FAIRPLAI : A Human-in-the-Loop Approach to Fair and Private Machine Learning

David Sanchez 1 ∗ , Holly Lopez 2 ∗ , Michelle Buraczyk 3 ∗ , Anantaa Kotal 1 · 2025

Read FAIRPLAI : A Human-in-the-Loop Approach to Fair and Private Machine Learning by David Sanchez 1 ∗ , Holly Lopez 2 ∗ , Michelle Buraczyk 3 ∗ , Anantaa Kotal 1 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-08702.txt.

Introduction

David Sanchez Jr. and three co-authors argue that fair and private machine learning cannot be fully automated, because fairness is a human judgment about context. Privacy tools can quietly worsen group disparities, and fairness fixes often need the very sensitive data that privacy rules lock away.

Their framework, FAIRPLAI, puts people inside the loop in three ways. It draws privacy-fairness frontiers so trade-offs among accuracy, privacy, and group outcomes stay visible. It lets stakeholders pick fairness criteria and operating points interactively. And it adds a differentially private auditing loop where reviewers inspect explanations and edge cases without exposing individual records. On benchmark datasets the approach keeps strong privacy while cutting fairness gaps versus automated baselines.

Catalog record

CREATORDavid Sanchez 1 ∗ , Holly Lopez 2 ∗ , Michelle Buraczyk 3 ∗ , Anantaa Kotal 1
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE46 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-08702

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

FAIRPLAI : A Human-in-the-Loop Approach to Fair and Private Machine Learning David Sanchez 1 ∗ , Holly Lopez 2 ∗ , Michelle Buraczyk 3 ∗ , Anantaa Kotal 1 Abstract As machine learning systems move from theory to practice, they are increasingly tasked with decisions that affect healthcare access, financial opportunities, hiring, and public services. In these contexts, accuracy is only one piece of the puzzle—models must also be fair to different groups, protect individual privacy, and remain accountable to stakeholders. Achieving all three is difficult: differential privacy can unintentionally worsen disparities, fairness interventions often rely on sensitive data that privacy restricts, and automated pipelines ignore that fairness is ultimately a human and contextual judgment . We introduce FAIRPLAI : (Fair and Private Learning with Active Human Influence) , a practical framework that integrates human oversight into the design and deployment of machine learning systems. FAIRPLAI works in three ways: (1) it constructs privacy–fairness frontiers that make trade-offs between accuracy, privacy guarantees, and group outcomes transparent; (2) it enables interactive stakeholder input, allowing decision-makers to select fairness criteria and operating points that reflect their domain needs; and (3) it embeds a differentially private auditing loop, giving humans the ability to review explanations and edge cases without compromising individual data security. Applied to benchmark datasets, FAIRPLAI consistently preserves strong privacy protections while reducing fairness disparities relative to automated baselines. More importantly, it provides a straightforward, interpretable process for practitioners to manage competing demands of accuracy, privacy, and fairness in socially impactful applications. By embedding human judgment where it matters most, FAIRPLAI offers a pathway to machine learning systems that are effective, responsible, and trustworthy in practice. Github: https://github.com/Li1Davey/Fairplai * * footnotetext: These authors contributed equally to this work I Introduction Machine learning has moved from theoretical development to widespread practical deployment in domains directly impacting people’s lives. Algorithms now assist in determining eligibility for loans, allocating healthcare resources, screening job candidates, and guiding public safety policies. In high-stakes applications, accuracy—while essential—is no longer sufficient. Increasingly, there is recognition that machine learning systems must also be fair , ensuring equitable treatment across groups; private , protecting sensitive individual data; and accountable , enabling trust and oversight. Yet growing evidence indicates that today’s systems often fall short of these standards. Recent research has demonstrated that machine learning systems can reproduce and even amplify existing social inequities when deployed in practice. For example, a large-scale evaluation of a widely used healthcare risk prediction system revealed systematic underestimation of the health needs of Black patients compared to white patients with equivalent medical conditions [ 38 ] . In employment settings, automated hiring platforms have been found to reflect gendered patterns in candidate selection. At the same time, facial recognition systems have shown disproportionately high error rates for women and individuals with darker skin tones [ 11 ] . These findings underscore how seemingly neutral models can generate disparate outcomes across demographic groups. At the same time, mounting evidence highlights the vulnerability of machine learning models to privacy attacks. Studies have shown that adversaries can determine whether an individual’s record was part of a model’s training set through membership inference attacks , or reconstruct sensitive personal attributes through model inversion . Such vulnerabilities expose individuals to risks that extend beyond technical failures, undermining the system’s reliability and the legitimacy of its deployment in sensitive domains. These challenges crystallize into what we term the fairness–privacy–accuracy paradox . Differential privacy provides strong protection against data leakage by injecting carefully calibrated randomness during training. However, this same noise can disproportionately harm underrepresented groups with weak signals, thereby increasing disparities. Conversely, fairness-enhancing methods often rely on access to sensitive demographic attributes such as race, gender, or age—information that privacy-preserving frameworks seek to limit or conceal. Meanwhile, optimizing purely for predictive accuracy without consideration of fairness or privacy can exacerbate both inequities and exposure risks. Striving for one goal often undermines another, and balancing all three has proven deeply challenging. Much of the existing research on fairness and privacy attempts to navigate this paradox through algorithmic innovation—developing new objective functions, constraints, or optimization procedures. While these methods offer valuable theoretical insights, they frequently share a key limitation: they treat fairness and privacy as abstract optimization targets . In practice, however, fairness is not a single universal definition, but a collection of normative judgments that vary across contexts and stakeholders. Similarly, privacy is not solely a mathematical guarantee, but also a social expectation shaped by human values and institutional trust. Treating these concerns as purely technical quantities risks missing the fact that decisions about what is “fair enough” or “private enough” are fundamentally human judgments. This observation motivates our central research question : How can machine learning systems balance fairness, privacy, and accuracy in ways that reflect human notions of morality, security, and accountability, rather than abstract optimization alone? To address this question, we propose FAIRPLAI ( Fair and Private Learning with Active Human Influence ), a human-in-the-loop framework for fair and private machine learning. Instead of seeking a single automated solution to reconcile fairness and privacy, FAIRPLAI explicitly incorporates human oversight at key points in the machine learning lifecycle. It does so through three main components: 1. Privacy–fairness frontiers : FAIRPLAI generates interpretable visualizations that map the trade-offs between predictive accuracy, fairness disparities, and privacy budgets, enabling stakeholders to see how different design choices affect outcomes. 2. Interactive stakeholder selection : Rather than fixing one fairness definition in advance, FAIRPLAI allows domain experts, policymakers, or organizational leaders to select criteria (e.g., demographic parity, equalized odds, or counterfactual fairness) and operating points that best reflect their context. Through these mechanisms, FAIRPLAI reframes fairness and privacy not as constraints to be optimized away, but as values to be contextualized, negotiated, and overseen by humans . The goal is not a purely theoretical optimum, but a practical balance that aligns technical safeguards with how humans reason about fairness and security. The contributions of this paper are threefold: • We articulate the fairness–privacy–accuracy paradox, showing why existing automated methods struggle to reconcile these objectives in real-world settings. • We introduce FAIRPLAI , a human-in-the-loop framework that integrates differential privacy with fairness-aware modeling while making trade-offs explicit through privacy–fairness frontiers. • We empirically demonstrate that FAIRPLAI preserves rigorous privacy guarantees, reduces fairness disparities compared to automated baselines, and provides interpretable decision points for human oversight in deployment. By embedding human judgment directly into the learning process, FAIRPLAI advances machine learning systems that are not only accurate, but also fair, private, and accountable to the values of those they serve . II Problem Statement: The Fairness–Privacy–Accuracy Paradox Machine learning systems deployed in socially sensitive domains are judged not only by how well they predict, but also by how responsibly they handle data and how equitably they treat individuals and groups. This means that three objectives must be considered together: II-A Performance (Accuracy) Performance, often expressed as predictive accuracy, is the most established measure of machine learning. In supervised learning, models are trained on datasets consisting of input features and corresponding outcomes, and their quality is assessed by how well predictions align with the true outcomes. The simplest performance measure is accuracy, defined as the proportion of correct

Read the complete text