Construction of Model Behavior by Novices through Interaction with Machine Learning

Jury

• Genoveva Vargas-Solar – Researcher (HDR), Université Claude Bernard Lyon 1 – Reviewer & Examiner
• Simone Stumpf – Professor, University of Glasgow – Reviewer & Examiner
• Marianela Ciolfi Felice – Assistant Professor, KTH Royal Institute of Technology – Examiner
• Anastasia Bezerianos – Professor, Université Paris-Saclay – Examiner

Abstract

Machine learning (ML) systems increasingly mediate decisions across domains such as hiring, healthcare, criminal justice, and content moderation. While often framed as neutral or objective, these systems frequently reproduce the very inequalities they claim to address. Such harms are not isolated technical failures, but reflections of broader sociotechnical dynamics: models encode dominant norms shaped by the data, assumptions, and institutional practices that underlie their design. Addressing these dynamics requires moving beyond internal performance metrics and attending to how model behavior unfolds in context, including how models operate in real-world settings and how people interpret and respond to their outputs.

This thesis investigates how model behavior is not only examined but constructed through human interaction with ML models in real-world settings. It centers novice users who are typically excluded from formal development pipelines but directly impacted by model decisions, exploring how they interpret, evaluate, and potentially shape model behavior through interaction. Grounded in Human–Computer Interaction (HCI) and informed by Science and Technology Studies (STS) and AI ethics, this thesis adopts a human-centered lens to present two empirical studies that engage novice users in the development and post-deployment phases of the ML lifecycle. These studies examine how user interaction makes model behavior accessible for reflection, negotiation, critique, and, if possible, change.

The first study introduces Collaborative Interactive Machine Teaching (CIMT), where small groups of novice users train an image classifier using a web-based platform called TeachTOK. It examines how participants collaboratively plan teaching strategies, curate training data, and reflect on model behavior using real-time feedback. Findings show that collaboration was central to negotiating teaching objectives, such as representational diversity, and building a shared understanding of model behavior.

The second study investigates user-driven algorithm auditing as a post-deployment practice, in which non-experts evaluate a pretrained image captioning model for gender bias. Drawing on cognitive models of sensemaking, we designed an auditing interface with two additional tools: one that supports hypothesis testing through input manipulation, and another that enables grouping outputs to reveal patterns. A between-subjects online study examined how interface conditions shaped users’ sensemaking outcomes. Findings show that different tools influenced which harms users could detect, how they formed hypotheses, and how they developed confidence in their interpretations. 

Across both development and post-deployment phases, this thesis demonstrates how novices’ interaction with the ML model contributes to the construction of model behavior. In CIMT, participants shaped behavior through teaching and negotiation. In auditing, they surfaced harms and evaluated model behavior through iterative sensemaking. In both cases, users brought their own perspectives to bear on how models should behave, perspectives often excluded from conventional ML development and evaluation.

This work contributes empirical, technical, and methodological insights that have the potential to support participatory and reflexive approaches to machine learning within human-centered AI, and calls for future research grounded in sustained, real-world, and community-led practices.