33 Association of Research Libraries Research Library Issues 299 — 2019 Proofs Proofs as explanations are testable, demonstrable, traceable, and unambiguous. In the context of AI, they pertain primarily to rule-based expert systems or systems that use decision trees (that is, an explicit knowledge basis encoded in human interpretable statements). Proofs of algorithmic predictions require clear causal links and logic statements that unambiguously trace performance from data to decision. Such an examination is possible in AI systems that employ ruled-based or decision-tree models because the rationale is specifically coded and human readable. While the performance of rule-based and decision- tree systems is inferior to that of current machine-learning techniques, these systems are still in use where explicit causality and accountability can be documented (for example, in certain health care, insurance, and public sector applications), demonstrating that, in specific circumstances, explainability is preferred over performance. Validations Validations or verifications as explanations are conclusions about the veracity of the AI substantiated by evidence and/or reason. Verification confirms the AI performance against an external measure, standard, factual data, or third-party corroboration. Feature selection is an explanatory strategy that attempts to reveal the key factors (for example, hyperparameter weights) that had a primary role in the prediction of the algorithm. By isolating or adjusting these elements, it is possible to explain the key components of the decision. There are various feature selection techniques but all of them are “decompositional” in that they attempt to reduce the work of the algorithm to its component parts and then use those results as an explanation. Feature selection provides a verification that certain elements have a primary influence on the prediction thereby explaining why a certain outcome pertains but not another (a contrastive explanation). While such an explanation is used primarily