.2 An objectivist account

An objectivist account would mean that the intrinsic value in AI for good represents moral knowledge that is of objective value, independent of individual or cultural beliefs. In part, this is how AI4SG has positioned itself in relation to values. Cowls et al. argue AI4SG use the UNs SDGs because they are:

“..are internationally agreed goals for development, and have begun informing relevant policies worldwide, so they raise fewer questions about relativity and cultural dependency of values. Although they are of course improvable, they are nonetheless the closest thing we have to a humanity-wide consensus on what ought to be done” [17, p. 113]

On the surface, this seems to be a flexible solution to the difficulties in defining the intrinsic value of social goods. Certainly, some form of international agreement alludes to a universality of values and in theory offers a consensus on what “ought to be done”. However, the underlying assumption here is a universalisation of core values which alludes to a notion of value absolutism, where certain values are seen to be universalizable and ubiquitous in moral life [65]. Such an objectivist stance oversimplifies the contextual difficulties of understanding exactly how moral knowledge is affected by the use of AI. This is an area which has received little attention in the ethics of AI and before we make such considerations, it is important that these discussions are more fleshed out. Importantly, we are still no closer to understanding what a value system actually means. Where do we delineate the objective nature of evaluative statements as being what is good, independent of a group and when values are not merely an expression of preference?

4.5 The importance of metaethics to AI for good

Both the objectivist and subjectivist account highlight substantial difficulties in the metaethical problem of AI for good. How moral knowledge is represented in a value system significantly impacts what the normative force of a concept is. Furthermore, knowledge representation plays a crucial role in understanding the technical challenges we face in AI for good. The intrinsic value of good matters not only as a benchmark for ethical evaluation, but also for ethical representation. In this sense, metaethical analysis begins to illuminate what normativity means to the practical challenges of ethics. Fundamental to the normative and the metaethical problem of AI for good, is that these are challenges that have been continuously debated in moral philosophy and highlight that our approach to AI ethics must be far more innovative and conscious of these challenges before we make the statement that AI must be for good. It is not the purpose of this article to be pessimistic, bemoaning that ethical AI is not possible, yet it is vital that we do not uphold stances that are not only hard to maintain, but that are problematic for how the function of ethics is seen to operate. There is a sense that AI for good as a normative concept, focuses on defining what a good outcome is, in order to give AI’s capability a moral value [45, p. 853]. More care and attention needs to be paid to discussing the types of moral knowledge which AI systems will inherit, both in the practical use with humans and in the future when building AI systems that incorporate ethical theories. Whilst there has been some work done by scholars to bring our attention to the importance of metaethics in AI, as a form of research it still has a long way to go [69].

5 Conclusion: good going forward

This paper has argued that AI for good is a normative concept that is problematic for AI ethics. As a moral evaluation, AI for good makes implicit that: (i) we know what a good outcome is and ii) we know the process by which to achieve it. Claim (i) was analysed as the normative problem, arguing that while AI for good is often argued through a consequentialist way, the intrinsic value of good is unclarified and diffused. Moreover, the instrumental value of AI in achieving intrinsic worth was problematised further because what AI is good for is bound to its definition and context. Claim (ii) was assessed as a metaethical problem, focusing on how moral knowledge is represented in the function of values, through the metaethical concepts of subjectivism and objectivism. Value, in AI for good, is often thought to be evaluative, meaning if it was subjective it would be relative to the ends of different groups or individuals and what they consider to be good. Alternatively, the objective stance would mean that the deployment of AI systems would be a value judgement in which AI ought to be deployed. Indeed, it could be argued that AI for good is fundamentally bound to metaethics because, when faced with the physical aspect of designing and creating an AI system, we are forced to choose how knowledge of values is represented. What is at risk is that the process of value is distorted into a question not of what values are, but how they are represented. Value, both instrumental and intrinsic, in AI for good are often framed not in an attentiveness to how we understand the role of ethics, but how we should represent it in a physical space. The risk of this is that it diminishes the role of values to an assumed normative role, which left uncritically examined can lead to an arbitrary selection of values and understanding of normative ethics itself [47].

As a way forward, AI researchers should be encouraged to engage in metaethical debates about the role of moral knowledge and normativity as a way of critiquing our own values. One example could be Iris Murdoch, whose critical work on the sovereignty of good, illuminated an aversion to such unreflexive uses of good, arguing that good is “not an empty receptacle into which the arbitrary will places objects of its choice” [51, p. 507]. For Murdoch, understanding the concept of good required a moral perception that was not concerned with defining good outcomes or explaining what values are, but through a process of reflection and ‘unselfing’ oneself from the moral egoism and selfishness that is our human nature [60]. In this sense, Murdoch’s understanding of the sovereignty of good, goes against a universalist notion that is fundamental to how the normative concept of AI for good is conceived of as a whole [52]. Certainly, there is scope for future work in AI ethics on the relationship between normative concepts and a metaethical reflexive approach to what implicit notions they imply. The purpose of this article was to introduce both the normative and metaethical problems in AI for good and to argue that we need to ask more deeply in AI ethics, what does it actually mean to be good?