Many have recently argued that there are weighty reasons against making high-stakes decisions solely on the basis of recommendations from artificially intelligent (AI) systems. Even if deference to a given AI system were known to reliably result in the right action being taken, the argument goes, that deference would lack morally important characteristics: the resulting decisions would not, for instance, be based on an appreciation of right-making reasons. Nor would they be performed from moral virtue; nor would they have moral (...) worth. I argue that, even if these characteristics all have intrinsic value, that intrinsic value has no practical relevance to decisions about whether to defer to AI. I make that point by drawing on a lesson from the literature on moral testimony. Once it is granted that deference to a reliable source is the policy most likely to bring about right action, a refusal to defer carries with it a heightened risk of wronging and mistreating people. And that heightened risk of wrongdoing, I argue, cannot be justified by appeal to the intrinsic value of striving for a morally exemplary decision-making process. (shrink)

Dehumanization by algorithms raises important issues for business and society. Yet, these issues remain poorly understood due to the fragmented nature of the evolving dehumanization literature across disciplines, originating from colonialism, industrialization, post-colonialism studies, contemporary ethics, and technology studies. This article systematically reviews the literature on algorithms and dehumanization (n = 180 articles) and maps existing knowledge across several clusters that reveal its underlying characteristics. Based on the review, we find that algorithmic dehumanization is particularly problematic for human resource management (...) and the future of work, managerial decision-making, consumer ethics, hard- and soft-law regulation, and basic values, including privacy and consumer rights. From the literature synthesis, we also derive the following definition of algorithmic dehumanization: the act of using algorithms and data in a way that results in the intentional or unintentional treatment of individuals and/or groups as less than fully human, thereby violating human rights, including liberty, equality, and dignity. Ultimately, we present a dehumanization avoidance model that serves as a structured research agenda and practical guide to address the challenges raised by algorithmic dehumanization. Thus, the model indicates promising pathways for future research at the intersection of AI and society and facilitates reflection on organizational processes that support improving corporate capabilities and managerial decision-making to avoid algorithmic dehumanization. (shrink)

An AI-based “patient preference predictor” (PPP) is a proposed method for guiding healthcare decisions for patients who lack decision-making capacity. The proposal is to use correlations between sociodemographic data and known healthcare preferences to construct a model that predicts the unknown preferences of a particular patient. In this paper, I highlight a distinction that has been largely overlooked so far in debates about the PPP–that between algorithmic prediction and decision-making–and argue that much of the recent philosophical disagreement stems from this (...) oversight. I show how three prominent objections to the PPP only challenge its use as the sole determinant of a choice, and actually support its use as a source of evidence about patient preferences to inform human decision-making. The upshot is that we should adopt the evidential conception of the PPP and shift our evaluation of this technology towards the ethics of algorithmic prediction, rather than decision-making. (shrink)

This paper explores the complex nature of social biases, arguing for a functional framework that recognizes their unity and diversity. The functional approach posits that all biases share a common functional role in overcoming underdetermination. This framework, I argue, provides a comprehensive understanding of how all psychological biases, including social biases, are unified. I then turn to the question of disunity, demonstrating how psychological social biases differ systematically in the mental states and processes that constitute them. These differences indicate that (...) biases at various levels of the cognitive architecture require distinct treatment along at least two dimensions: epistemic evaluation and mitigation strategies. By examining social biases through this dual lens of unity and diversity, we can more effectively identify when and how to intervene on problematic biases. Ultimately, this approach provides a nuanced understanding of the nature of social bias, offering practical guidance for addressing existing biases and proactively managing emerging biases in both human and artificial minds. (shrink)

We argue that the later Wittgenstein’s philosophy of language and mathematics, substantially focused on rule-following, is relevant to understand and improve on the Artificial Intelligence (AI) alignment problem: his discussions on the categories that influence alignment between humans can inform about the categories that should be controlled to improve on the alignment problem when creating large data sets to be used by supervised and unsupervised learning algorithms, as well as when introducing hard coded guardrails for AI models. We cast these (...) considerations in a model of human–human and human–machine alignment and sketch basic alignment strategies based on these categories and further reflections on rule-following like the notion of meaning as use. To sustain the validity of these considerations, we also show that successful techniques employed by AI safety researchers to better align new AI systems with our human goals are congruent with the stipulations that we derive from the later Wittgenstein’s philosophy. However, their application may benefit from the added specificities and stipulations of our framework: it extends on the current efforts and provides further, specific AI alignment techniques. Thus, we argue that the categories of the model and the core alignment strategies presented in this work can inform further AI alignment techniques. (shrink)

How feasible are conceptual engineering projects of social concepts that aim for the engineered concept to be widely adopted in ordinary everyday life? Predominant frameworks on the psychology of concepts that shape work on stereotyping, bias, and machine learning have grim implications for the prospects of conceptual engineers: conceptual engineering efforts are ineffective in promoting certain social-conceptual changes. Specifically, since conceptual components that give rise to problematic social stereotypes are sensitive to statistical structures of the environment, purely conceptual change won’t (...) be possible without corresponding world change. This tradition, however, tends to ignore that concepts don’t only encode statistical, but also causal information. Paying attention to this feature of concepts, I argue, shows that conceptual engineering is not only possible. There is an imperative to conceptually-engineer. (shrink)

This paper begins with the thought that there is something out of place about offloading inquiry into one's own mind to AI. The paper's primary goal is to articulate the unease felt when considering cases of doing so. It draws a parallel between the use of algorithms in the criminal law: in both cases one feels entitled to be treated as an exception to a verdict made on the basis of a certain kind of evidence. Then it identifies an account (...) of first‐person authority that can make good on this: agentialism. Thus, the paper constitutes an argument in favor of an agentialist treatment of self‐knowledge and first‐person authority. (shrink)

This chapter is a survey of the ethics of belief. It begins with the debate as it first emerges in the foundational dispute between W. K. Clifford and William James. Then it surveys how the disagreements between Clifford and James have shaped the work of contemporary theorists, touching on topics such as pragmatism, whether we should believe against the evidence, pragmatic and moral encroachment, doxastic partiality, and doxastic wronging.

The ongoing explosion of interest in artificial intelligence is fueled in part by recently developed techniques in machine learning. Those techniques allow automated systems to process huge amounts of data, utilizing mathematical methods that depart from traditional statistical approaches, and resulting in impressive advancements in our ability to make predictions and uncover correlations across a host of interesting domains. But as is now widely discussed, the way that those systems arrive at their outputs is often opaque, even to the experts (...) who design and deploy them. Is it morally problematic to make use of opaque automated methods when making high-stakes decisions, like whether to issue a loan to an applicant, or whether to approve a parole request? Many scholars answer in the affirmative. However, there is no widely accepted explanation for why transparent systems are morally preferable to opaque systems. We argue that the use of automated decision-making systems sometimes violates duties of consideration that are owed by decision-makers to decision-subjects, duties that are both epistemic and practical in character. Violations of that kind generate a weighty consideration against the use of opaque decision systems. In the course of defending our approach, we show that it is able to address three major challenges sometimes leveled against attempts to defend the moral import of transparency in automated decision-making. (shrink)

The possibilities of unconscious perception and unconscious bias prompt parallel debates about unconscious mental content. This chapter argues that claims within these debates alleging the existence of unconscious content are made fraught by ambiguity and confusion with respect to the two central concepts they involve: consciousness and content. Borrowing conceptual resources from the debate about unconscious perception, the chapter distills the two conceptual puzzles concerning each of these notions and establishes philosophical strategies for their resolution. It then argues that empirical (...) evidence for unconscious bias falls victim to these same puzzles, but that progress can be made by adopting similar philosophical strategies. Throughout, the chapter highlights paths forward in both debates, illustrates how they serve as fruitful domains in which to study the relationship between philosophy and empirical science, and uses their combined study to further understanding of a general theory of unconscious content. (shrink)

In domains as disparate as playing Go and predicting the structure of proteins, artificial intelligence (AI) technologies have begun to perform at levels beyond which any humans can achieve. Does this fact represent something lamentable? Does superhuman AI performance somehow undermine the value of human achievements in these areas? Go grandmaster Lee Sedol suggested as much when he announced his retirement from professional Go, blaming the advances of Go-playing programs like AlphaGo for sapping his will to play the game at (...) a high level. In this paper, I attempt to make sense of Sedol’s lament. I consider a number of ways that the existence of superhuman-performing AI technologies could undermine the value of human achievements. I argue there is very little in the nature of the technology itself that warrants such despair. (Compare: does the existence of a fighter jet undermine the value of being the fastest human sprinter?) But I also argue there are several more localized domains where these technologies threaten to displace human beings from being able to achieve valuable things at all. This is a particular worry for those in unequal societies, I argue, given the difficulty of many achievements and the corresponding amount of resources needed to achieve great things. (shrink)

I argue that, given the way that AI models work and the way that ordinary human rationality works, it is very likely that people are anthropomorphizing AI, with potentially serious consequences. There are good reasons to doubt that LLMs have anything like human understanding, and even if they have representations or meaningful contents in some sense, these are unlikely to correspond to our ordinary understanding of natural language. However, there are natural, and in some ways quite rational, pressures to anthropomorphize (...) or personify LLMs and other AI systems in biased ways. This includes not only the classical or obvious ways of personifying AI — taking them to be sentient or have consciousness and understanding — but also taking AI to simulate or function like understanding, track the meaning of our language and our reasoning or logic, or have a model of the world our language is about. These more subtle or unobvious forms of anthropomorphism or personification can contribute to automation bias, AI Hype, and all their attendant risks, and encourage stronger anthropomorphisms as multi-modal AI assistants are developed further. (shrink)

In response to widespread use of automated decision-making technology, some have considered a right to explanation. In this paper I draw on insights from philosophical work on explanation to present a series of challenges to this idea, showing that the normative motivations for access to such explanations ask for something difficult, if not impossible, to extract from automated systems. I consider an alternative, outcomes-focused approach to the normative evaluation of automated decision-making, and recommend it as a way to pursue the (...) goods originally associated with explainability. (shrink)

Morality bears on what we should forget. Some aspects of our identity are meant to be forgotten and there is a distinctive harm that accompanies the permanence of some content about us, content that prompts a duty to forget. To make the case that forgetting is an integral part of our moral duties to others, the paper proceeds as follows. In §1, I make the case that forgetting is morally evaluable and I survey three kinds of forgetting: no-trace forgetting, archival (...) forgetting, and siloing. In §2, I turn to how we practice these forms of forgetting in our everyday lives and the goods these practices facilitate by drawing on examples ranging from the expunging of juvenile arrest records to the right to privacy. In §3, I turn to how my account can help us both recognize and address a heretofore neglected source of harm caused by technology and big data. In §4, I end by addressing the concern that we lack control over forgetting and thus can't be required to forget. I argue this challenge can be answered, but there’s a harder challenge that can’t. Forgetting is under threat. To address this challenge and preserve forgetting, we must change the world. (shrink)

Prediction-based decisions, which are often made by utilizing the tools of machine learning, influence nearly all facets of modern life. Ethical concerns about this widespread practice have given rise to the field of fair machine learning and a number of fairness measures, mathematically precise definitions of fairness that purport to determine whether a given prediction-based decision system is fair. Following Reuben Binns (2017), we take ‘fairness’ in this context to be a placeholder for a variety of normative egalitarian considerations. We (...) explore a few fairness measures to suss out their egalitarian roots and evaluate them, both as formalizations of egalitarian ideas and as assertions of what fairness demands of predictive systems. We pay special attention to a recent and popular fairness measure, counterfactual fairness, which holds that a prediction about an individual is fair if it is the same in the actual world and any counterfactual world where the individual belongs to a different demographic group (cf. Kusner et al. (2018)). (shrink)

There is a current debate about if, and in what sense, machine learning systems used in the medical context need to be explainable. Those arguing in favor contend these systems require post hoc explanations for each individual decision to increase trust and ensure accurate diagnoses. Those arguing against suggest the high accuracy and reliability of the systems is sufficient for providing epistemic justified beliefs without the need for explaining each individual decision. But, as we show, both solutions have limitations—and it (...) is unclear either address the epistemic worries of the medical professionals using these systems. We argue these systems do require an explanation, but an institutional explanation. These types of explanations provide the reasons why the medical professional should rely on the system in practice—that is, they focus on trying to address the epistemic concerns of those using the system in specific contexts and specific occasions. But ensuring that these institutional explanations are fit for purpose means ensuring the institutions designing and deploying these systems are transparent about the assumptions baked into the system. This requires coordination with experts and end-users concerning how it will function in the field, the metrics used to evaluate its accuracy, and the procedures for auditing the system to prevent biases and failures from going unaddressed. We contend this broader explanation is necessary for either post hoc explanations or accuracy scores to be epistemically meaningful to the medical professional, making it possible for them to rely on these systems as effective and useful tools in their practices. (shrink)

Some artificial intelligence systems can display algorithmic bias, i.e. they may produce outputs that unfairly discriminate against people based on their social identity. Much research on this topic focuses on algorithmic bias that disadvantages people based on their gender or racial identity. The related ethical problems are significant and well known. Algorithmic bias against other aspects of people’s social identity, for instance, their political orientation, remains largely unexplored. This paper argues that algorithmic bias against people’s political orientation can arise in (...) some of the same ways in which algorithmic gender and racial biases emerge. However, it differs importantly from them because there are strong social norms against gender and racial biases. This does not hold to the same extent for political biases. Political biases can thus more powerfully influence people, which increases the chances that these biases become embedded in algorithms and makes algorithmic political biases harder to detect and eradicate than gender and racial biases even though they all can produce similar harm. Since some algorithms can now also easily identify people’s political orientations against their will, these problems are exacerbated. Algorithmic political bias thus raises substantial and distinctive risks that the AI community should be aware of and examine. (shrink)

Firms use algorithms to make important business decisions. To date, the algorithmic accountability literature has elided a fundamentally empirical question important to business ethics and management: Under what circumstances, if any, are algorithmic decision-making systems considered legitimate? The present study begins to answer this question. Using factorial vignette survey methodology, we explore the impact of decision importance, governance, outcomes, and data inputs on perceptions of the legitimacy of algorithmic decisions made by firms. We find that many of the procedural governance (...) mechanisms in practice today, such as notices and impact statements, do not lead to algorithmic decisions being perceived as more legitimate in general, and, consistent with legitimacy theory, that algorithmic decisions with good outcomes are perceived as more legitimate than bad outcomes. Yet, robust governance, such as offering an appeal process, can create a legitimacy dividend for decisions with bad outcomes. However, when arbitrary or morally dubious factors are used to make decisions, most legitimacy dividends are erased. In other words, companies cannot overcome the legitimacy penalty of using arbitrary or morally dubious factors, such as race or the day of the week, with a good outcome or an appeal process for individuals. These findings add new perspectives to both the literature on legitimacy and policy discussions on algorithmic decision-making in firms. (shrink)

Law-enforcement agencies are increasingly able to leverage crime statistics to make risk predictions for particular individuals, employing a form of inference that some condemn as violating the right to be “treated as an individual.” I suggest that the right encodes agents’ entitlement to a fair distribution of the burdens and benefits of the rule of law. Rather than precluding statistical prediction, it requires that citizens be able to anticipate which variables will be used as predictors and act intentionally to avoid (...) them. Furthermore, it condemns reliance on various indexes of distributive injustice, or unchosen properties, as evidence of law-breaking. (shrink)

Data‐driven algorithms are widely used to make or assist decisions in sensitive domains, including healthcare, social services, education, hiring, and criminal justice. In various cases, such algorithms have preserved or even exacerbated biases against vulnerable communities, sparking a vibrant field of research focused on so‐called algorithmic biases. This research includes work on identification, diagnosis, and response to biases in algorithm‐based decision‐making. This paper aims to facilitate the application of philosophical analysis to these contested issues by providing an overview of three (...) key topics: What is algorithmic bias? Why and how can it occur? What can and should be done about it? Throughout, we highlight connections—both actual and potential—with philosophical ideas and concerns. (shrink)

One of the main lines of research in algorithmic fairness involves individual fairness (IF) methods. Individual fairness is motivated by an intuitive principle, similar treatment, which requires that similar individuals be treated similarly. IF offers a precise account of this principle using distance metrics to evaluate the similarity of individuals. Proponents of individual fairness have argued that it gives the correct definition of algorithmic fairness, and that it should therefore be preferred to other methods for determining fairness. I argue that (...) individual fairness cannot serve as a definition of fairness. Moreover, IF methods should not be given priority over other fairness methods, nor used in isolation from them. To support these conclusions, I describe four in-principle problems for individual fairness as a definition and as a method for ensuring fairness: (1) counterexamples show that similar treatment (and therefore IF) are insufficient to guarantee fairness; (2) IF methods for learning similarity metrics are at risk of encoding human implicit bias; (3) IF requires prior moral judgments, limiting its usefulness as a guide for fairness and undermining its claim to define fairness; and (4) the incommensurability of relevant moral values makes similarity metrics impossible for many tasks. In light of these limitations, I suggest that individual fairness cannot be a definition of fairness, and instead should be seen as one tool among several for ameliorating algorithmic bias. (shrink)

Predictive algorithms are playing an increasingly prominent role in society, being used to predict recidivism, loan repayment, job performance, and so on. With this increasing influence has come an increasing concern with the ways in which they might be unfair or biased against individuals in virtue of their race, gender, or, more generally, their group membership. Many purported criteria of algorithmic fairness concern statistical relationships between the algorithm’s predictions and the actual outcomes, for instance requiring that the rate of false (...) positives be equal across the relevant groups. We might seek to ensure that algorithms satisfy all of these purported fairness criteria. But a series of impossibility results shows that this is impossible, unless base rates are equal across the relevant groups. What are we to make of these pessimistic results? I argue that none of the purported criteria, except for a calibration criterion, are necessary conditions for fairness, on the grounds that they can all be simultaneously violated by a manifestly fair and uniquely optimal predictive algorithm, even when base rates are equal. I conclude with some general reflections on algorithmic fairness. (shrink)

As inductive decision-making procedures, the inferences made by machine learning programs are subject to underdetermination by evidence and bear inductive risk. One strategy for overcoming these challenges is guided by a presumption in philosophy of science that inductive inferences can and should be value-free. Applied to machine learning programs, the strategy assumes that the influence of values is restricted to data and decision outcomes, thereby omitting internal value-laden design choice points. In this paper, I apply arguments from feminist philosophy of (...) science to machine learning programs to make the case that the resources required to respond to these inductive challenges render critical aspects of their design constitutively value-laden. I demonstrate these points specifically in the case of recidivism algorithms, arguing that contemporary debates concerning fairness in criminal justice risk-assessment programs are best understood as iterations of traditional arguments from inductive risk and demarcation, and thereby establish the value-laden nature of automated decision-making programs. Finally, in light of these points, I address opportunities for relocating the value-free ideal in machine learning and the limitations that accompany them. (shrink)

In analyzing oppressive systems like racism, social theorists have articulated accounts of the dynamic interaction and mutual dependence between psychological components, such as individuals’ patterns of thought and action, and social components, such as formal institutions and informal interactions. We argue for the further inclusion of physical components, such as material artifacts and spatial environments. Drawing on socially situated and ecologically embedded approaches in the cognitive sciences, we argue that physical components of racism are not only shaped by, but also (...) shape psychological and social components of racism. Indeed, while our initial focus is on racism and racist things, we contend that our framework is also applicable to other oppressive systems, including sexism, classism, and ableism. This is because racist things are part of a broader class of oppressive things, which are material artifacts and spatial environments that are in congruence with an oppressive system. (shrink)

The use of machine learning systems for decision-support in healthcare may exacerbate health inequalities. However, recent work suggests that algorithms trained on sufficiently diverse datasets could in principle combat health inequalities. One concern about these algorithms is that their performance for patients in traditionally disadvantaged groups exceeds their performance for patients in traditionally advantaged groups. This renders the algorithmic decisions unfair relative to the standard fairness metrics in machine learning. In this paper, we defend the permissible use of affirmative algorithms; (...) that is, algorithms trained on diverse datasets that perform better for traditionally disadvantaged groups. Whilst such algorithmic decisions may be unfair, the fairness of algorithmic decisions is not the appropriate locus of moral evaluation. What matters is the fairness of final decisions, such as diagnoses, resulting from collaboration between clinicians and algorithms. We argue that affirmative algorithms can permissibly be deployed provided the resultant final decisions are fair. (shrink)

In this paper we make the case for the emergence of novel kind of bias with the use of algorithmic decision-making systems. We argue that the distinctive generative process of feature creation, characteristic of machine learning (ML), contorts feature parameters in ways that can lead to emerging feature spaces that encode novel algorithmic bias involving already marginalized groups. We term this bias _assembled bias._ Moreover, assembled biases are distinct from the much-discussed algorithmic bias, both in source (training data versus feature (...) creation) and in content (mimics of extant societal bias versus reconfigured categories). As such, this problem is distinct from issues arising from bias-encoding training feature sets or proxy features. Assembled bias is not epistemically transparent in source or content. Hence, when these ML models are used as a basis for decision-making in social contexts, algorithmic fairness concerns are compounded. (shrink)

In response to widespread use of automated decision-making technology, some have considered a right to explanation. In this article, I draw on insights from philosophical work on explanation to present a series of challenges to this idea, showing that the normative motivations for access to such explanations ask for something difficult, if not impossible, to extract from automated systems. I consider an alternative, outcomes-focused approach to the normative evaluation of automated decision making and recommend it as a way to pursue (...) the goods originally associated with explainability. (shrink)

Algorithm-assisted decision procedures—including some of the most high-profile ones, such as COMPAS—have been described as unfair because they compound injustice. The complaint is that in such procedures a decision disadvantaging members of a certain group is based on information reflecting the fact that the members of the group have already been unjustly disadvantaged. I assess this reasoning. First, I distinguish the anti-compounding duty from a related but distinct duty—the proportionality duty—from which at least some of the intuitive appeal of the (...) former illegitimately derives. Second, I distinguish between different versions of the anti-compounding duty, arguing that, on some versions, uses of algorithm-assisted decision procedures rarely clash with the anti-compounding duty. Third, drawing on examples of algorithm-assisted decision procedures, I present three objections to the idea that there is a reason not to compound injustice. The most important of these is that one can compound injustice in a non-disrespectful way, and that the wrongfulness of non-disrespectfully compounding injustice is fully explained by the proportionality duty. (shrink)

One of David Estlund’s key claims in Utopophobia is that theories of justice should not bend to human motivational limitations. Yet he does not extend this view to our cognitive limitations. This creates a dilemma. Theories of justice may ignore cognitive as well as motivational limitations—but this makes them so unrealistic as to be unrecognizable as theories of justice. Theories may bend to both cognitive and motivational limitations—but Estlund wants to reject this view. The other alternative is to find some (...) non-ad hoc way to distinguish cognitive from motivational limitations. I argue that this strategy will not work. Just as a person’s cognitive limitations may block her motives no matter how much she perseveres, so too motivational limitations may be genuine inabilities. Even ideal theories of justice must bend to even ordinary motivational limitations when they truly cause us to be unable to comply with requirements. (shrink)

Algorithm Decision Making (ADM) systems designed to augment or automate human decision-making have the potential to produce better decisions while also freeing up human time and attention for other pursuits. For this potential to be realised, however, algorithmic decisions must be sufficiently aligned with human goals and interests. We take a Principal-Agent (P-A) approach to the questions of ADM alignment and trust. In a broad sense, ADM is beneficial if and only if human principals can trust algorithmic agents to act (...) faithfully on their behalf. This mirrors the challenge of facilitating P-A relationships among humans, but the peculiar nature of human-machine interaction also raises unique issues. The problem of asymmetric information is omnipresent but takes a different form in the context of ADM. Although the decision-making machinery of an algorithmic agent can in principle be laid bare for all to see, the sheer complexity of ADM systems based on deep learning models prevents straightforward monitoring. We draw on literature from economics and political science to argue that the problem of trust in ADM systems should be addressed at the level of institutions. Although the dyadic relationship between human principals and algorithmic agents is our ultimate concern, cooperation at this level must rest against an institutional environment which allows humans to effectively evaluate and choose among algorithmic alternatives. (shrink)

A significant and important ethical tension in resource allocation and public health ethics is between utility and equity. We explore this tension between utility and equity in the context of health AI through an examination of a diagnostic AI screening tool for diabetic retinopathy developed by a team of researchers at Duke-NUS in Singapore. While this tool was found to be effective, it was not equally effective across every ethnic group in Singapore, being less effective for the minority Malay population (...) than for the Chinese majority. We discuss the problematic normative nature of bias in health AI and explore the ways in which bias can interact with various forms of social inequalities. From there, we examine the specifics of the diabetic retinopathy case and weigh up specific trade-offs between utility and equity. Ultimately, we conclude that it is ethically permissible to prioritise utility over equity where certain criteria hold. Given that any medical AI is more likely than not to have lingering bias due to bias in the training data that may reflect other social inequalities, we argue that it is permissible to implement an AI tool with residual bias where: (1) its introduction reduces the influence of biases (even if overall inequality is worsened), and/or (2) where the utility gained is significant enough and shared across groups (even if unevenly). (shrink)