Text-to-geometry AI can generate visually convincing CAD models, but producing the correct shape does not necessarily mean creating a functional engineering component. Maor Farid, co-founder and CEO of Leo AI, uses the simple lock washer to illustrate this fundamental limitation of current generative design systems, tells Machine Design.
A lock washer appears geometrically simple: a ring with a split and twist. Yet its dimensions encode specific engineering functions. Its free height determines the preload stored when a bolt is tightened, while its ends interact with mating surfaces to resist loosening under vibration. Even small dimensional or material changes can prevent the washer from performing correctly.
Current text-to-geometry systems primarily learn relationships between language and visible shapes. CAD files contain geometry but rarely explain why engineers selected particular dimensions. Consequently, an AI system can reproduce the appearance of a lock washer without understanding which dimensions determine its mechanical behavior. Ironically, complex components such as brackets may hide these shortcomings because many dimensions can vary without compromising functionality. Simple standardized components leave much less room for error.
The problem becomes more serious because incorrect geometry may not produce an immediate warning. Bad software code often generates an error, but flawed CAD geometry can render and export successfully, pass visual inspection, and fail only during testing or service.
Farid argues that engineering-grade AI needs three capabilities. Models should learn from physics, standards, load cases, and failure modes rather than geometry alone. Generated designs should provide sources, such as relevant standards, tables, or internal drawings, so engineers can verify decisions. AI systems should also recognize when available information is insufficient and decline to generate unsupported designs.
For engineers evaluating text-to-geometry tools, simple components may therefore provide the toughest tests. A system that can explain and source the functional decisions behind an ordinary part demonstrates something more valuable than impressive visualization: genuine engineering reasoning.
