FOR(GEEKS)

AI

AI can reshape cities, but planners need guardrails

A Nature Cities review finds AI could improve flood planning and public participation, but warns of bias, errors and weak accountability.

· 6 min read

Image: Techxplore

Urban planners are already testing generative AI for flood modeling, neighborhood design and public consultation—but a review of more than 100 studies says those systems need firm safeguards before they influence high-stakes decisions. Techxplore reports that the review, published in Nature Cities, examines how AI is entering urban science and practice.

The authors, Esteban Moro, a Northeastern professor of physics, and Ryan Wang, a professor of civil and environmental engineering and a member of the Network Science Institute, focused on two increasingly common categories: distribution-fitting generative models and foundation models. The research covered urban science, computational social science and geospatial AI, a field that combines artificial intelligence with mapping and satellite imagery.

From flood maps to synthetic neighborhoods

Distribution generative models can create new images or renderings after learning from large collections of examples. In an urban setting, that could mean generating a previously unseen section of a city from satellite imagery of the surrounding area.

“One example in urban settings: You could send a model a distribution of satellite images around a city, and it will be able to produce another part of the city.”

— Esteban Moro, Northeastern professor of physics

The review identifies flood prediction as one practical use. A model trained on images of flooding from different regions could generate a representation of how flooding might appear in a particular location, such as Boston’s city center. That kind of output could help planners visualize scenarios and consider mitigation strategies, although the researchers stress that generated material cannot replace validation.

The same class of tools is also being used to create renderings of potential neighborhoods. Such images may help planners communicate proposed changes, but they are representations generated from prior data—not direct measurements of what will happen in a real city.

Foundation models operate differently. They are primarily instruction-based and learn through human language, including systems such as ChatGPT, Claude and Google’s NotebookLM. In urban planning, the review describes their use in translating dense policy documents into plain English so that members of the public can understand them more easily.

AI agents join planning meetings

One of the more unusual examples in the review involved AI agents representing residents from two neighborhoods in Beijing during urban planning meetings. The agents spoke about issues that mattered to those communities and, according to the report, influenced subsequent land-development plans.

The study found improvements in measures including participant satisfaction and inclusion. Moro described the case as an example of AI supporting participation rather than simply automating a planning decision.

“The study found improvements in measures, including participant satisfaction and inclusion.”

— Esteban Moro, Northeastern professor of physics

That distinction matters. An AI system that summarizes policy or presents a resident perspective is being used differently from one that independently approves an infrastructure plan. The review presents both as emerging applications, but argues that human judgment must remain central when decisions affect safety, housing or public resources.

Bias and false information become governance risks

The researchers warn that urban AI systems can inherit and amplify biases in their training data. If historical maps, planning records or imagery reflect unequal investment or exclusionary decisions, a model trained on those materials may reproduce those patterns in its outputs.

Generative systems can also produce false or misleading information. In low-stakes design work, an inaccurate rendering may be corrected during review. In disaster response or large-scale infrastructure planning, the consequences are more serious. The authors describe inaccurate statements and the difficulty of interpreting how a model reached an answer as governance risks, not merely technical imperfections.

To address those problems, Moro and Wang propose an ethical framework built around five themes:

The framework responds to a gap the researchers see between adoption and accountability. Communities are using AI tools, but the review argues that they have not always carried out a sufficiently deep examination of what would make those systems acceptable to researchers, practitioners and the people affected by their decisions.

“Communities have adopted a lot of these AI technologies, but we haven’t seen a deep reflection on what is needed for these technologies to be accepted both at the level of our research [and] at the level of practitioners.”

— Esteban Moro, Northeastern professor of physics

Validation cannot be an optional cost

Ryan Wang said responsibility for reliability checks should be shared by AI providers and the urban planners who implement their systems. A city cannot treat a model as a neutral black box simply because it produces polished maps, text or images.

“While we use it, it’s important to evaluate it honestly. We can’t just say there’s a new toy. Let’s use it.”

— Ryan Wang, Northeastern professor of civil and environmental engineering

The review’s recommendation is straightforward: communities should build robust validation and reliability checks into their use of AI. Those checks could be expensive because they require human reviewers, domain expertise and processes for challenging or correcting model outputs.

Moro noted the economic tension. Letting a large language model operate in the background and accepting its answer at face value is cheaper than paying people to verify each response. But that cost saving shifts risk onto residents and public agencies when the system is used for emergency planning or major infrastructure decisions.

“The human experience is really valuable. We see a lot of initiatives to incorporate humans in the loop.”

— Esteban Moro, Northeastern professor of physics

The review is identified as “Generative AI in urban science and practice,” by Yunke Zhang and colleagues, and was published in Nature Cities in 2026. It does not present a single urban AI system ready for broad deployment or establish that generated flood scenarios are accurate enough to guide construction on their own. Its central finding is narrower and more consequential: AI can expand the tools available to planners, but the institutions using it must also fund the human oversight needed to check those tools.