Generative artificial intelligence in engineering design: Between optimisation and imagination

Generative artificial intelligence has moved from a research curiosity to a central preoccupation of the engineering-design community in the space of a few years. This commentary offers a critical map of the field, distinguishing three overlapping waves: the established discipline of topology optimization and generative design; deep generative models—generative adversarial networks, variational autoencoders, and, most recently, diffusion models—that synthesize designs from data; and the large-language-model turn that has brought foundation models into the conceptual and documentary phases of design. I argue that the most consequential questions raised by generative design are not about raw capability but about evaluation, trust, and the division of labor between humans and machines. Off-the-shelf machine learning metrics reward statistical resemblance rather than engineering validity, large models hallucinate and violate physical constraints, and the manufacturability of visually impressive outputs is frequently doubtful. Reviewing the evidence, I contend that generative tools are best understood as instruments that expand the designer’s search over possibilities rather than as autonomous designers, and that the field’s central task is to build the evaluation, verification, and accountability infrastructure that would make generative design trustworthy. I conclude with an agenda for this task.