Generative AI models have reshaped how we write, code, and illustrate—but not how we design. Current models produce statistically plausible outputs, yet engineering design requires precise constraint satisfaction, human-centered integration, and generality across domains. First, I examine the tension between probabilistic models and the reliability demands of high-stakes engineering. I introduce a counterexample-driven training paradigm that yields order-of-magnitude gains in constraint satisfaction and data efficiency. Second, I illustrate how AI's disconnect from engineering and design practice limits its adoption. I present human-centered benchmarks and methods that forego end-to-end design in favor of small, surgical design modifications. Third, I showcase synthetic-data pipelines that allow models to acquire general engineering knowledge, enabling a single model to solve disparate design problems. Together, these advances demonstrate AI systems that satisfy constraints, operate within design workflows, and generalize across domains—extending generative AI beyond statistical plausibility toward engineering-grade intelligence.