Investigating Machine Learning Accuracy, Type, and Explainability on Trust-Relevant Criteria: A Human-Centered Approach

As human-machine learning (ML) interaction increases in relevance, system designers must understand the influence of performance- and non-performance-based ML systems’ attributes on users’ trust. This understanding will ensure ML-enabled systems are appropriately relied upon, leading to performance benefits in human-ML interactions. An iterative approach was taken across three online experiments, in which participants evaluated the performance of ML-enabled classifiers varying in terms of their accuracy, type, and explainability. In many cases, all three factors interacted to shape individuals’ subjective perceptions of the algorithms and their objective task performance, demonstrating a benefit of accurate, white-box models paired with explainable artificial intelligence tools. Taken together, ML algorithm accuracy, type, and explainability affect subjective and objective trust-relevant criteria. Implications for this work concern human-centered design and an appreciation of users’ expectations and interpretations of these complex systems.