Urban planners in São Paulo face the same pressure as those in Berlin: do more with less time, test more ideas faster, and justify every decision. Enter generative AI—ChatGPT and similar tools that can summarize dense research papers in seconds, draft policy scenarios, and sketch out design alternatives that might take a team weeks to explore by hand. For researchers working under deadline, it feels like a breakthrough. But urban designers and researchers warn that efficiency without scrutiny may quietly reshape how cities are imagined, without anyone fully noticing.
Generative AI is remarkably useful in urban design work. Machine learning algorithms can analyze video footage to understand pedestrian movement and traffic patterns, helping planners design safer, more efficient streets. Other systems simulate urban heat distribution or air quality across neighborhoods, enabling climate-conscious decisions about building density and green space. Still others let designers quickly test different land use scenarios—changing building heights, block sizes, or park locations—to compare outcomes before any concrete is poured. These are genuine advances, not science fiction.
Yet beneath this promise lies a question that matters: are these tools enhancing urban knowledge, or reshaping it in ways we do not fully understand? Urban design is not a neutral technical field. It has evolved differently across the world—shaped by post-war reconstruction in Europe, urban renewal policies in the United States, and entirely different pressures in Asia and Africa. Every city planning theory emerged to solve specific problems for specific communities in specific places. When large language models enter this space, they risk flattening that hard-won, context-dependent knowledge into generic, universally applicable recommendations that work nowhere well.
Recent research by urban designers and planners examined precisely this risk. Their key finding: while LLMs can genuinely speed up writing, support analysis, and help explore ideas, they also carry serious dangers—especially when their outputs are treated as fully correct or used without considering local context. The models generate plausible text based on patterns in training data, not verified truth. They can sound confident about zoning or housing policy while being entirely wrong. Sometimes they fabricate citations that do not exist, subtly embedding false authority into planning debates.
The researchers propose what they call cornerstones for responsible use—not rigid rules, but practical guides. Research direction must remain with the human planner, not the model. Engagement with AI should be critical and active, not passive acceptance of whatever appears on screen. Knowledge must stay grounded in the specific city, community, and context where it will actually be used, not treated as generalizable. And skepticism should come first: planners should not trust these tools quickly, especially in high-stakes decisions about housing, zoning, or development that affect real people's lives.
Cities are not datasets. They are lived environments shaped by history, power, culture, and the messy reality of who gets heard and who does not. AI can help designers test ideas and handle complexity. But the moment urban planners begin by asking the model what to study or how to frame a problem, they risk producing generic solutions that miss what actually matters to the people who live there. The goal is not to reject these powerful tools, but to keep humans firmly in control of the questions they answer.