The Immortal Science of ML - Machine Learning & the Theory-Free Ideal - Paper

Aug 5, 2026

Upload by Quinn

This paper argues that machine learning methods do not enable a truly “theory-free” form of inductive inference, contrary to claims of epistemic distinctness. The core issue is a misguided conception of scientific objectivity: the theory-free ideal, framed in analogy to the value-free ideal. Using two case studies, the paper shows how removing theory mediation cannot be sustained, and how the theory-free ideal undermines the epistemic standing of scientific work that employs ML.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages25

Reading Time1 h 3 min

AccessFree to download

Introduction

ML’s promise of theory-free science

What “theory-free” is taken to deny

Why theory-free induction fails

Induction, theory, and empirical support

Epistemic status and the theory-free ideal

The Immortal Science of ML: Machine Learning & the Theory-Free Ideal Mel Andrews July 31, 2025 Abstract This paper contends with the notion that the methods of machine learning (ML) are unique among the tools of science in enabling a form of theory-free inductive inference. I contest these assertions of epistemic distinctness, attributing the prevalence of these views to an untenable conception of scientific objectivity: what I term a theory-free ideal, in homage to its normative counterpart. ML, as a formal method of induction, must rely on conceptual or theoretical resources to get inference off the ground. By means of two case studies, I argue that the theory-free ideal has a deleterious effect on the epistemic standing of ML-involving science. 1 Introduction In the decade elapsed since the deep learning revolution, machine-learning (ML) techniques have established a firm foothold throughout the sciences. Successes of such methods have ranged from real-time particle-event sortation at the Large Hadron Collider (Duarte et al., 2018) to DeepMind’s Nobel prize-winning accomplishments with its AlphaFold 2.0 system for protein-structure prediction (Jumper et al., 2021) . According to certain spokespeople from science and engineering communities, however, these achievements are trivial in comparison to the potential such methods hold for scientific advancement. Various commentators have asserted with some frequency that ML will instigate profound—even “revolutionary”—changes to the nature of science and the knowledge it produces (Anderson, 2008; Boge, 2022; Hey et al., 2009; MayerSch¨onberger & Cukier, 2013; Society & Institute., 2019; Spinney, 2022; Sre´ckovi´cet al., 2022) . The people behind such claims see ML methods as holding the potential to retire or else displace the role of theorizing in science (Anderson, 2008; Mayer-Sch¨onberger & Cukier, 2013; Spinney, 2022; Sre´ckovi´c et al., 2022) . Desai et al. (2022) refer to this conception of an ML-enabled scientific paradigm as “the epistemically revolutionary new frontier raised by data science: the socalled ‘theory-free’ paradigm in scientific methodology”(p.469) . Some of these statements regarding the scientific usage of ML echo proclamations that were once made of domain-generic statistical analyses: Levins and Lewontin (1985) write of the motivation behind principle component analysis (PCA) and regression techniques that researcher’s “assumption is that they are approaching the data in a theory-free manner and that data will ‘speak to them’ through the correlation analysis” (Levins & Lewontin, 1985, 156) . We see this sentiment echoed today in the assertion that big data analytic tools promise to allow the raw data to “speak for themselves”(Anderson, 2008, 1) . This theory-freedom is intended as a negation of theory-mediation, theorydrivenness, theory-involvement, and theory-ladeness. It is also, we will see from an examination of the source literature, a denial that methods rest essentially on domain-knowledge or prior conceptualization of the target phenomena, or that they should be understood as representing features of target systems in any inference-licensing respect. “Theory” is hence to be understood in a broad and colloquial sense, as incorporating domain knowledge or conceptualization of target phenomena. Subscribers to the theory-free ideal seek to purge science of what they see as epistemically compromising arbitrariness and subjectivity. This subjective element is brought on board when human critical thinking or conceptualization of target phenomena play an essential role in shaping an empirical research program. Taken at face value, the claim that ML could enable a form of theory-free inductive inference—or even inductive inference that differs qualitatively in its degree of theoretical support—does not withstand scrutiny. That inductive inference rests on pregiven conceptual infrastructure is, I take it, an effective premise of all disc

FAQs

What does the paper mean by the “theory-free ideal” in ML-involving science?

It refers to a meta-narrative claiming that ML can retire or displace theorizing, analogous to a value-free ideal. The paper treats this as a problematic conception of scientific objectivity.

Why does the paper reject the idea of genuinely theory-free inductive inference?

It argues that inductive inference requires conceptual infrastructure and both theoretical and empirical support. The notion that induction shifts toward being less theoretically grounded is treated as incoherent.

What effect does the theory-free ideal have on the epistemic standing of science using ML?

Through two case studies, the paper contends that endorsing the theory-free ideal is detrimental, because it reifies potential misuse into an account of how ML methods work and how they should work.

Aug 5, 2026

Upload by Quinn