Is Machine Learning Unsafe and Irresponsible in Social Sciences? Paradoxes and Reconsidering from Recidivism Prediction Tasks - read online

Aug 4, 2026

The paper examines contested questions behind using computational machine learning methods for high-stakes social science predictions, focusing on recidivism prediction tasks. It challenges prevalent criticisms by identifying the underlying paradoxes in predictive effectiveness, generalization, and the operational trust such systems require. It reviews evidence on limited accuracy gains and persistent challenges to interpretability, then argues for a renewed paradigm integrating computational tools with conventional social science approaches.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages20

Reading Time50 min

AccessFree to download

Introduction

Machine learning in social science and high-stakes settings

Roadblocks from opacity and trust concerns

Critiques and paradoxes in recidivism prediction

Algorithmic accuracy limits

Interpretability and explainability trade-offs

Reconsidering the computational paradigm

Integrating computational methods with social science approaches

Title: Is Machine Learning Unsafe and Irresponsible in Social Sciences? Paradoxes and Reconsidering from Recidivism Prediction Tasks Authors: Jianhong Liu*, Dianshi Li Affiliations: Faculty of Law, University of Macau; Macau, China *Corresponding author. Email: [jliu@um.edu.mo](jliu@um.edu.mo) Abstract: The paper addresses some fundamental and hotly debated issues for highstakes event predictions underpinning the computational approach to social sciences. We question several prevalent views against machine learning and outline a new paradigm that highlights the promises and promotes the infusion of computational methods and conventional social science approaches. Main Text: Recent advances and breakthroughs in artificial intelligence, such as machine learning (ML), particularly deep learning models, have shown their superior predictive performance and out-of-sample generalization capabilities (He et al., 2016; Vaswani et al., 2017) . This progression has triggered social scientists to ponder the potential contributions of these technologies in their respective fields (Dwyer et al., 2018; Sun et al., 2020) . Over the past decade, an array of applications has emerged, certain applications have blossomed into burgeoning disciplines, including computational psychology and computational education (Kučak et al., 2018; Rothenberg et al., 2023) However, the deployment of these technologies has hit roadblocks in domains such as healthcare and criminal justice, where the opacity of machine learning models has raised concerns due to the high-risk nature of these domains themselves (Kirchner et al., 2016; Neri et al., 2020) . Specifically, the public’s trust to predictions in these high-stakes arenas—those with profound implications for individuals ’ or collectives’trajectories—are exceptionally susceptible to the ramifications of erroneous decision-making and obscured decision processes (Rudin et al., 2022; van Dijck, 2022) , as erroneous judgments in these contexts might result in patients receiving inappropriate medical interventions or the inadvertent release of high-risk criminals, leading to potentially grave consequences (Garrett & Rudin, 2022; Rudin, 2019) . Despite considerable interdisciplinary efforts to integrate machine learning technologies into social science (Corbett-Davies et al., 2023; Simmler et al., 2022) , recent scholarship has cast doubt on the very foundations of computational approach to social science. Researchers hailing from distinct academic traditions have grappled with finding a harmonious balance between data-driven methodologies and theorydriven frameworks (Angelino et al., 2017; Dwyer et al., 2018; Rothenberg et al., 2023; Sun et al., 2020) . The focus of this multidisciplinary discourse has been multi-faceted, encompassing the quest for algorithmic accuracy (Ozkan et al., 2020; Singh & Mohapatra, 2021) , the imperative for interpretability in algorithmic decision-making (Rudin et al., 2022; Tolan et al., 2019) , and the ethical dimensions related to procedural justice (Kaissis et al., 2020; Neri et al., 2020) . Nevertheless, emerging literature has not only questioned the efficacy of these crossdisciplinary endeavors but also challenged the very underpinnings of the whole computational approaches to social science disciplines (Dressel & Farid, 2018; Kirchner et al., 2016; Ozkan et al., 2020; Tolan et al., 2019; Wexler, 2017) . The views against machine learning approaches are primarily on two dimensions. Firstly, the question of algorithmic accuracy remains a significant hurdle. A thorough systematic review reveals that most studies report only modest performance metrics. The Area Under the Receiver Operating Characteristic (AUC-ROC) scores typically hover around 0.741 (Travainiet al., 2022). Dressel & Farid (2018), using merely two features, achieved an accuracy of 68%, equivalent to the performance of an untrained human (Dressel & Farid, 2018) . Similarly, Etzler et al. (2023)

FAQs

Why does the paper focus on recidivism prediction tasks in social sciences?

Recidivism prediction represents a high-stakes setting where prediction errors and hidden decision processes can have serious consequences for individuals and communities, making the trust and risks central issues.

What are the two main dimensions of concern raised against machine learning approaches?

The paper highlights (1) hurdles in algorithmic accuracy and (2) difficulties in achieving acceptable interpretability or explainability for algorithmic decision-making.

How does the paper propose to move beyond prevailing skepticism?

It adopts a stance different from much of the existing literature by revisiting the foundations and outlining a new paradigm that emphasizes both the promises of computational methods and the infusion of conventional social science approaches.

Aug 4, 2026

Upload by Olivia Brown