To whom to explain and what? : Systematic literature review on empirical studies on Explainable Artificial Intelligence (XAI)
Year: 2021
Type: review
Study design: Systematic Review
Abstract: Expectations towards artificial intelligence (AI) have risen continuously because of machine learning models’ evolution. However, the models’ decisions are often not intuitively understandable. For this reason, the field of Explainable AI (XAI) has emerged, which tries to create different techniques to help users understand AI better. As AI’s use spreads more broadly in society, it becomes like a co-worker that people need to understand. For this reason, AI-human interaction in research is of broad and current interest. This thesis outlines the current empirical XAI research literature themes from the human-computer interaction (HCI) perspective. This study's method is an explorative, systematic literature review carried out following the PRISMA (Preferred Research Items for Systematic Reviews) method. In total, 29 articles that concluded an empirical study into XAI from the HCI perspective were included in the review. The material was collected based on database searches and snowball sampling. The articles were analyzed based on their descriptive statistics, stakeholder groups, research questions, and theoretical approaches. This study aims to determine what factors made users consider XAI transparent, explainable, or trustworthy and to whom the XAI research was intended. Based on the analysis, three stakeholder groups to whom the current XAI literature was aimed for emerged: end-users, domain experts, and developers. This study’s findings show that domain experts’ needs towards XAI vary greatly between domains, whereas developers need better tools to create XAI systems. The end-users, on their part, considered case-based explanations unfair and wanted to have explanations that “speak their language”. Also, the results indicate that the effect of current XAI solutions on users’ trust towards AI systems is relatively small or even non-existing. The studies’ direct theoretical contributions and the number of theoretical lenses used were both found out to be relatively low. This thesis’s most immense contribution is to provide a synthesis of the extant empirical XAI literature from the HCI perspective, which previous studies have rarely brought together. Continuing this thesis, researchers can further investigate research avenues such as explanation quality methodologies, algorithm auditing methods, users’ mental models, and prior conceptions about AI.
Source: UTUPub (University of Turku)
Author: Miika Tiainen
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W3165689925