Generative artificial intelligence is moving into the streets, buildings and infrastructure of modern cities, promising to change how urban life is observed, modelled and managed. A new Review in Nature Cities examines what these systems can realistically contribute to urban science and practice—and where excitement is running ahead of evidence. The authors argue that the most important question is not whether AI can produce impressive images or fluent explanations, but whether it can generate reliable knowledge for decisions that affect millions of people.
Cities are unusually difficult environments for artificial intelligence because they are shaped by countless interacting systems. Transport, housing, energy use, public health, employment, weather and social behavior all influence one another, while the data describing them come in radically different forms. Satellite images, traffic sensors, administrative records, mobile-phone traces, social-media posts and interviews may each reveal only one part of the urban picture. Generative AI is designed to work across such heterogeneous information, potentially connecting visual, numerical, textual and geospatial data in ways that conventional analytical tools often cannot.
One of the most visible applications is the synthesis and interpretation of urban imagery. Generative models can create realistic scenes of streets, buildings and public spaces, while computer-vision systems can analyze photographs, satellite images and video to identify features such as land use, road conditions, vegetation or informal settlements. These capabilities could help planners explore how a redesigned street might look, compare alternative development scenarios or detect changes across large areas. Yet visual realism is not the same as geographic accuracy. An image may appear convincing while misrepresenting scale, materials, accessibility or the people who actually use a place.
The Review also explores the possibility of using generative models to represent human behavior. AI systems can simulate how different groups might respond to changes in transport prices, housing policies, public spaces or emergency conditions. In principle, such models could help researchers test scenarios before implementing costly interventions. They might also support “what-if” analysis in urban digital twins—computational representations of cities that combine physical infrastructure with streams of real-world data. However, a model that produces plausible behavior is not automatically a model that predicts behavior. Human decisions are influenced by culture, inequality, habits, institutions and unexpected events that may be poorly captured in training data.
This distinction between plausibility and validity is central to the emerging field. Large language models and other generative systems are trained to identify patterns and produce likely outputs, rather than to establish causal relationships. They can summarize evidence, generate hypotheses and translate complex information into accessible language, but they may also invent sources, reproduce statistical biases or present uncertain conclusions with unwarranted confidence. In urban policy, these failures can have concrete consequences. A flawed recommendation about flood protection, public transport or housing allocation may reinforce existing inequalities or direct resources away from communities that need them most.
Validation therefore becomes more demanding than simply checking whether an AI-generated output looks persuasive. Researchers need to compare model predictions with observations collected in the relevant city and under conditions similar to those in which the system will be used. They must test performance across neighborhoods, demographic groups and time periods, while examining whether results remain stable when data sources change. Independent replication, uncertainty estimates and transparent documentation are essential. A model that performs well in one city may fail in another because urban systems are shaped by distinct laws, infrastructure, histories and patterns of social behavior.
Data governance presents another major challenge. Urban AI systems may depend on sensitive information, including location traces, income records, health data, surveillance footage and details about movement through public space. Even when datasets are anonymized, combining multiple sources can make individuals or small groups identifiable. Generative models may also memorize or reproduce information from their training data. Responsible deployment therefore requires privacy protections, secure data handling, clear rules about consent and limits on secondary use. Technical safeguards alone cannot resolve these questions; cities must also establish legal and institutional accountability.
The authors emphasize that generative AI should support, rather than replace, human judgment. Planners, engineers, social scientists, public officials and local residents bring contextual knowledge that may be absent from a model’s data. Human oversight is especially important when AI outputs influence decisions about vulnerable populations or the distribution of public services. Participatory processes could allow communities to challenge assumptions, identify missing information and evaluate proposed interventions. This approach treats AI as a tool for expanding the range of scenarios people can examine, not as an authority that determines what a city should become.
Used carefully, the technology could accelerate several parts of the urban research cycle. Generative systems may help organize fragmented datasets, translate technical findings, create preliminary design alternatives and reveal connections across disciplines. They could make complex planning information easier for the public to understand and enable faster communication between researchers and decision-makers. But the Review makes clear that speed and convenience are not substitutes for evidence. The strongest applications will be those in which AI-generated insights are combined with conventional statistics, field observations, expert review and direct engagement with affected communities.
The arrival of generative AI offers cities a powerful new set of instruments, but not a shortcut around uncertainty. Its future in urban science will depend on whether institutions can evaluate systems honestly, recognize their limits and build safeguards before deployment becomes routine. The central test will be practical and ethical at once: can these models help cities understand change without obscuring who produced the knowledge, whose experience is missing and who remains responsible when an automated recommendation goes wrong?
Subject of Research: Generative artificial intelligence in urban science and urban practice
Article Title: Generative AI in urban science and practice
Article References: Zhang, Y., Xu, F., Wang, Q.R. et al. Generative AI in urban science and practice. Nature Cities (2026). https://doi.org/10.1038/s44284-026-00492-2
Image Credits: AI Generated
DOI: https://doi.org/10.1038/s44284-026-00492-2
Keywords: generative AI, urban science, urban planning, smart cities, human behavior modelling, urban digital twins, multimodal data, AI governance, responsible AI, city management
