The Proxy Problem: Fairness and Artificial Intelligence

2021

Automated decision-making systems implemented in public life are typically standardized. One algorithmic decision-making system can replace thousands of human deciders. Each of the humans so replaced had her own decision-making criteria: some good, some bad, and some arbitrary. Is such arbitrariness of moral concern? We argue that an isolated arbitrary decision need not morally wrong the individual whom it misclassifies. However, if the same algorithms are applied across a public sphere, such as hiring or lending, a person could be excluded from a large number of opportunities. This harm persists even when the automated decision-making systems are "fair" on standard metrics of fairness. We argue that such arbitrariness at scale is morally problematic and propose technically informed solutions that can lessen the impact of algorithms at scale and so mitigate or avoid the moral harms we identify.

Business & Information Systems Engineering, 2023

Amicus Curiae, 2019

This article discusses conceptions of fairness in algorithmic decision-making, within the context of the UK’s legal system. Using practical operational examples of algorithmic tools, itargues that such practices involve inherent technical trade-offs over multiple, competing notions of fairness, which are further exacerbated by policy choices made by those public authorities who use them. This raises major concerns regarding the ability of such choices to affect legal issues in decision-making, and transform legal protections, without adequate legal oversight, or a clear legal framework. This is not to say that the law does not have the capacity to regulate and ensure fairness, but that a more expansive idea of its function is required.

2023

How to protect people from algorithmic harms? A promising solution, although in its infancy, is algorithmic impact assessment (AIA). AIAs are iterative processes used to investigate the possible short and long terms societal impacts of AI systems before their use, but with ongoing monitoring and periodic revisiting even after their implementation. When conducted in a participatory and transparent fashion, they could create bridges across the legal, social and computer science domains, promoting the accountability of the entity performing them as well as public scrutiny. They could enable to re-attach the societal and regulatory context to the mathematical definition of fairness, thus expanding the formalistic approach thereto. Whilst the regulatory framework in the European Union currently lacks the obligation to perform such AIA, some other provisions are expected to play a role in AI development, leading the way towards more widespread adoption of AIA. These include the Data Protection Impact Assessment (DPIA) under the General Data Protection Regulation (GDPR), the risk assessment process under the Digital Services Act (DSA) and the Conformity Assessment (CA) foreseen under the AI Regulation proposal. In this paper, after briefly introducing the plurality of definitions of fairness in the legal, social and computer science domains, and explaining to which extent the current and upcoming legal framework mandates the adoption of fairness metrics, we will illustrate how AIA could create bridges between all these disciplines, allowing us to build fairer AI solutions. We will then recognise the role of DPIA, DSA risk assessment and CA by discussing the contributions they can offer towards AIA but also identify the aspects lacking therein. We will then identify how these assessment provisions could aid the overall technical discussion of introducing and assessing fairness in AI-based models and processes.

AI and Ethics, 2021

There is growing concern that decision-making informed by machine learning (ML) algorithms may unfairly discriminate based on personal demographic attributes, such as race and gender. Scholars have responded by introducing numerous mathematical definitions of fairness to test the algorithm, many of which are in conflict with one another. However, these reductionist representations of fairness often bear little resemblance to real-life fairness considerations, which in practice are highly contextual. Moreover, fairness metrics tend to be implemented within narrow and targeted fairness toolkits for algorithm assessments that are difficult to integrate into an algorithm’s broader ethical assessment. In this paper, we derive lessons from ethical philosophy and welfare economics as they relate to the contextual factors relevant for fairness. In particular we highlight the debate around the acceptability of particular inequalities and the inextricable links between fairness, welfare and a...

IEEE Technology and Society Magazine

 Operating at a large scale and impacting large groups of people, automated systems can make consequential and sometimes contestable decisions. Automated decisions can impact a range of phenomena, from credit scores to insurance payouts to health evaluations. These forms of automation can become problematic when they place certain groups or people at a systematic disadvantage. These are cases of discrimination-which is legally defined as the unfair or unequal treatment of an individual (or group) based on certain protected characteristics (also known as protected attributes) such as income, education, gender, or ethnicity. When the unfair treatment is caused by automated decisions, usually taken by intelligent agents or other AI-based systems, the topic of digital discrimination arises. Digital discrimination is prevalent in a diverse range of fields, such as in risk assessment systems for policing and credit scores [1], [2]. Digital discrimination is becoming a serious problem, as more and more decisions are delegated to systems increasingly based on artificial intelligence (AI) techniques such as machine learning. Although a significant amount of research has been undertaken from different disciplinary angles to understand this challenge-from computer science to law to sociology-none of these fields have been able to resolve the problem on their own terms. For instance, computational methods to verify and certify bias-free data

Res Publica, 2023

Minds and machines, 2024

Our goal in this paper is to establish a set of criteria for understanding the meaning and sources of attributing (un)fairness to AI algorithms. To do so, we first establish that (un)fairness, like other normative notions, can be understood in a proper primary sense and in secondary senses derived by analogy. We argue that AI algorithms cannot be said to be (un)fair in the proper sense due to a set of criteria related to normativity and agency. However, we demonstrate how and why AI algorithms can be qualified as (un)fair by analogy and explore the sources of this (un)fairness and the associated problems of responsibility assignment. We conclude that more user-driven AI approaches could alleviate some of these difficulties.

Proceedings of the Conference on Fairness, Accountability, and Transparency

A key goal of the fair-ML community is to develop machine-learning based systems that, once introduced into a social context, can achieve social and legal outcomes such as fairness, justice, and due process. Bedrock concepts in computer science-such as abstraction and modular design-are used to define notions of fairness and discrimination, to produce fairness-aware learning algorithms, and to intervene at different stages of a decision-making pipeline to produce "fair" outcomes. In this paper, however, we contend that these concepts render technical interventions ineffective, inaccurate, and sometimes dangerously misguided when they enter the societal context that surrounds decision-making systems. We outline this mismatch with five "traps" that fair-ML work can fall into even as it attempts to be more context-aware in comparison to traditional data science. We draw on studies of sociotechnical systems in Science and Technology Studies to explain why such traps occur and how to avoid them. Finally, we suggest ways in which technical designers can mitigate the traps through a refocusing of design in terms of process rather than solutions, and by drawing abstraction boundaries to include social actors rather than purely technical ones. CCS CONCEPTS • Applied computing → Law, social and behavioral sciences; • Computing methodologies → Machine learning;

AI & Ethics, 2022

The use of predictive machine learning algorithms is increasingly common to guide or even take decisions in both public and private settings. Their use is touted by some as a potentially useful method to avoid discriminatory decisions since they are, allegedly, neutral, objective, and can be evaluated in ways no human decisions can. By (fully or partly) outsourcing a decision process to an algorithm, it should allow human organizations to clearly define the parameters of the decision and to, in principle, remove human biases. Yet, in practice, the use of algorithms can still be the source of wrongful discriminatory decisions based on at least three of their features: the data-mining process and the categorizations they rely on can reconduct human biases, their automaticity and predictive design can lead them to rely on wrongful generalizations, and their opaque nature is at odds with democratic requirements. We highlight that the two latter aspects of algorithms and their significance for discrimination are too often overlooked in contemporary literature. Though these problems are not all insurmountable, we argue that it is necessary to clearly define the conditions under which a machine learning decision tool can be used. We identify and propose three main guidelines to properly constrain the deployment of machine learning algorithms in society: algorithms should be vetted to ensure that they do not unduly affect historically marginalized groups; they should not systematically override or replace human decision-making processes; and the decision reached using an algorithm should always be explainable and justifiable.