As artificial intelligence (AI) becomes increasingly integrated into products and services across industries, ensuring responsible development and deployment has become a critical challenge. Several tools, processes, and principles have been developed with the aim of supporting responsible AI (RAI) development in industry practice. However, research has shown a persistent gap between the aspirational goals of these RAI interventions and the practical realities faced by RAI practitioners in designing and building responsible AI systems.

To bridge this gap, in the first part of my thesis, I conducted a set of empirical studies with individual AI practitioners, cross-functional AI teams, and end users, to contribute a deeper understanding of RAI practices and challenges in industry settings. Through these investigations, I identified a set of design opportunities for RAI interventions that are actually useful and usable for practitioners on the ground. To synthesize key insights into the evolving landscape of industry RAI, I conducted a systematic literature review of empirical studies on industry RAI practices over the past decade, highlighting concrete takeaways for RAI researchers, practitioners, and policymakers.

Building on insights from these studies, in the second part of my thesis, I design, develop, and evaluate a set of tools and processes to better support RAI practices in real-world industry settings, particularly in the contexts of AI auditing, red-teaming, and impact assessment. For example, through an iterative co-design process, I developed a Societal Impact Assessment template to support industry AI and computing researchers in grappling with the potential negative societal impact of their work. I developed WeAudit, a generalizable workflow and a corresponding system that scaffolds end users in iteratively investigating, deliberating on, and reporting perceived harms in generative AI systems. Finally, I built PersonaTeaming Playground, a system that incorporates user personas into automated red-teaming algorithms to support current red-teaming practices and explore human–AI collaboration in the context of generative AI evaluation.

Weaving together field research with the design of novel HCI systems, this thesis advances actionable insights and practical solutions to support HCI and AI researchers, practitioners, and policymakers in developing safer and more responsible AI systems.

270 pages

Thesis Committee:
Kenneth Holstein (Co-Chair)
Motahhare Eslami (Co-Chair)
Jason I. Hong
Jenn Wortman Vaughan (Microsoft Research)
Solon Barocas (Microsoft Research)

Brad A. Myers, Head, Human-Computer Interaction Institute
Martial Hebert, Dean, School of Computer Science

Return to: SCS Technical Report Collection
School of Computer Science homepage

This page maintained by reports@cs.cmu.edu