Shaking up the dogma - Solving trade-offs without (moral) values in machine learning - Draft

Aug 3, 2026

The paper examines how ethical and epistemological considerations intertwine across machine-learning practice, focusing on the accuracy–interpretability trade-off as a paradigmatic case. It surveys the common view that resolving conflicting desiderata requires non-epistemic values, drawing on arguments from inductive risk. In contrast, it argues that under sufficiently specified deployment purposes, the trade-off can be resolved using purely epistemic grounds. The work analyzes the structure of trade-offs and formalizes accuracy and interpretability, then derives resolution strategies based on selecting the appropriate epistemic frame.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages16

Reading Time40 min

AccessFree to download

Abstract

1. Introduction

Ethical and epistemological entanglement

Value-ladenness and division of labor

False dogmas and reconciled trade-offs

Accuracy versus interpretability as an epistemic issue

Shaking up the dogma: Solving trade-offs without (moral) values in machine learning *Draft* Thomas Grote; Cluster of Excellence: “Machine Learning – New Perspectives For Science”; University ofTübingen; [thomas.grote@uni-tuebingen.de](thomas.grote@uni-tuebingen.de). Oliver Buchholz; Chair of Bioethics; ETH Zürich; oliver.buchholz@hest.ethz.ch Abstract The field of machine learning intricately links ethical and epistemological considerations in many contexts which raises the question as to their precise relation. This paper tries to provide a partial answer by focusing on one particular context, namely, the trade-off between accuracy and interpretability, which can be considered a prime example for the entanglement of ethics and epistemology in machine learning. At its core, the trade-off states that any choice of a machine learning model needs to balance the conflicting desiderata of achieving accurate predictions and an interpretable functionality. On a widely shared view inspired by the argument from inductive risk, this balancing of conflicting desiderata can only be resolved by appeal to non-epistemic values. By contrast, we argue that, in certain settings, the accuracy-interpretability trade-off can be resolved on purely epistemic grounds. To that end, we closely analyze the general nature of trade-offs as well asthe notions of accuracy and interpretability. This allows us to derive strategies for resolving the accuracy-interpretability trade-off that center around choosing the right epistemic frame for a given machine learning application and, thus, do not require non-epistemic considerations. We conclude by sketching the implications of this result for the general relation of ethical and epistemological considerations in ML. Keywords: machine learning; trade-offs; accuracy; interpretability; values in science: 1. Introduction Ethical and epistemological issues are often enmeshed in machine learning. It is not uncommon for ethicists to wrestle with epistemic concepts like accuracy, reliability, opacity, or uncertainty alongside genuinely morally normative concepts like autonomy or fairness. Nevertheless, with few exceptions, the exact relationship between ethics and epistemology is rarely ever spelled out (Russo et al., 2024; Grote, 2024; Sterkenburg, 2024) . This comes at the expense of an unclear understanding concerning the scope and methodology of ethics in machine learning. This paper isan attempt to make progress in this regard. Specifically, we study the relationship between ethics and epistemology on the basis of trade-offs in machine learning. According to a widely held view, inspired by the literature on inductive and epistemic risk (see, e.g., Ward, 2021), machine learning models are value laden in that different trade-offs arise in the design and development process (Biddle, 2022; Nyrup, 2022; Johnson, 2023) . These trade-offs amount to a choice between two desiderata that cannot be mutually satisfied. It is furthermore assumed that many of these trade-offs are inescapable: they occur necessarily and can only be resolved by means of value judgements, which then reflect in the model. If we accept these assumptions, then this results in a division of labor between ethicists and epistemologists in the design and development of machine learning models. On the one hand, the task of theepistemologist is to install appropriate safeguards so that the machine learning model meets a particular epistemic desideratum. On the other, the ethicist’s task is to balance different value choices underlying certain trade-offs. Even still, it is becoming apparent that the picture is more complicated than just suggested. Many trade-offs turn out to be false dogmas upon closer examination. For example, Beigang (2023) has argued that a trade-off between two statistical fairness notions, ‘equalized odds’ and ‘predictive parity’, can be modified by way of causal inference techniques so that they are universally c

FAQs

What central trade-off does the paper focus on?

The paper focuses on the trade-off between accuracy and interpretability in machine learning, treating it as an example of how ethics and epistemology can become entangled.

What is the commonly shared view about resolving accuracy–interpretability conflicts?

A widely shared view holds that resolving such conflicts requires appeal to non-epistemic values, since conflicting desiderata cannot be jointly satisfied and can only be settled via value judgments.

How does the paper argue the accuracy–interpretability trade-off can be resolved?

It argues that in well-specified deployment settings, the trade-off can be resolved on purely epistemic grounds by choosing the right epistemic frame for the machine-learning application, avoiding non-epistemic considerations.

Aug 3, 2026

Upload by Theodore