Five years of roundtable dialogue chart a responsible path for Urban AI

Artificial intelligence is reshaping how cities are planned, but a new review article argues that the technology’s future in planning depends less on algorithms than on institutions, communities and human judgment. The paper, published in Frontiers of Urban and Rural Planning, synthesizes five years of interdisciplinary dialogue from the Urban AI in Planning roundtable, most recently convened at the 2025 Annual Conference of the Association of Collegiate Schools of Planning. Initiated in 2021 and supported by the Urban Data Science Working Group since 2022, the roundtable has brought together scholars and practitioners from urban planning, geography, architecture, AI, transportation and professional design practice to ask a deceptively simple question: how can AI be integrated into planning responsibly?

The authors, led by Xinyue Ye with contributors including James Nguyen H Spencer, Wei Li, Yuhao Kang, Ahoura Zandiatashbar and Junghwan Kim, applied qualitative thematic analysis to the recurring discussions across the roundtable’s five-year history. Rather than documenting a single panel, the paper distills broader principles from the collective dialogue. Three interconnected themes emerged. First, Urban AI works best as a decision-support capacity amplifier that complements rather than replaces human judgment. Second, uneven geographies of data availability, privacy expectations and model performance create disparities between urban, rural and resource-constrained settings. Third, interdisciplinary collaboration, planning education and community engagement are essential for building trustworthy and socially responsive AI systems. Across all three themes, persistent concerns include data bias, transparency, governance and ethical accountability.

The first principle, human-AI complementarity, was framed most concretely through the experience of small rural towns. Wei Li described communities that lack dedicated planners, where a single staff member may spend only a fraction of their time on planning while juggling other duties. A community meeting might generate fifty pages of notes that must be processed into something meaningful for a planning document, a task that is nearly impossible without help. Generative AI and large language models can summarize such material into key points, freeing planners to focus on human-centered conversations with stakeholders. Practitioners, Li reported, tell him they finally have time to do the work planners are supposed to do. The panelists were careful to stress that AI should strengthen rather than outsource critical thinking, since models tend to produce responses that please users rather than challenge them, rarely saying no and potentially trapping users in a narrow bubble of agreeable information.

The discussion also confronted a persistent misconception about what AI systems actually know. James Spencer argued that the central issue is not the sophistication of AI itself but the nature of the data on which it relies. There is a common belief that generative AI draws on everything on the internet, yet the internet is neither a complete repository of knowledge nor a source containing the most meaningful information for planning or governance. In an urban resilience project in Lake Charles, Louisiana, thousands of property records contain critical attributes, such as the number of bedrooms, that are absent from public online databases. Spencer’s conclusion was blunt: unless communities invest in producing, organizing and maintaining the foundational data that AI requires, the potential of generative models in planning will remain fundamentally limited, and possibly misguided.

The second principle, geographic equity, emerged from evidence that AI performance itself is spatially uneven. Junghwan Kim cited his own research showing that generative AI repeatedly performs worse on spatial questions in low-density rural areas than in high-density urban ones. When he asked ChatGPT about environmental justice issues across more than 3,000 United States counties, the responses for urban jurisdictions were substantially more location-specific than those for rural counties. The same pattern appeared in a seemingly simple task such as identifying neighboring counties, where the model performed reliably for urban contexts but far less accurately in rural regions. Kim warned that if planners in small and mid-sized communities rely on these tools without recognizing their spatial limitations, the resulting analyses and decisions may be distorted by incomplete or inaccurate information.

Privacy expectations also diverge sharply by place. Wei Li described a project involving an autonomous mobility service in a community of about 6,000 residents, where the vehicle’s distinctive appearance made users readily identifiable. When it stopped in a driveway, neighbors often inferred who was riding and for what purpose, creating discomfort for individuals using the telemedicine function who preferred not to disclose their health needs. The experience demonstrated that assumptions based on conventional data or big-city planning practices do not always translate to rural communities, and that issues of privacy, autonomy and personal dignity may differ substantially from what textbooks or metropolitan-derived frameworks present. Spencer added a methodological caution from peri-urban settlements, where a car visible in satellite imagery may function not only as transportation but as a store of wealth, meaning that what appears to be an indicator of transportation access may also be a proxy for socioeconomic status. Careful qualitative interpretation, he argued, is needed before classifying ambiguous visual information into coded datasets.

The third principle, collaborative governance, was illustrated through examples of cross-disciplinary and community partnership. Yuhao Kang described his work at the MIT Senseable City Lab partnering with the city planning office in Stockholm, where planners articulated problems and he translated their needs into computational terms. The collaboration succeeded because expertise was complementary and interdependent: he lacked their practical and institutional understanding of urban systems, and they lacked his AI expertise. Kang also noted that computer science is advancing methods intended to address some of the concerns raised, including synthetic data generation and federated learning for privacy protection and emerging algorithms capable of mitigating spatial unevenness in imbalanced datasets, while acknowledging that algorithms alone cannot resolve the broader ethical challenges inherent in planning applications.

Community engagement emerged as the most emphatic recommendation of the discussion. Wei Li reported that public anxiety about AI, including fears about job loss and opaque decision making, often stems from systems developed by a small number of corporations without meaningful community involvement. To bridge this gap, his team invited high school students, retired teachers and local residents to participate in training autonomous driving models by labeling objects such as cars, dogs and road markings. Even if their contributions were imperfect, participants felt ownership of the model and recognized that the technology deployed in their community was partly shaped by their input, fostering greater trust and acceptance. The shift from technology deployed to communities toward technology developed with communities, Li argued, fundamentally changes the social contract surrounding AI adoption. Spencer connected this to a real-data, SimCity-like platform integrating flood modeling, storm modeling, real estate information, demographic trends and infrastructure configurations, whose legitimacy depends on whether stakeholders actively participate in shaping the model itself.

Education featured prominently as a pathway toward responsible adoption. Kim described teaching AI literacy, encompassing the ability to understand, use and evaluate AI systems and to engage ethically with issues such as algorithmic bias, as a foundational element that should inform pedagogical strategies across housing, transportation and environmental planning curricula. Zandiatashbar emphasized that students must first understand what datasets actually represent, their limitations and biases, before working with large-scale models, and that relying solely on black-box AI models is not sufficient. The panelists also noted that generative AI itself can lower disciplinary barriers by functioning as a real-time translator across disciplinary languages, clarifying unfamiliar concepts during meetings and making interdisciplinary dialogue smoother.

The paper concludes that the central challenge facing Urban AI is no longer whether AI can support planning tasks, but how planning institutions can govern, interpret and integrate AI into democratic decision-making processes. As AI becomes embedded within the operational fabric of cities, in domains such as transportation, energy systems, environmental monitoring and public safety, the authors argue that society will require robust coordination and governance mechanisms allowing models, agents, data streams, devices and human actors to interact reliably. Technologies such as GeoAI, spatial knowledge graphs and digital twin environments are identified as important foundations for the knowledge infrastructure this will require. The evolution of responsible Urban AI, the authors write, depends as much on institutional capacity, collaborative governance and public trust as on advances in computational methods. The study is based on qualitative insights from the roundtable discussion, with no primary datasets generated or analyzed, and its conclusions reflect the perspectives of the participating scholars and practitioners rather than empirical measurements of AI outcomes in planning practice.

Subject of Research: Responsible integration of artificial intelligence into urban planning practice, education and community decision making

Article Title: From vision to practice: five years of responsible Urban AI and community insight

Article References: Ye, X., Spencer, J. N. H., Li, W., Kang, Y., Zandiatashbar, A., Kim, J., & Feng, X. (2026). From vision to practice: five years of responsible Urban AI and community insight. Frontiers of Urban and Rural Planning, 4(1), Article 20. https://doi.org/10.1007/s44243-026-00093-6

Image Credits: AI Generated

DOI: 10.1007/s44243-026-00093-6

Keywords: Urban AI, urban planning, AI ethics, geographic equity, collaborative governance, AI literacy, rural data gaps, community engagement, decision support, digital twins, planning education, AI bias