December 2021
Acknowledgements
My dissertation is dedicated to my grandmother, Mrs. L. N. Papneja, and to my wife, Devshree.
My journey towards the dissertation began in 2003 during my undergraduate years, when I enrolled in a course offered by the AI Institute on Knowledge Based Systems. There, in the basement of the Boyd Graduate Studies Building, is where I built my first interactive agent using Exsys Corvid – a rule-based animal identification system with an inference engine that consisted of a knowledge base, logic blocks, and had an explanation facility that explained to the user how it had reached its conclusion. Having studied literature on information privacy during my doctoral studies, I came up with the idea of using explanations as a mechanism to reduce the user’s privacy concerns in interactions with conversational agents (chatbots). I identified self-disclosure as a dependent variable, and conducted a thorough literature review of self-disclosure to conversational agents. I synthesized my findings in the form of an emergent theoretical framework. Likewise, I researched literature on explanations, identified the constructs and came up with the moderated-mediation theoretical model, as well as the arguments for the hypotheses, including the arguments relating to perceived relevance and informational justice.
I came up with the idea of using two different information sensitivity contexts, the two scenarios, the dozen or so questions for each scenario, and the two types of explanations for each question. I designed and conducted the explanations reliability test, reworded the explanations, and retested them. I found and adapted each measure, designed the online experiment, and created the pre- and post-experiment surveys using Qualtrics. I learned how to use IBM Watson Assistant, and built the six different chatbots needed for the six different experimental conditions. I filed for, and obtained (after a revision), the IRB approval.
In the middle of conducting my pilot study, I came across the two references – MacCallum et al. (1999) and Wolf et al. (2013) – that talked about sample size requirements in studies involving latent factors. Given that my empirical study had 16 latent factors, each with at least three measurement items, I realized that a sample size of 60 would not be enough. I increased the sample size to 150 and let the pilot study run for a few more days on the Prolific platform. After analyzing and writing up the results of the pilot in November 2021, I was told that I had executed on what I had promised as part of my proposal defense, and could use my pilot results as the results of my dissertation and be done. Instead, I chose to add a couple of questions to the sensitive scenario (and correspondingly to the non-sensitive scenario) in order to make the difference in perceived sensitivity between the sensitive and non-sensitive scenarios more stark. I then ran the main study on Prolific with a proper sample size of 650 (588 after exclusions).
My dissertation and PhD are the result of several such self-realizations and self-efforts, and I planned, executed, and wrote every word of the ~33,000-word document as best as I could within imposed constraints. Any methodological rigor in my empirical study came only from reading other studies, and from my time as an Industrial Engineering major at Georgia Tech, a time I really missed throughout my PhD.
I am thankful to the Terry College of Business and the University of Georgia. I am thankful to my family and friends, and to former professors and classmates from the University of Notre Dame and at Georgia Tech for their constant encouragement and guidance. I am thankful to my wife, her parents, and her sisters for helping keep my morale up through difficult times. I am thankful to my former managers and co-workers at Hewitt Associates and Goldman Sachs for helping me realize the broader perspective, and to see the forest through the trees in difficult times.