Prospering on an urban planet

Introduction

From London, UK (DeepMind), to Menlo Park, USA (Meta), San Francisco, USA (OpenAI), Beijing, China (Baidu), and numerous others, cities are hubs for artificial intelligence (AI) development and practical applications. Investments in AI in cities (urban AI) are exponentially increasing. For example, some sources estimate that AI in sustainable urban planning will increase by 20% annually, from $12 billion in 2022 to over $54 billion by 2030.¹

The efficiency and capacities of AI, such as the rapid cycle from data collection to model output, influence the very core of urban research, policy,
and practice. AI can support urban planning for example by summarizing, analyzing, and providing feedback on plans, helping to incorporate diverse perspectives on urban challenges and AI applications. It also shows promise in enabling two-way information exchange between planning authorities and the public, thereby supporting participatory planning.² Multi-stakeholder partnerships and platforms emerge that can jointly contribute to further urban AI research, develop industry solutions, and support policy. Examples include the UN and government-led, Africa-focused initiative AI Hub for Sustainable Development,³ think tanks such as the Paris-based Urban AI4, and transdisciplinary academic institutions such as MILA⁵ in Quebec, Canada.

However, research on AI and the digitalization of urban areas, the so-called “smart city” discourse, has been fragmented, mainly focused on the built and engineered environment, and mainly highlighted positive contributions of technologies and successful projects.⁶

Urban AI research, development, and deployment need to recognize the complex interactions between society, ecology, and technology in urban areas. Cities function as interacting social-ecological-technological systems (SETS).^9,10 They are shaped by patterns of architectural design, people’s habits such as movements and activities, the norms and rules of the governing institutions, and the resources locally and from afar that sustain the urban fabric. As data-generating, pattern-generated systems, cities can
both inform and benefit from AI. For example, in Singapore, the advanced digital twin Virtual Singapore⁷ integrates data from both above- and below-ground infrastructure. It is one of the application areas highlighted in the National AI Strategy,⁸ providing managers with diverse datasets and supporting a wide range of uses. These include educating the public about urban nature, managing traffic in real time, and enabling long-term planning through scenario modeling, such as evaluating strategies for responding to
disease outbreaks.

Cities are both contributors to and recipients of global trends such as climate change, and are projected to experience increasingly turbulent futures. Changes such as increasing temperatures, increasing frequency and intensity of floodings, increasingly expansive and dense built-up areas, and biodiversity decline often co-occur⁹ and create compound effects that planning has traditionally struggled to manage. For example, urban greenspaces are widely recognized for providing essential services, such as shade and rainwater absorption. However, their average share of urban landscapes decreased from 19.5% in 1990 to 13.9% in 2020.¹⁰ Meanwhile, it’s estimated that by 2040, only 1% of the urban populations could escape temperature increases.¹⁰ Urban populations are projected to continue to increase. From 2018 to 2050, 2.5 billion more people are projected to be living in cities, primarily in low- to lower-middle-income countries in Asia and sub-Saharan Africa.¹¹

It is crucial to harness the potential in the burgeoning emergence of AI to support urban sustainability and ensure planetary stability. On an urban planet, urban resilience is central to planetary sustainability. The often superhuman capacities of urban AI could provide unprecedented potential to support development toward desirable urban and planetary futures. However, responsible use of AI,¹¹ and recognition of the diversity and complexity of urban SETS in research and development, is critical to ensuring urban development is equitable, resilient, and sustainable.

AI for Urban Data

With AI, urban governance is becoming increasingly autonomous and data-centric.¹² The influence of AI on urban societies, infrastructure, and institutions, and how cities are monitored, controlled, and planned, is rapidly growing. In China, for example, the (Alibaba) City Brain is an AI-powered, computer vision system intended to optimize traffic flows and support traffic planning.¹³ Cities and urban regions generate massive amounts of data through sensors, cameras, and monitoring systems, requiring increasing use of AI to manage data, make decisions, and integrate data for interoperability. It is not only cities that generate data; people do as well. Digital footprints from social media platforms, mobile apps, and other user-generated content are increasingly leveraged as data sources for AI models applied in urban analytics and planning.¹⁴ At the same time, cities are also massive consumers of data to guide urban planning, development, policy, management, and more, thus serving as key systems for the continued development of AI.¹⁵

AI has shown unprecedented capacities to analyze different and large quantities of data, often in real time, and dealing with non-linear data, such as the impact of urban features on surface temperatures.¹⁶ Urban AI data has made valuable contributions, for example, to mapping,¹⁷ identifying,^18,19 assessing,²⁰ and monitoring^21,22 urban green-blue and built infrastructure, and understanding user perceptions.^19,23,24 Increasingly high-resolution remote sensing (satellite) data is improving the precision of AI models and supports more informed planning. Lesser-known areas can be better understood, such as informal urban areas (often called slums or favelas when mainly inhabited by the financially poor), which often have limited population data and can be difficult to access.²⁵ AI models can produce urban plans and greenspace recommendations with superhuman efficiency, for example suggesting designs that increase people’s proximity to greenspaces.²⁶

AI can support combinations of data, such as local sensors and remote sensing, for mapping urban areas, for example night lights, vegetation cover, and socioeconomic factors such as population density.²⁷ As was shown in the Three Gorges Reservoir Area in China, AI has the capacity to identify complex social-ecological relationships in urbanizing regions, for example assessing the presence and behavior of ecosystem services bundles and their drivers.²⁸

AI-Assisted Urban Predictive Modeling

AI is increasingly used to predict and simulate complex urban systems, enabling more accurate forecasting of trends such as land-use/ cover changes and land surface temperature,²⁹ traffic demand and flow,³⁰ and energy and water demand.^31–33 Increasingly spatiotemporally precise urban models contribute to, for example, forecasting weather, hazards, and their impact. The capability stems from AI’s ability to a) process vast amounts of diverse data, b) combine geospatial, weather, human, infrastructure, and critical facilities data,³⁴ and c) identify complex patterns. Advances in explainable AI can identify, for example, which urban infrastructure features or inequalities lead to hazard predictions (e.g., air pollution or urban heat³³). The advances allow at least a partial view or some light into the AI “black box” of algorithmic processing.

AI can generate forecasts and nowcasts,³⁵ and increasingly realistic models of future hazard and flood damage patterns.³⁶ As AI modeling moves beyond simple random-forest models, and incorporates the spatial and temporal structure of urban environmental cities through graph networks with spatial and temporal convolutions (or transformers), urban digital twins begin to better approximate real urban environments.³⁴ For example, flood depths and extent are predicted not just from physically based models or weather data and remote sensing, but are updated with human telemetry (e.g., movement or cell phone use) and urban resident reports of traffic to cities or on social media.³⁵ AI can combine traffic, flood depth, power grid, human movement, and critical facilities data to assess if evacuation orders were effective, where people moved, and which hospitals or food facilities are compromised or stressed (e.g., in Hurricane Beryl, see Resilitix³⁷).

The generative capabilities of urban AI extend beyond mere predictions to actively shape the future of our built environments based on predictive insights. For instance, Mustafa et al. (2020)³⁸ demonstrated this by using an AI model to generate urban layout designs specifically engineered to be more resistant to urban flooding. This model achieved flood mitigation by optimizing variables such as the redistribution of green areas and the redesign of road networks to reduce water depth. This showcases how urban AI moves beyond simply predicting problems to also predict, design, and optimize potential solutions. Research is still limited regarding impact forecasting, communication of information (such as early warning signals), and the potential of support for rapid decision-making.³⁹

AI-Assisted Urban DecisionMaking

Urban AI offers capabilities that can deepen the understanding of urban areas by merging different data for sensing, imaging, and mapping, often in real time,⁴⁰ and presenting complex and detailed data analyses and predictions to support decision-making⁴¹ (see Theme box 3, ClimateIQ: Empowering Communities to Prepare for Climate Threats with Hyperlocal data). Its support spans crucial decision-making phases— for example, from problem identification, where AI-powered analyses of real-time traffic sensor data can swiftly pinpoint congestion patterns⁴²; option generation and evaluation, where AI models can simulate the impact of new zoning regulations on housing affordability⁴³; resource allocation, where AI optimizes the deployment of limited urban resources based on data-driven insights⁴⁴; to performance monitoring and adaptive management, such as analyzing bus passenger ridership data to optimize demand.⁴⁵

Beyond these direct decision-making applications, urban AI contributes to a deeper understanding of urban built-up structures,⁴⁶ processes such as land-use change,^47,48 and proposals for innovations, such as new and resource-efficient approaches to urban agriculture.⁴⁹ Foundation models using self-supervised learning can facilitate the development of other digital applications⁵⁰ and could potentially revolutionize decision-making by offering area-specific machine capacities.³⁹ After large language models (LLMs), which typically deal with text, large multimodal models (LMMs) are emerging that can combine text, images, and audio.⁵¹ The improved observation, imaging, and analysis capabilities of urban AI provide potential to increase the understanding of urban challenges and identify (multifunctional) solutions appropriate for the local context.

Potential

We present some key areas where urban AI can support prosperity on an urban planet:

Spatially precise, detailed, and comprehensive SETS modeling. The enhanced forecast capacities and detailed understanding of potential urban scenarios provided by some AI systems can support solutions designed for the local area and that take into account the interactions among multiple urban systems domains.⁴⁰ Graph networks are presenting new opportunities, allowing modeling of the spatial temporal and connected structure of urban elements across diverse data, such as traffic flow simulations. Importantly, graph networks allow using urban data for more realistic urban digital twins and improved urban structures.

Powerful scenarios and predictions, from spatiotemporally precise hazard prediction to impact forecast, communication, and decision support.³⁹ AI acting inside digital twins offers unparalleled potential to examine how climate change and other urban stressors can be mitigated and adapted to.

Improving social sensing capabilities. AI capacities to preprocess and incorporate social sensing and human telemetry data, such as Flickr photos, are improving. The capacities can mitigate some of the known biases related to, for example, social media, and contribute to more nuanced understandings of how urban spaces are used and function.

Limitations

Urban AI research incorporates different biases. Urban AI studies tend to focus on a few specific social media apps and their users, large cities, and cities in upper-middle- to high-income countries. Urban areas tend to be highly heterogeneous, which can increase the risk of so-called transfer bias,¹¹ in which models trained on data from one region risk producing incorrect and potentially damaging recommendations when applied to another region. For example, sudden and intense precipitation might have little impact in one urban area, whereas in another it could cause flooding or landslides.⁵² AI has also been found to reinforce existing socioeconomic differences, in part due to unequal representation in training data.^53–55 Such bias risks limiting the potential of AI models to identify sustainable urban solutions.

Studies have mostly focused on the built-up and engineered environment. Green-blue infrastructure and ecosystem services are still relatively under-explored. There remains untapped opportunity to use AI to better understand and address biodiversity loss and ecological resilience, which is critical to social and infrastructure resilience.

Limited preparedness. Studies indicate a low awareness among urban planners of how to utilize AI.⁵⁶ This lack of knowledge and preparedness often stems from curriculum gaps in higher education, a scarcity of accessible and tailored training for professionals, limited exposure to compelling real-world case studies, and resource constraints within planning departments. Furthermore, for many, AI can still be perceived as a “black box,” contributing to skepticism.⁵⁷ The lack of knowledge and preparedness could become a problem not necessarily through low uptake, but through damaging uptake, for example of AI systems with limited capacities to deal with given tasks or specific SETS contexts.

AI’s ability to manage interconnected shocks in urban SETS remains underexplored. While AI typically learns from routine patterns, like traffic flows,⁵⁸ urban futures are expected to be increasingly turbulent and diverge from historical trends (see chapter 1, Preparing for a Future of Interconnected Shocks).^59,60 Research has also primarily focused on isolated elements of SETS, such as the effect of expanding urban areas on the provisioning of ecosystem services. Predicting the behavior and impact of shocks is especially challenging due to the complexity and variability of urban SETS, which differ across socio-economic, climatic, ecological, and infrastructural conditions, even within short distances or timeframes. Urban AI modeling holds promise for identifying responses to change, tipping points (when change becomes irreversible), feedback loops (self-reinforcing desired or undesired behaviors), cascading effects across SETS–and strategies for adapting to, and mitigating turbulence. Yet despite its potential to support sustainable and livable urban futures, this research area remains largely untapped.