Generative molecular design has rapidly expanded the capacity of computational systems to propose novel chemical structures, optimize predicted molecular properties, and explore chemical spaces that would be difficult to enumerate manually. Yet the pharmaceutical meaning of these capabilities remains unsettled. A structurally valid or computationally optimized molecule is not necessarily synthesizable, pharmacologically selective, developable, safe, or therapeutically superior. This critical review examines how evidence for generative molecular design should be interpreted across progressively demanding stages of pharmaceutical relevance. The selected literature was appraised according to the definition of model success, representation and data dependence, objective-function validity, synthetic feasibility, medicinal-chemistry suitability, experimental confirmation, developability evidence, comparator quality, reproducibility, and downstream progression. The synthesis shows that generative systems have convincingly demonstrated molecular representation, novelty generation, distribution learning, and optimization against specified computational objectives. Evidence is also emerging that synthesis-aware workflows can produce experimentally realizable and biologically active compounds in selected discovery programmes. However, much of the literature remains concentrated at the level of benchmark performance, predictor-mediated optimization, or isolated prospective examples. Evidence for consistent gains in selectivity, safety, pharmacokinetics, lead quality, candidate progression, or clinical benefit is substantially weaker. The article therefore proposes a graded evidentiary interpretation in which plausible molecular generation, predicted pharmaceutical utility, experimental confirmation, preclinical viability, and improved medicines are treated as distinct claim levels. This framework is not presented as a validated standard, but as a critical tool for preventing computational success from being mistaken for pharmaceutical improvement. Progress will depend on stronger comparators, prospective failure-inclusive evaluation, synthesis execution, multidimensional property testing, calibrated uncertainty, independent replication, and longitudinal reporting beyond initial compound generation.