User Agency in AI-infused Systems

Author:

Adenuga, Iyadunni

Graduate Program:

Informatics

Degree:

Doctor of Philosophy

Document Type:

Dissertation

Date of Defense:

May 28, 2024

Committee Members:

Dongwon Lee, Professor in Charge/Director of Graduate Studies
Jack Carroll, Major Field Member
Hadi Hosseini, Major Field Member
Christopher Dancy, Outside Unit & Field Member
Jonathan Dodge, Chair & Dissertation Advisor

Keywords:

Human Agency
Human-Centric AI-infused Systems
Scale Validation
Explainable AI
User-centered Design

Abstract:

AI-infused systems are ubiquitous in today's environments. These systems are usually automated and opaque, so they sometimes hamper people's innate and unique attribute of feeling agency (i.e., feeling in control). Enhancing human agency in any sphere leads to increased satisfaction and productivity, as well as better performance. These qualities are necessary in current sociotechnical systems if people are to easily adopt and appropriately trust these AI-infused systems. Prior work to make AI-infused systems less opaque, which are usually system-centric, focuses on providing explanations. More human-centric approaches are questioning the usefulness and actionability of these explanations. Our approach to this issue involves teasing out agency from its assumed co-occurrence with explanations. The main question we attempt to solve in this dissertation is how to design AI-infused systems to accommodate and enhance human agency needs. First, we perform a thoughtful examination of AI explanations and human agency and outline possible forms of relationship between the two phenomena. Using this knowledge, we designed and built an agentic text summarization system called Living Documents. Text summarization is an example of an AI-infused system with a one-track operation such that an input generates a static output. We included several interactions and studied their effects on people's feelings of control. This introduced an additional research question of how to accurately assess people's agency needs so that designers are able to meet them. We created a 4-part agency scale, which we validated in multiple studies. The 4-part agency scale includes: 1) Need for Agency in everyday situations- NfAE 2) Sense of Agency in everyday situations-SoAE 3) Need for Agency with related technology-NfAT and 4) Sense of Agency with this technology-SoAT. The first three parts (NfAE, SoAE, and NfAT) of this scale help to gather demographical information about the agency needs and expectations of the intended target users. The fourth part (SoAT) helps to assess how well the created system designs match these expectations. Furthermore, we were able to collate and present some design recommendations that would aid interaction designers in the development of agentic forms of AI-infused systems. This dissertation ends with concluding remarks that include a brief discussion of a preliminary study's results that highlight the need for agency for all of the different user groups in a typical sociotechnical environment like ridesharing, as well as a broad description of my future research agenda.