Artificial intelligence explanations are increasingly presented in pharmaceutical research as molecular saliency maps, ranked descriptors, pathway-level importance scores, covariate effects, counterfactual examples, and local prediction narratives. Although these outputs can make model behaviour more visible, visibility alone does not establish that an explanation is pharmacologically meaningful, scientifically dependable, or suitable for influencing a development decision. This article addresses the unresolved distinction between visually interpretable model output and pharmacologically useful explanation. It develops a proposed explanation-evaluation theory in which usefulness is defined relationally: an explanation must answer a specified scientific question for an identified user, support a bounded pharmaceutical decision, remain faithful to the model being interpreted, demonstrate context-appropriate stability, align cautiously with relevant chemical, biological, or pharmacometric knowledge, and communicate uncertainty without converting association into mechanism or prediction into experimental confirmation. The central contribution is the proposed Pharmacological Usefulness Test for AI Explanations, a non-compensatory evaluation construct in which visual attractiveness, predictive performance, user satisfaction, or apparent mechanistic plausibility cannot offset failure in a critical evidentiary dimension. The theory is designed to support comparative evaluation across molecular, omics, and pharmacometric models while preserving domain-specific evidence requirements. It also identifies failure modes arising from unstable attribution, inappropriate explanatory questions, mechanistic overinterpretation, uncalibrated confidence, automation bias, and insufficient human contestability. The proposed test is conceptual rather than empirically validated and does not establish clinical utility, regulatory acceptability, universal thresholds, or deployment readiness. Its intended value is to organize future explanation studies around pharmaceutical questions, evidentiary boundaries, and decision consequences rather than the persuasive appearance of explanatory graphics.