Reimagining Participation in Machine Learning Research and Development

Abstract

Products and services enabled by machine learning (ML) models are ubiquitous in our everyday lives. ML-enabled systems provide us directions while we walk or drive, recommendations when we watch, listen or shop, information when we want to remember or wonder, and increasingly, content when we want to create. The pervasive influence of these systems has prompted calls to involve end users and stakeholders in ML research and development. While participatory approaches have featured in computing research for decades, participatory ML has recently emerged as a distinct field that aims to involve end users and stakeholders in developing systems that learn from data. However, concerns about "participation-washing'' have accompanied the turn towards participation -- the practice of superficially involving stakeholders to create an illusion of meaningful input while withholding substantive power over processes and outcomes (Sloane 2022, p.1). Despite these concerns, systematic research on how participation functions across scales of ML research and development and how technical and institutional structures shape participatory processes remains limited.

This thesis examines participation in ML research and development through four interconnected studies, progressing from broad historical analysis to contemporary case studies. The first study, a systematic review and thematic analysis of participatory computing research, identifies how researchers historically navigated power dynamics with participants. Based on interviews, the second study investigates the strategies that participation brokers use and the challenges they confront when translating participant knowledge into inputs for ML development. The final two studies critically analyse participation in large language model (LLM) deployments: one demonstrates how feedback interfaces constrain user input through a survey and affordances analysis, while the other examines ChatGPT's societal impacts through three disruptive events, exposing the limitations of individualised participation for addressing collective concerns.

These studies demonstrate a fundamental tension: while participatory projects promise to integrate complementary forms of knowledge -- researchers' technical expertise and participants' contextual knowledge -- scaling ML-enabled systems transforms participation from relational to procedural. Academic and corporate incentives systematically privilege technical expertise, creating power asymmetries between researchers and participants that intensify as participation becomes tightly formatted through technical interfaces. Such formatting primarily serves organisational data collection interests and prevents collective action, thereby reducing participation to what Kelty (2020, p. 9) calls the "lowest common denominator of contributory autonomy.''

Based on these insights, I argue that meaningful participation in ML research and development requires structural rather than technical interventions. Simply expanding participation through existing channels risks entrenching current power asymmetries. Instead, I propose a dual approach: First, reforming research environments for community-specific projects through new funding models, evaluation metrics, and publication formats that value relational work and contextual knowledge equally with technical contributions, while evolving broker roles from facilitators to advocates who champion community interests. Second, for large-scale systems, moving beyond consensus-seeking towards productive contestation, where collectives forming around shared concerns about LLMs can challenge not only model outputs but also the politics of development and contexts of deployment. As ML-enabled systems increasingly become fixtures of our digital lives, this thesis offers insights and pathways for users, stakeholders and collectives to meaningfully shape their research and development.