Artificial intelligence has become embedded in target identification, molecular representation, compound generation, virtual screening, preclinical prediction, and selected development workflows. Yet the significance of these applications remains difficult to judge because computational performance, experimental usefulness, translational value, and routine pharmaceutical readiness are frequently discussed as though they were equivalent outcomes. This critical review examines where artificial intelligence materially changes drug-discovery practice and where its contribution remains conditional, indirect, or unconfirmed. The review applies a stage-specific analytical approach that distinguishes data availability from data suitability, benchmark performance from prospective usefulness, molecular plausibility from developability, prediction from explanation or causality, and experimental confirmation from clinical translation. The evidence indicates that artificial intelligence has its most defensible value in bounded tasks involving large or structurally informative datasets, clearly specified objectives, rapid candidate prioritization, and experimental feedback. Examples include relational target analysis, focused molecular generation, structure-enabled screening, and the prioritization of compounds for assay testing. However, many reported gains remain partly attributable to dataset composition, automation, benchmark construction, chemical similarity, or evaluation choices rather than to a transferable algorithmic advantage. Evidence becomes progressively thinner when claims move from discovery-task acceleration to mechanistic validity, preclinical predictiveness, clinical benefit, regulatory acceptability, or improved portfolio success. The principal conceptual contribution of this review is a proposed value-and-limit interpretation in which claims are judged according to their application stage, validation setting, experimental confirmation, and permissible influence on pharmaceutical decisions. The synthesis is necessarily constrained by heterogeneous study designs, uneven reporting, selective publication of successful campaigns, and limited independent prospective evidence. Responsible adoption therefore requires evidence proportional to the consequence of the decision being supported rather than reliance on a generic designation of artificial-intelligence capability.