Designing AI Experiences: Boundary Representations, Collaborative Processes, and Data Tools

Abstract

Artificial Intelligence (AI) has transformed our everyday interactions with technology through automation, intelligence augmentation, and human-machine partnership. Nevertheless, we regularly encounter undesirable and often frustrating experiences due to AI. A fundamental challenge is that existing software practices for coordinating system and experience designs fall short when creating AI for diverse human needs, i.e., ``human-centered AI'' or HAI. ``AI-first'' development workflows allow engineers to first develop the AI components, and then user experience (UX) designers create end-user experiences around the AI's capabilities. Consequently, engineers encounter end-user blindness when making critical decisions about AI training data needs, implementation logic, behavior, and evaluation. In the conventional ``UX-first'' process, UX designers lack the needed technical understanding of AI capabilities (technological blindness) that limits their ability to shape system design from the ground up. Human-AI design guidelines have been offered to help but neither describe nor prescribe ways to bridge the gaps in needed expertise in creating HAI.

In this dissertation, I investigate collaboration approaches between designers and engineers to operationalize the vision for HAI as technology inspired by human intelligence that augments human abilities while addressing societal needs. In a series of studies combining technical HCI research with qualitative studies of AI production in practice, I contribute (1) an approach to software development that blurs rigid design-engineering boundaries, (2) a process model for co-designing AI experiences, and (3) new methods and tools to empower designers by making AI accessible to UX designers. Key findings from interviews with industry practitioners include the need for ``leaky'' abstractions shared between UX and AI designers. Because modular development and separation of concerns fail with HAI design, leaky abstractions afford collaboration across expertise boundaries and support human-centered design solutions through vertical prototyping and constant evaluation. Further, by observing how designers and engineers collaborate on HAI design in an in-lab study, I highlight the role of design `probes' with user data to establish common ground between AI system and UX design specifications, providing a critical tool for shaping HAI design. Finally, I offer two design methods and tool implementations --- Data-Assisted Affinity Diagramming and Model Informed Prototyping --- for incorporating end-user data into HAI design.

HAI is necessarily a multidisciplinary endeavor, and human data (in multiple forms) is the backbone of AI systems. My dissertation contributions inform how stakeholders with differing expertise can collaboratively design AI experiences by reducing friction across expertise boundaries and maintaining agency within team roles. The data-driven methods and tools I created provide direct support for software teams to tackle the novel challenges of designing with data. Finally, this dissertation offers guidance for imagining future design tools for human-centered systems that are accessible to diverse stakeholders.