Machine Learning as Performative Materialist Practice: Thirteen Theses on the Epistemology, Methodology, and Politics of Applied ML

TL;DR

This paper critiques traditional epistemological assumptions in applied machine learning within institutional contexts, proposing a performative materialist framework that emphasizes models as intervention tools rather than truth representations.

Contribution

It introduces a unified performative materialist framework for understanding ML, challenging conventional views and offering practical guidelines rooted in this perspective.

Findings

Models are seen as intervention instruments, not truth representations.

Data products form complex adaptive systems that coevolve with targets.

Validity is measured by real-world effects, not formal model properties.

Abstract

Machine learning practice in institutional decision-support contexts -- government, public policy, public health, criminal justice, resource allocation -- rests on a set of largely unexamined epistemological commitments inherited from classical statistics and computer science: that models represent stable regularities, that validation can be context-free, that performance metrics are politically neutral, and that feature importance reveals system structure. This paper challenges these commitments through a unified framework of performative materialist ML, articulated as thirteen theses. Drawing on Pickering's cybernetic ontology, the performativity literature from economic sociology (Callon, MacKenzie), Simon's bounded rationality, the formalization of performative prediction (Perdomo et al., 2020), and fifteen years of applied ML experience in government and public policy, we argue that: (1) ML models are best understood not as truth-seeking representations but as temporally situated compressions that function as instruments of intervention; (2) the full data product is a complex adaptive system that coevolves with its target and navigates a multi-objective space no single algorithm can optimize; (3) validity is fundamentally performative, measured by effects in the world rather than formal properties of the model; (4) the choices embedded in objective functions, fairness criteria, and resource thresholds are political decisions belonging to stakeholders, not technicians. We show how these theses unify several practical prescriptions -- temporal cross-validation, precision and recall at k, pipeline-aware fairness auditing, satisficing over optimizing -- as consequences of a coherent materialist epistemology rather than isolated best practices