1
Journey to XAI: An Evolution Perspective

Ruby Chanda1* and Sarika Sharma2

1Symbiosis Institute of Management Studies, Symbiosis International (Deemed) University, Maharashtra, Pune, India

2Symbiosis Institute of Computer Studies and Research, Symbiosis International (Deemed) University, Maharashtra, Pune, India

Abstract

In the last couple of years, multiple industries such as finance, healthcare, transport, entertainment, and education have radically changed with the exponential growth of artificial intelligence (AI). But on the other hand, concerns about bias, opacity, and unpredictability of AI systems have deepened as they are increasingly involved in important decision-making processes. Due to these intricacies, the concept of explainable artificial intelligence (XAI) has become a vital area of research and development for the assurance of AI systems that are accountable, understandable, and transparent. This chapter presents a comprehensive analysis of the origins and the maturity of XAI, including its reasons, aims, main milestones, proper methods, possible legal effects, potential obstacles, and coming trends. XAI begins with an overview of rule-based AI systems, which were dominant in the old times, and then proceeds to explain different types of AI, which have since replaced these outdated algorithms. The main nodes of the contemporary stage proved AI mechanisms to be unambiguous commands and algorithms following clear instructions and logic for decision-making; thus, this process turned out to be neither lengthy nor complex. On the other hand, the next-to-nothing capacity of the rule-based system was unsuccessful in addressing such intricate issues as the ones with unique demands and exceptional factors. The black-box AI, which is subsequently created on the basis of deep learning and machine learning technologies, forms the narrative's central theme. The advent of these systems made image recognition, natural language processing, and self-driving vehicles feasible in businesses, which was previously a far-fetched dream for several business owners. Yet, due to the fact that they are by their nature unclear and impenetrable, the issues of equal justice, correctness, and trustworthiness of those black-box AI systems are intensively being questioned. The necessity of XAI is recognized because of this issue, their target being to build XAI models for advancing the understanding of AI algorithms and gaining basic trust from people in the AI systems. Among the numerous XAI strategies built up after the dedication of resources include feature significance analysis, attention mechanisms, model-agnostic approaches, rule extraction, local interpretability tools, and counterfactual explanations. By using the case studies and empirical reviews, researchers have proved that such approaches bring useful outlines to the process of decision-making in AI. Besides that, the scope of XAI amalgamates with ethical and regulatory considerations as well. It should be noted that the General Data Protection Regulation of the European Union outlines the right to an account of automated decision-making process underlining the need for responsibility and transparency in AI presentations. The institute of electrical and electronics engineers (IEEE) and association for computing machinery (ACM) ethical standards made clear the fact that the responsible AI, where fairness, transparency, accountability, and privacy necessarily follow, needed to be brought into practice. Undoubtedly, we have come a long way, but there are some difficulties in the prospects of using effective XAI that still needs to be addressed. This obstacle still has many barriers to overcome, such as the danger of antagonism, the lack of scalability, the extremely high computational complexity, and AI comprehension by people. Nevertheless, the process to XAI goes in the path of interdisciplinary alliance, innovation, and development, by which one can realize hope. To begin with, the progress in XAI will be expected to shape the AI revolution in the future by bringing stakeholders an adequate degree of control over using the technologies and avoiding the worst outcomes. This chapter is founded on an explanative and integrated description of the development of XAI over the years, ranging from rule-based AI systems in earlier stages to the super-complex black-box models of today. We stress the exceptional goals reached in XAI research, outline ways of improving interpretability, and define the essential factors that push explainability movement. We will also address the gap points of the present XAI era and those expected for the future, and we will look at the ethical and legal concerns raised with XAI usage in the past. This chapter reveals transparency and accountability as one of AI systems essential, and discussing XAI implementation changes future of AI. With this specific goal, it develops by closely monitoring the progress of XAI.

Keywords: XAI, technology, AI, evolution, black box

1.1 Introduction

1.1.1 Definition of Explainable Artificial Intelligence

The term "explainable artificial intelligence" (XAI) describes the potential biases and expected outcomes of an artificial intelligence (AI) model. It helps define model accuracy, equity, openness, and AI-driven decision-making outcomes. When implementing AI models, an organization must be able to communicate AI to its stakeholders in order to gain their trust. Explainability in AI makes it easier for an organization to follow a responsible AI development approach.

The term "explainable artificial intelligence" describes a group of strategies and tactics intended to assist human users in comprehending and having faith in the results generated by machine learning (ML) algorithms. In contrast to ML's "black-box" theory, which holds that even the system's designers are unable to completely explain how the AI makes decisions, XAI seeks to make these processes understandable and transparent. XAI represents the implementation of the social right to explanation. The challenge for humans as AI develops is to comprehend and trace the algorithm's steps.

This leaves the entire mathematical process as an uninterpretable black box, as it is frequently called. These black-box models are built straight out of the data. Furthermore, neither the AI algorithm's inner workings nor how it arrived at a certain conclusion can be understood or explained by the engineers or data scientists who developed it.

In other words, XAI is a term that frequently overlaps with interpretable AI and explainable ML. It describes AI systems that allow humans to maintain intellectual control over them or the techniques used to get there [1]. The primary emphasis is typically on AI's thinking, which is made clearer and comprehensible in its judgments or forecasts [2, 3]. XAI counteracts the black-box tendency of ML, wherein even the AI's developers are unable to provide an explanation for a specific decision [4, 5].

1.1.2 Importance of Explainability in AI Systems

The ability of the human intellect to defend judgments to others is one of its most important aspects. People who keep their goals and thoughts to themselves are likely to be viewed as "strange fellows," so it is crucial in social circumstances as well. In the classroom, students want to comprehend the reasoning of their lecturers. Furthermore, establishing trust often requires that one explain their choices to others, such as when a doctor explains a therapeutic option to a patient.

XAI is a resurgent research trend as the need to support these rules and principles and the explainable decision-making system and research grows. Recent developments in AI and ML, their application to various fields, and lingering worries about the models' unethical usage, lack of transparency, and unwanted biases are the main causes of the revived interest in XAI research.

The Industrial Control System has several practical applications that significantly boost industrial production efficiency through automated machinery and manufacturing processes. Nevertheless, because of the model's and the decisions' lack of explainability and transparency, the usage of the black box is still not recommended in this situation. Wang et al. and Longo et al. [6, 7] state that XAI includes ML or AI systems/tools for deciphering the inner workings of black models (e.g., what the models have learned) and/or for elucidating specific predictions [8]. The level of transferable qualitative information about the relationship between model input and prediction-that is, the appropriate/selective causes of the event-in a recipient-friendly manner is generally referred to as the explainability of an AI model's prediction.

In the literature, the terms "explainability" and "interpretability" are used interchangeably. Therefore, it is clear that explainability in the context of an intelligent system (i.e., an AI-based system) goes beyond interpretability in terms of significance, thoroughness, and forecast accuracy.

The proliferation of XAI techniques has made it more difficult to weigh the benefits, drawbacks, and comparative advantages of the many fields. The different XAI techniques also differ greatly from one another. A technique could be ante hoc (i.e., part of the pretraining phase), postdoc (i.e., work on a previously trained model), surrogate (i.e., use a simple model to mimic the prediction of a...