Generative AI has become a prominent approach in de novo drug design because it can propose new chemical structures rather than only screen existing libraries. Across the past decade, the field has expanded from early autoencoder and reinforcement learning systems to graph, transformer, and diffusion-based molecular generators. This systematic review evaluated generative AI models applied to de novo molecular design. The review focused on model families, molecular representations, synthetic feasibility, pharmacological validation, evaluation metrics, and translational readiness. A systematic review was conducted according to PRISMA 2020 principles using searches of PubMed, Scopus, IEEE Xplore, and Web of Science. Records were screened by two reviewers, eligible studies were extracted using a structured form, and findings were synthesized narratively. The evidence base showed rapid methodological growth and frequent reporting of chemically valid, novel, and diverse molecules. However, synthetic feasibility was inconsistently integrated, and prospective experimental pharmacological validation was reported in only a small subset of studies. Generative AI for de novo drug design is technically sophisticated but remains incompletely translated into experimentally validated lead discovery. The main barriers are weak synthesis-aware optimization, inconsistent benchmarking, limited external validation, and scarce biological testing.