Generative AI tools are inserting themselves into this translation chain. Not at the end, processing finished designs, but at the beginning — taking language directly as design input and producing spatial and structural proposals from it.

This changes the brief from a document into a prompt. And prompts behave differently from documents in ways the profession has not yet fully absorbed.

A document is static. A prompt is iterative. You refine it based on outputs. The design process becomes a dialogue between the practitioner’s intent and the model’s interpretation — faster, more exploratory, but also more susceptible to the model’s biases and training data than a human designer working from first principles.

The structural implications are significant. If early-stage massing and spatial decisions are being made by language models that have been trained on existing building typologies, the resulting proposals will cluster around what has been built before. Genuinely novel structural approaches — the ones that emerge from a specific site condition, a specific material constraint, a specific performance requirement — are less likely to surface from a model that is essentially averaging prior art.

The prompt is powerful. It is also conservative in ways that are not always visible. Knowing when to fight the model’s suggestion is the new form of structural judgment.