Fairness - from the ethical principle to the practice of Machine Learning development - as an ongoing agreement with stakeholders

Aug 5, 2026

Upload by Logic

This paper explains why bias in Machine Learning (ML) cannot be fully eliminated and develops an end-to-end approach to turn justice and fairness ethical principles into day-to-day ML development through an ongoing agreement with stakeholders. The proposed iterative pro-ethical process targets asymmetric power in fairness decisions, enabling ML teams to identify, mitigate, and continuously monitor bias across system-development steps. It also offers ways to communicate imperfect bias trade-offs to users.

LanguageEnglish

CategoryTechnology

FormatPDF

Pages17

Reading Time43 min

AccessFree to download

Introduction

Bias and discrimination in AI

From abstract ethics to operational practice

Technical bias mitigation and its limitations

Need for practical, stakeholder-oriented processes

Fairness: from the ethical principle to the practice of Machine Learning development as an ongoing agreement with stakeholders Georgina Curto Flavio Comim University of Notre Dame Universitat Ramon Llull, IQS School of Management Notre Dame, USA Barcelona, Spain [gcurtore@nd.edu](gcurtore@nd.edu) [flavio.comim@iqs.url.edu](flavio.comim@iqs.url.edu) 22 March 2023 Abstract This paper clarifies why bias cannot be completely mitigated in Machine Learning (ML) and proposes an end-to-end methodology to translate the ethical principle of justice and fairness into the practice of ML development as an ongoing agreement with stakeholders. The pro-ethical iterative process presented in the paper aims to challenge asymmetric power dynamics in the fairness decision making within ML design and support ML development teams to identify, mitigate and monitor bias at each step of ML systems development. The process also provides guidance on how to explain the always imperfect trade-offs in terms of bias to users. Keywords: Bias, Artificial Intelligence, Trustworthy AI, fairness, discrimination, pro-ethical design 1. Introduction Discrimination and bias in AI are still unresolved topics in the current information civilization (Zuboff 2019) . While AI has the potential to facilitate the achievement of all United Nations Sustainable Goals, it can also widen existing social gaps by reproducing and often aggravating societal bias (Vinuesa et al. 2020) . ML systems in particular have been found to often exacerbate representational and allocational harm to vulnerable salient groups (Suresh and Guttag 2021) . As a result, these groups not only receive demeaning treatment, but also less resources and opportunities. AI-amplified bias has been identified in critical services such as education, health and justice (Floridi 2020) . ML bias has other particularities that deserve special attention. Users are often not aware of it and developers cannot always explain it (Barocas & Selbst 2016) . In this context, governmental efforts to regulate AI have gained traction in the past few years (White House 2016; European Commission 2021) . In addition, there has been a proliferation of ethical guidelines (Algorithm Watch 2021) in what has been described as a “moral panic”(Ess 2020) . These have been found to converge on specific topics (Hagendorff 2020; Zeng et al. 2019) and have been summarized as 5 ethical principles: transparency, justice and fairness, non-maleficence, responsibility and privacy (Jobin 2019).The scope of this article focuses on the principle of justice and fairness and facilitates the principle of transparency by guiding which specific steps of the fairness decision making process should be openly disclosed. In the nascent field of AI ethics, these ethical principles have been qualified as appropriate but too abstract. AI development teams find them difficult to apply in practice. Existing AI ethics guidelines focus the effort on the “what” and fall short on clarifying the operationalization of AI ethics (Floridi 2019; Morley, Elhalal, et al. 2021; Morley, Kinsey, et al. 2021; Vakkuri et al. 2020; Vakkuri and Kemell 2019) . As a result, counterproductive practices such as ethics shopping, ethics bluewashing, ethics lobbying, ethics dumping or ethics shirking are prone to flourish (Floridi 2019) . An urgency has been identified to translate theoretical principles into practical inclusive processes (Harrison et al. 2020) . This article has four objectives and the first one is to answer the question: how can the principle of justice and fairness be applied into the practice of ML development? In parallel to the production of ethical guidelines, the principle of fairness has been tackled from the technical perspective by mitigating bias. A large body of work in recent years has been produced on the bias identification and debiasing of ML systems, especially on Natural Language Processing (NLP) (Bolukbasi et al. 2016; Calis

FAQs

Why can’t bias be completely mitigated in machine learning development?

Bias cannot be fully eliminated because algorithms and learning processes can reproduce or amplify bias, and social impact is hard to predict before deployment.

What does the paper propose to translate fairness ethics into ML practice?

An end-to-end, iterative pro-ethical methodology that treats fairness as an ongoing agreement with stakeholders and supports teams in managing fairness decisions throughout development.

How does the proposed process help teams handle bias across ML system development?

It guides teams to identify, mitigate, and monitor bias at each step of ML system development, including how to explain imperfect bias trade-offs to users.

Aug 5, 2026

Upload by Logic