Opening the Black-Box of AI: Challenging Pattern Robustness and Improving Theorizing through Explainable AI Methods

Machine Learning (ML) algorithms, as approach to Artificial Intelligence (AI), show unprecedented analytical capabilities and tremendous potential for pattern detection in large data sets. Despite researchers showing great interest in these methodologies, ML remains largely underutilized, because the algorithms are a black-box, preventing the interpretation of learned models. Recent research on explainable artificial intelligence (XAI) sheds light on these models by allowing researchers to identify the main determinants of a prediction through post-hoc analyses. Thereby, XAI affords the opportunity to critically reflect on identified patterns, offering the opportunity to enhance decision making and theorizing based on these patterns. Based on two large and publicly available data sets, we show that different variables within the same data set can generate models with similar predictive accuracy. In exploring this issue, we develop guidelines and recommendations for the effective use of XAI in research and particularly for theorizing from identified patterns