Opening the black box: Operational principles, tools and frameworks that advance explainable artificial intelligence (XAI) models

As artificial intelligence (AI) models are increasingly becoming permeated across various domains, there are instances where they are generating hallucinations, misinformation and erroneous outputs. Various stakeholders, particularly the regulatory ones, are encouraging the developers of machine learning (ML) systems to clarify or justify their models' decisions, actions or predictions in a way that is understandable to their users. In this light, this article raises awareness on Explainable Artificial Intelligence (XAI) principles that are intended to increase transparency, accountability and fairness about the modus operandi of machine learning algorithms. A systematic review of the extant literature identifies key tools, frameworks and best practices that enhance the interpretability of AI models, including open-source techniques like SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME), among others. The synthesis of the findings also shed light on XAI challenges and limitations of black-box models. This contribution advances a conceptual framework for the responsible implementation of XAI and offers practical guidelines that promote the interpretability of AI systems, whilst addressing their opacity, as well as their biased outcomes. It puts forward theoretical and managerial implications as well as future research avenues.