The literature on algorithmic fairness has examined exogenous sources of biases such as shortcomings in the data and structural injustices in society. It has also examined internal sources of bias as evidenced by a number of impossibility theorems showing that no algorithm can concurrently satisfy multiple criteria of fairness. This paper contributes to the literature stemming from the impossibility theorems by examining how informational richness affects the accuracy and fairness of predictive algorithms. With the aid of a computer simulation, we (...) show that informational richness is the engine that drives improvements in the performance of a predictive algorithm, in terms of both accuracy and fairness. The centrality of informational richness suggests that classification parity, a popular criterion of algorithmic fairness, should be given relatively little weight. But we caution that the centrality of informational richness should be taken with a grain of salt in light of practical limitations, in particular, the so-called bias-variance trade off. (shrink)

Journal of Political Philosophy, EarlyView.

Philosophy and Phenomenological Research, Volume 104, Issue 3, Page 728-746, May 2022.

Philosophical Issues, Volume 31, Issue 1, Page 355-373, October 2021.

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)

A popular informal argument suggests that statistics about the preponderance of criminal involvement among particular demographic groups partially justify others in making defensive mistakes against members of the group. One could worry that evidence-relative accounts of moral rights vindicate this argument. After constructing the strongest form of this objection, I offer several replies: most demographic statistics face an unmet challenge from reference class problems, even those that meet it fail to ground non-negligible conditional probabilities, even if they did, they introduce (...) new costs likely to cancel out any justificatory contribution of the statistic, but even if they didn’t, demographic facts are the wrong sort to make a moral difference to agents’ negative rights. I conclude that the popular argument should be rejected, and evidence-relative theories do not have the worrisome implication. (shrink)

Many oppose the use of profile evidence against defendants at trial, even when the statistical correlations are reliable and the jury is free from prejudice. The literature has struggled to justify this opposition. We argue that admitting profile evidence is objectionable because it violates what we call “equal protection”—that is, a right of innocent defendants not to be exposed to higher ex ante risks of mistaken conviction compared to other innocent defendants facing similar charges. We also show why admitting other (...) forms of evidence, such as eyewitness, trace, and motive evidence, does not violate equal protection. (shrink)

A challenge we face in a world that has been shaped by, and continues to be shaped by, racist attitudes and institutions is that the evidence is often stacked in favor of racist beliefs. As a result, we may find ourselves facing the following conflict: what if the evidence we have supports something we morally shouldn’t believe? For example, it is morally wrong to assume, solely on the basis of someone’s skin color, that they’re a staff member. But, what if (...) you’re in a context where, because of historical patterns of discrimination, someone’s skin color is a very good indicator that they’re a staff member? When this sort of normative conflict looms, a conflict between moral considerations on the one hand and what you epistemically ought to believe given the evidence on the other, what should we do? It might be unfair to assume that they’re a staff member, but to ignore the evidence would mean risking inaccurate beliefs. Some, notably Tamar Gendler (2011), have suggested that we simply face a tragic irresolvable dilemma. In this chapter, I consider how these cases of conflict arise and I canvass the viability of suggested resolutions of the conflict. In the end, I argue that there’s actually no conflict here. Moral considerations can change how we epistemically should respond to the evidence. (shrink)

If someone says, “You’ve stereotyped me,” we hear the statement as an accusation. One way to interpret the accusation is as follows: you haven’t seen or treated me as an individual. In this essay, I interpret and evaluate a theory of wrongful stereotyping inspired by this thought, which I call the failure-to-individualize theory of wrongful stereotyping. According to this theory, stereotyping is wrong if and only if it involves failing to treat persons as individuals. I argue that the theory—however one (...) interprets it—is inadequate. Either the theory will not reliably identify all cases of wrongful stereotyping or it will fail to adequately explain why they are wrong. I conclude that it does not follow that we must entirely jettison the objection that stereotyping fails to treat persons as individuals. What follows is only that the objection must play a more circumscribed role in a theory of when and why stereotyping is wrong. (shrink)

This paper develops a precise understanding of the thesis of moral encroachment, which states that the epistemic status of an opinion can depend on its moral features. In addition, I raise objections to existing accounts of moral encroachment. For instance, many accounts fail to give sufficient attention to moral encroachment on credences. Also, many accounts focus on moral features that fail to support standard analogies between pragmatic and moral encroachment. Throughout the paper, I discuss racial profiling as a case study, (...) arguing that moral encroachment can help us identify one respect in which racial profiling is epistemically problematic. (shrink)

According to the doxastic wrongs thesis, holding certain beliefs about others can be morally wrongful. Beliefs which take the form of stereotypes based on race and gender and which turn out to be false and are negatively valenced are prime candidates for the charge of doxastic wronging: it is no coincidence that most of the cases discussed in the literature involve false beliefs. My aim in this paper is to show that the thesis of doxastic wrongs does not turn on (...) the truth-value or valence of beliefs. (shrink)

Using pure statistical evidence about a group to judge a particular member of that group is often found objectionable. One natural explanation of why this is objectionable appeals to the moral notion of respecting individuality: to properly respect individuality, we need individualized evidence, not pure statistical evidence. However, this explanation has been criticized on the ground that there is no fundamental difference between the so-called “individualized evidence” and “pure statistical evidence”. This paper defends the respecting-individuality explanation by developing an account (...) of what it means to respect individuality. It combines an idealistic account of respecting individuality and a prioritization account of respecting individuality, and offers a principled way to distinguish individualized evidence from non-individualized evidence. (shrink)

Why do organizations conduct job interviews? The traditional view of interviewing holds that interviews are conducted, despite their steep costs, to predict a candidate’s future performance and fit. This view faces a twofold threat: the behavioral and algorithmic threats. Specifically, an overwhelming body of behavioral research suggests that we are bad at predicting performance and fit; furthermore, algorithms are already better than us at making these predictions in various domains. If the traditional view captures the whole story, then interviews seem (...) to be a costly, archaic human resources procedure sustained by managerial overconfidence. However, building on T. M. Scanlon’s work, we offer the value of choice theory of interviewing and argue that interviews can be vindicated once we recognize that they generate commonly overlooked kinds of noninstrumental value. On our view, interviews should thus not be entirely replaced by algorithms, however sophisticated algorithms ultimately become at predicting performance and fit. (shrink)

As machine learning informs increasingly consequential decisions, different metrics have been proposed for measuring algorithmic bias or unfairness. Two popular “fairness measures” are calibration and equality of false positive rate. Each measure seems intuitively important, but notably, it is usually impossible to satisfy both measures. For this reason, a large literature in machine learning speaks of a “fairness tradeoff” between these two measures. This framing assumes that both measures are, in fact, capturing something important. To date, philosophers have seldom examined (...) this crucial assumption, and examined to what extent each measure actually tracks a normatively important property. This makes this inevitable statistical conflict – between calibration and false positive rate equality – an important topic for ethics. In this paper, I give an ethical framework for thinking about these measures and argue that, contrary to initial appearances, false positive rate equality is in fact morally irrelevant and does not measure fairness. (shrink)

I discuss the epistemic status of algorithmic predictions in the legal realm. My main claim is that algorithmic predictions do not give us knowledge, not even probabilistic knowledge. The situation, however, is relevantly different from the one in which we find ourselves at the time of assessing statistical evidence in general, and it is rather related to the fact that algorithmic fairness in legal contexts is essentially undetermined. In the light of this, we have to settle for justified beliefs and (...) justified credences. I end by drawing some morals for the Knowledge Norm of action. (shrink)

Recommendation systems (RSs) predict what the user likes and recommend it to them. While at the onset of RSs, the latter was designed to maximize the recommendation accuracy (i.e., accuracy was their only goal), nowadays many RSs models include diversity in recommendations (which thus is a further goal of RSs). In the computer science community, the introduction of diversity in RSs is justified mainly through economic reasons: diversity increases user satisfaction and, in niche markets, profits.I contend that, first, the economic (...) justification of diversity in RSs risks reducing it to an empirical matter of preference; second, diversity is ethically relevant as it supports two autonomy rights of the user: the right to an open present and the right to be treated as an individual. So far, diversity in RSs has been morally defended only in the case of RSs of news and scholarly content: diversity is held to have a depolarizing effect in a democratic society and the scientific community and make the users more autonomous in their news choices. I provide a justification of diversity in RSs that embraces all kinds of RSs (i.e., a holistic moral defense) and is based on a normative principle founded on the agency of the user, which I call the right to be an exception to predictions. Such a right holds that the proper treatment of a RS user qua agent forbids providing them with recommendations based only on their past or similar users’ choices. (shrink)

In this paper we argue that transparency of machine learning algorithms, just as explanation, can be defined at different levels of abstraction. We criticize recent attempts to identify the explanation of black box algorithms with making their decisions (post-hoc) interpretable, focusing our discussion on counterfactual explanations. These approaches to explanation simplify the real nature of the black boxes and risk misleading the public about the normative features of a model. We propose a new form of algorithmic transparency, that consists in (...) explaining algorithms as an intentional product, that serves a particular goal, or multiple goals (Daniel Dennet’s design stance) in a given domain of applicability, and that provides a measure of the extent to which such a goal is achieved, and evidence about the way that measure has been reached. We call such idea of algorithmic transparency “design publicity.” We argue that design publicity can be more easily linked with the justification of the use and of the design of the algorithm, and of each individual decision following from it. In comparison to post-hoc explanations of individual algorithmic decisions, design publicity meets a different demand (the demand for impersonal justification) of the explainee. Finally, we argue that when models that pursue justifiable goals (which may include fairness as avoidance of bias towards specific groups) to a justifiable degree are used consistently, the resulting decisions are all justified even if some of them are (unavoidably) based on incorrect predictions. For this argument, we rely on John Rawls’s idea of procedural justice applied to algorithms conceived as institutions. (shrink)

This study concerns the sociotechnical bases of human autonomy. Drawing on recent literature on AI ethics, philosophical literature on dimensions of autonomy, and on independent philosophical scrutiny, we first propose a multi-dimensional model of human autonomy and then discuss how AI systems can support or hinder human autonomy. What emerges is a philosophically motivated picture of autonomy and of the normative requirements personal autonomy poses in the context of algorithmic systems. Ranging from consent to data collection and processing, to computational (...) tasks and interface design, to institutional and societal considerations, various aspects related to sociotechnical systems must be accounted for in order to get the full picture of potential effects of AI systems on human autonomy. It is clear how human agents can, for example, via coercion or manipulation, hinder each other’s autonomy, or how they can respect each other’s autonomy. AI systems can promote or hinder human autonomy, but can they literally respect or disrespect a person’s autonomy? We argue for a philosophical view according to which AI systems—while not moral agents or bearers of duties, and unable to literally respect or disrespect—are governed by so-called “ought-to-be norms.” This explains the normativity at stake with AI systems. The responsible people (designers, users, etc.) have duties and ought-to-do norms, which correspond to these ought-to-be norms. (shrink)

Arguments on legal evidence rely on generalizations, that link a certain circumstance to a certain hypothesis and warrants the claim that the circumstance makes the hypothesis more probable. Some generalizations are acceptable and others are unacceptable. A generalization can be unacceptable on at least four different grounds. A false generalization is unacceptable because membership in the reference class does not increase the probability of the hypothesis. A non-robust generalization is unacceptable because it uses a reference class that is too heterogeneous. (...) A biastriggering generalization is unacceptable because decision makers are inclined to overestimate the evidentiary value of membership in the reference class. A discriminating generalization is unacceptable because it puts members in the reference class in an unfair disadvantage. Research funded by the Swedish Research Council. (shrink)