The philosophical basis of algorithmic recourse*

admin管理员组

文章数量:1579932

The philosophical basis of algorithmic recourse*

Suresh Venkatasubramanian1

suresh@cs.utah.edu University of Utah

ABSTRACT

Philosophers have established that certain ethically important values are modally robust in the sense that they systematically deliver correlative benefits across a range of counterfactual scenarios. In this paper, we contend that recourse - the systematic process of reversing unfavorable decisions by algorithms and bureaucracies across a range of counterfactual scenarios - is such a modally robust good. In particular, we argue that two essential components of a good life - temporally extended agency and trust - are underwritten by recourse.

We critique existing approaches to the conceptualization, operationalization and implementation of recourse. Based on these criticisms, we suggest a revised approach to recourse and give examples of how it might be implemented - especially for those who are least well off1.

KEYWORDS

recourse, algorithmic decision making, precarity, robust goods

ACM Reference Format:

Suresh Venkatasubramanian and Mark Alfano. 2020. The philosophical basis of algorithmic recourse. In Conference on Fairness, Accountability, and Transparency (FAT* ’20), January 27-30, 2020, Barcelona, Spain. ACM, New York, NY, USA, 10 pages. https://doi/10.1145/3351095.3372876

• 1 INTRODUCTION

Human agents are distinctive among animals in the amount of longterm planning they engage in. We make plans that may come to fruition days, weeks, years, or even decades in the future. In some cases we even plan for events that will occur only after our own deaths. Such planning is remarkable not just for the amount of time involved but also for the level of recursive means-end reasoning involved. If your ultimate aim is, for instance, to vacation next year in Hawaii, you might go about it by saving money in order to be

‘This research was partially funded by an Australian Research Council Discovery Project (code: DP190101507) and by the National Science Foundation under grant IIS-1633724

^Both authors contributed equally to this research.

1We focus on the least well off because this is arguably the most defensible principle from an ethical and political point of view.[27]

Mark Alfano I

mark.alfano@gmail Delft University of Technology Macquarie University

able to purchase a ticket. And you might go about saving money by getting a second job. And you might go about getting a second job by receiving certification to work that job. And you might go about receiving certification by taking vocational training courses. In this scenario, you take training in order to receive certification in order to get a second job in order to save money in order to purchase airfare in order to go to Hawaii. Such plans are only likely to succeed in a sufficiently well-ordered system, in which the reasons things might go wrong are foreseeable and understandable, and in which errors can be identified and rectified. If you could not trust that the vocational training would be sufficient to get certified, it would not make sense to plan in this way. Likewise, if you could not trust that hyperinflation would not destroy your savings, it would not make sense to plan in this way. The kind of agency that we both expect to be able to exercise and in many cases actually do exercise presupposes that our society is organized in a sufficiently regular, understandable, and corrigible way, which makes it possible to trust that the elaborate, temporally-extended planning we engage in is likely to be successful.

Among the things people typically make long-term plans for are various essential primary goods, such as housing. Computer scientists interested in algorithmic fairness have tended to focus on the distribution of such goods. In this paper, we are also concerned with their nature. In particular, we are interested in the fallback mechanisms and dispositions that people may be able to take advantage of when they lack an important primary good. In recent years, social scientists have begun to study the growing instability surrounding access to various primary goods. Researchers sometimes speak of the problem of precarity[5, 11, 21], which broadly speaking can be characterized as a state of precarious existence (or precarious access to resources like employment, housing, health care and so on) in which small “shocks” can remove access to such critical resources. Someone who suffers a precarious existence lacks financial and social security, which impinges on their ability to engage in temporally extended agency. They may have housing and a steady job today, but if anything were to go wrong in their life (e.g., a chronic illness, an unexpected financial burden, a parking ticket), they would lose their housing or job. A recent study in the state of New Jersey found that loss of driving privileges due to license suspension (which is often used as a punishment for reasons unrelated to driver safety) led to severe collateral impacts: 42% of the people whose license had been suspended reported reported losing their jobs [6]. This is unsurprising in a state with inadequate public transport: if you can’t get to work, it’s hard to hold down a job. Furthermore, of those who reported losing their job, 45% reported being unable to find a new job, and 88% of those who did find new employment reported a decrease in income.

In a similar vein, a recent survey by bankrate found that only 41% of American adults would be able to cover an unexpected cost of $500 from their existing finances2. Likewise, the Federal Reserve Board [25] found that 46% of adults in the United States could not cover an emergency expense of $400 without having to sell something or borrow money. Someone who suffers from precarity in this way will find it hard to engage in temporally extended agency and to trust that their plans will come to fruition. They face a life of constant worry and stress, and such worry and stress can have knock-on effects that feed back into the precariousness of their existence. For example, stress and anxiety may lead someone to snap at their boss, which could get them fired. Elizabeth Anderson [2, pg. 63] estimates that approximately 80% of American workers - essentially, all those who are “neither securely self-employed nor upper-level managers” - are just “one arbitrary and oppressive managerial decision away” from being fired, demoted, or otherwise mistreated by the pervasive “authoritarian governance in our work and off-hours lives.”

• 1.1 The modally robust good of recourse

The examples described in the previous section and others like them suggest that people will often need some way to reverse unfavorable decisions that would otherwise impair their ability not only to accomplish one particular goal but also to accomplish all of the other goals that it is a means to. For example, someone who is counting on a loan in order to purchase a car in order to be able to drive to a well-paying job in order to take care of their family might be denied that loan. In such a case, the denial affects not just their immediate financial situation but their whole life plan. If someone cannot trust that they will have some way of overcoming challenges that thwart the crucial means to their long-term ends, they will have little reason to try to engage in the temporally-extended agency characteristic of mature adults.

We live in a world in which many decisions that significantly affect our ability to exercise temporally-extended agency are made by algorithms and bureaucracies3. These algorithms and bureaucracies establish a system of incentives and disincentives that apply to both the ends that people might pursue for their own sake and the means to those ends. If you want to enter a profession, you typically need to receive some sort of certification. If you want to make a large purchase, you may need to take out a loan (that you can pay back at a reasonable interest rate in a reasonable amount of time). If you want to travel internationally, you need to get a passport and potentially also a visa. Across a vast range of sectors, decisions that fundamentally affect people’s lives and their ability to engage in long-term planning are made by algorithms and bureaucracies. Sometimes, those decisions are unfavorable. When they are, the subject of the decision can only reasonably plan their subsequent course of action if they know what it would take to receive a more favorable decision. After all, a desired or hoped-for end can only become the target of a plan if the agent is able to select a means to that end. Moreover, this need to be able to plan applies not just to one-off cases, but generally over the course of one’s life. As such, someone can be positioned in such a way that they know or reasonably expect that, were things to go wrong, they would be able to set them right again. Such positioning refers not only to the way things currently are but also to how they might be across a range of counterfactual scenarios.

As such, we need some way to ensure that people both have some way of getting unfavorable decisions reversed and know, in general, that they will have a way of getting unfavorable decisions reversed. Let us define the enjoyment of recourse as being in such a position. Recourse systematically delivers the benefit of reversing harmful decisions by algorithms and bureaucracies across a range of counterfactual scenarios4. If someone enjoys recourse, then not only are they able to get a single decision reversed, but they also enjoy the power to reverse decisions across a range of counterfactual scenarios. As such, someone who enjoys recourse need not passively suffer the slings and arrows of outrageous fortune, but is positioned to take up arms against a sea of troubles. They do not suffer from what Condorcet [3, 11:161, 191] considered one of the most debilitating aspects of poverty: “the idea of being counted for nothing, of being delivered up, without defense, to all vexations and all outrages.”

It is illuminating in this context to refer to recent work by Philip Pettit [26], who has argued that a wide range of ethically important values are modally robust5. For a good to be modally robust in Pettit’s sense, it must systematically deliver some other benefit in a range of counterfactual scenarios. For instance, according to Pettit, people value the non-robust good of favor, and therefore also value the robust good of friendship, which delivers favor in a range of counterfactual scenarios. If someone is your friend, not only do they favor you now, but also they would be disposed to favor you in a range of nearby possible worlds. Friends are disposed to put one another back on course rather than simply abandoning each other when the going gets tough [1], and there are derogatory natural language expressions (e.g., ‘fair-weather friend’) for people whose favor cannot be counted on in a broad enough range of counterfactual scenarios. Beyond friendship, Pettit argues, people value a variety of other robust goods. The virtue of honesty is a robust good that delivers the non-robust benefit of truth-telling in a range of counterfactual scenarios. If someone is honest, you can trust them to tell you the truth when they have no incentive to lie, but also to tell you the truth were lying to be to their benefit. Likewise, the robust good of respect delivers the non-robust benefit of non-interference in a range of counterfactual scenarios.

According to Pettit, robust goods are valuable because they are “resilient enough to survive situational shifts” (p. 24), and thus deliver their correlative non-robust goods both “as things actually are” and “as they would be under certain variations” (p. 46). For this reason, when we are assured that someone embodies a robust good, we can live free from anxiety and fear that the correlative non-robust benefit (and everything that depends on it) will suddenly be snatched away without notice or warning. Robust goods thus systematically deliver, as a side-effect, peace of mind and warrant for trust.

While Pettit’s account focuses primarily on robust goods as they are embodied in individual humans, it is also possible for a social group or an institution to embody a robust good. For example, a fail-safe nuclear reactor is a complex socio-technical system in which multiple layers of safeguards are put in place. When such a reactor is working as designed, it delivers two non-robust benefits in a range of counterfactual scenarios: namely, electrical power and safety from radiation. If something were to go wrong - either mechanically or via human error - in a fail-safe reactor, multiple alerts and protective actions would be triggered that would (at least if it works as designed) set the reactor on a course towards equilibrium.

In addition, whereas Pettit’s account focuses only on robust goods, it is possible in similar fashion to define modally robust ills as ills that deliver non-robust harms in a range of counterfactual scenarios. For example, malevolence towards someone is a robust ill because it delivers harm in a range of counterfactual scenarios. If someone harbors malevolence towards you, then not only are they going to harm you in the actual world when it is easy for them, but also they will go out of their way to harm you in nearby possible worlds where they face obstacles to harming you. And, just as robust goods can be embodied by both individuals and institutions, so robust ills can be embodied by both individuals and institutions. For example, Kate Manne [23] argues that misogyny

本文标签： BASISPhilosophicalrecourseAlgorithmic