VALUES IN MACHINE LEARNING: WHAT FOLLOWS FROM UNDERDETERMINATION? - Tom F. Sterkenburg - Paper

Aug 2, 2026

Upload by Oliver

Inductive underdetermination is often taken to imply that machine learning algorithms are necessarily value-laden. This paper makes the notion of “value-ladenness” for “machine learning algorithms” precise, then challenges a general inference from underdetermination to that conclusion. It argues that learning procedures need not inherit value judgments merely from their inductive character, and that establishing value-ladenness requires detailed engagement with specific algorithmic designs and their interaction of epistemic and non-epistemic factors.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages27

Reading Time1 h 8 min

Introduction

Bias and value-ladenness in inductive inference

Paper roadmap

The underdetermination argument

Johnson’s representative formulation

Canons of inductive inference and non-epistemic values

VALUES IN MACHINE LEARNING: WHAT FOLLOWS FROM UNDERDETERMINATION? TOM F. STERKENBURG Abstract. It has been argued that inductive underdetermination entails that machine learning algorithms must be value-laden. This paper offers a more precise account of what it would mean for a “machine learning algorithm” tobe “value-laden,” and, building on this, argues that a general argument from underdetermination does not warrant this conclusion. 1. Introduction Machine learning is in many ways biased. Much contemporary work in computer science and philosophy alike is devoted to charting the various types and entry points of algorithmic bias in machine learning pipelines (d’Alessandro et al. , 2017; Danks and London, 2017; Hellstr¨om et al. , 2020; Mehrabi et al. , 2021; Fazelpour and Danks, 2021) . Several authors (Karaca, 2021; Biddle, 2022 , 2023; Birhaneet al. , 2022; Nyrup, 2022; Sullivan, 2022 , 2023) have also made a connection to the philosophy of science literature on the role of non-epistemic value judgmentsin scientific inference (Douglas, 2016; Elliott and Steel, 2017; Elliott, 2022) . Some have adopted arguments from this literature to reason more fundamentally that machine learning algorithms must be value-laden (Dotan, 2021; Johnson, 2024) . Johnson (2024, p. 28), in particular, poses the question “whether it is really possible for [machine learning] algorithms to be value-free even in principle.” Setting aside the “[p]roblematic social patterns [...] necessarily encoded in the data on which algorithms operate,” and setting aside even the “all-too-human nature of the engineers themselves,” she asks “whether values are constitutive of the very operation of algorithmic decision-making, such that on no idealized conception could [machine learning algorithms] be value-free” (ibid.) . In addressing this question, Johnson adopts general arguments from the philosophy of science against the so-called value-free ideal. These arguments rely on the inductive nature of scientific inference, and the fundamental problem of the underdetermination of inductive conclusions by the available data; characteristics that are shared by machine learning algorithms. Johnson writes that “[t]hese arguments result in the view that both scientific and algorithmic decision procedures are deeply value-laden”(2024, p. 30) . Yet there is something unsatisfying about the lesson that machine learning algorithms must be the product of value-laden choices beginning to end. For one thing, it seems odd to be led to the conclusion that, simply in virtue of their being procedures for inductive learning, standard learning algorithms like stochastic gradient Date: December 20, 2024 . This is a preliminary version. I welcome feedback. 2 STERKENBURG descent or Bayesian updating must already be inherently value-laden. More generally, when one opens a textbook on machine learning, one finds various theoretical and methodological—apparently epistemic—motivations for this or that algorithm. What seems in order, in the spirit of work connecting the ethics and epistemology of artificial intelligence (Russo et al. , 2023; Grote, forthcoming), is a more careful picture of how both epistemic and non-epistemic factors come together in the design of machine learning algorithms. One step towards such a picture is to show why a general argument from underdetermination does not already settle the matter that learning algorithms must be value-laden. That is the work of this paper. To be clear, I do not seek to defend a claim that machine learning algorithms are not, in fact, value-laden. In the course of my analysis I indicate more specific paths for exposing value-ladenness, particularly there in algorithm design where epistemic considerations must meet practical demands. My point is that these paths require more detailed engagement with the actual learning algorithms, and as such require more work than a general argument from underdetermination. The plan

FAQs

What question does the paper address about underdetermination and machine learning?

It questions whether inductive underdetermination really entails that machine learning algorithms must be value-laden.

How does the author refine what “value-laden” means for machine learning algorithms?

The paper offers a more precise account of what it would mean for a “machine learning algorithm” to be “value-laden,” rather than relying on a general slogan.

What is the paper’s main conclusion about arguments from underdetermination?

A general argument from underdetermination does not, by itself, warrant the conclusion that learning algorithms are value-laden; the claim requires more detailed work about actual algorithms.

Aug 2, 2026

Upload by Oliver