Graduate Thesis Or Dissertation


Assessing and Finding Faults in AI: Two Empirical Studies Public Deposited

Downloadable Content

Download PDF


Attribute NameValues
  • With the advent of Artificial Intelligence (AI) in every sphere of life in today's day and age, it has become increasingly important for non-AI experts to be able to comprehend the underlying logic of how AI systems work, assess them and find faults in these systems, particularly when they are used in high risk scenarios such as in military strategies and medical applications. Recent developments to address the need to open the black boxes of these AI-powered systems have led to the emergence of AI explanations. There now exist myriad successful explanation methods and tools that attempt to explore and explain how AI systems work. However, a key problem with such work is the lack of a process that users can follow to navigate AI systems along with their explanation. This problem becomes increasingly evident with non-AI experts, due to their lack of context and depth of knowledge of the subject. To address this challenging problem, my colleagues and I propose a new process called AAR/AI or After-Action Review for Artificial Intelligence that aims to bridge this gap between AI systems and non-AI experts. AAR/AI, inspired by the US Defense debriefing strategy called AAR, is a process for understanding, analyzing and navigating sequential decision making environments. This thesis details two human-subjects studies my colleagues and I conducted, one qualitatively and the other quantitatively, to evaluate the effectiveness of AAR/AI in assessing an AI system and in identifying and localizing faults in it. The studies recommend that not only does AAR/AI assist non-AI experts to effectively navigate an AI system and keep their thoughts organized and logical, it also helps them identify and localize faults in it. Participants that used AAR/AI to localize faults did so with far more precision and recall than those that did not. I believe that this is a crucial step towards building democratic and explainable AI systems, and making them accessible to a larger audience that is not familiar with them.
Resource Type
Date Issued
Degree Level
Degree Name
Degree Field
Degree Grantor
Commencement Year
Committee Member
Academic Affiliation
Rights Statement
Peer Reviewed



This work has no parents.

In Collection: