AI for Cities: Patterns, Gaps, and Future Directions

Abstract

This study presents a systematic review of how artificial intelligence is being applied to cities, mapping the thematic structure, methodological choices, and sustainability orientation of the field across 7,660 peer-reviewed articles published between January 2020 and December 2025. The corpus was assembled from Scopus following PRISMA 2020 guidelines, restricted to English-language journal articles, and screened against pre-specified inclusion criteria. Of the retained articles, 5,441 were classified as empirical and form the basis of the methodological and thematic analyses; the remaining 2,219 are conceptual or review pieces. Classification used a hybrid pipeline that combined keyword extraction, large language model-assisted coding under deterministic decoding settings, and expert adjudication, with inter-coder reliability for a 10% double-coded sample reaching substantial-to-almost-perfect agreement across all primary dimensions.

The first headline finding is the rate of expansion. The annual output of urban-AI articles roughly quadrupled, from 574 in 2020 to 2,395 in 2025, with research originating from 109 countries. Three countries, China, India, and the United States, account for more than half of the corpus, and China alone contributes 37.2 percent. Although the volume signals a maturing field, the geographic concentration and the English-language filter mean that local governance practices and Chinese-, Lusophone-, and Spanish-language scholarship are systematically underrepresented in the international literature.

The second finding concerns what AI is actually being asked to do in urban contexts. Prediction and classification together account for over 40 percent of empirical applications, while decision-support tasks, the category most aligned with the needs of policymakers, represent fewer than one percent. Methods native to urban complexity, such as graph neural networks (n=111) and physics-informed neural networks, remain rare. The field has largely imported architectures developed for computer vision and natural language processing instead of designing tools around the relational, multi-scale, and socio-technical structure of urban systems.

The third finding emerges from a correspondence analysis whose first two principal axes accounted for 41.7 and 23.4 percent of total inertia. The horizontal axis separates technical-predictive research from socio-spatial integration, and the vertical axis separates black-box from interpretable methods. Climate and Environment work clusters with interpretable, ensemble-based methods; Social Equity work clusters with simulation and decision-oriented analysis; Digital and Smart Cities work, by contrast, sits in the technical-predictive, black-box quadrant. Funding patterns broadly mirror this geometry: Chinese funding clusters with deep-learning, SDG 11 work, while EU Horizon and North American funding shows broader thematic and methodological diversity.

The fourth finding is a sustainability paradox. The technically most advanced cluster, Digital and Smart Cities, has the weakest sustainability integration: only 37 percent of its articles show strong integration, compared with 48 percent for Resilience and Safety and a clear majority for Climate and Environment. We argue that this reflects an implicit framing in which technical optimization, including lower energy intensity per task, higher throughput, and reduced latency, is treated as a proxy for sustainability while distributional outcomes, rebound effects, and the lifecycle costs of underlying digital infrastructure remain unassessed. We present this mechanism as an interpretive hypothesis suggested by the cross-tabulations rather than a tested causal claim.

The fifth finding concerns research design. Cross-sectional designs dominate at 74.3 percent, longitudinal studies remain rare, and equity-focused work is uncommon outside the Social Equity cluster. Without temporal depth, claims about whether AI interventions deliver durable sustainability gains rest on projected rather than evaluated outcomes; without equity instrumentation, distributional consequences for informal settlements and historically marginalized communities remain invisible. These design constraints also limit what can be said about how benefits and burdens are distributed in space and over time, an important blind spot for any sustainability evaluation.

From these patterns we derive a research agenda organized around three imperatives. The first is architectural innovation: building urban-native models such as graph neural networks aligned with infrastructure topology and physics-informed architectures that encode established urban theory, including gravity models, scaling laws, and spatial interaction frameworks. The second is governance integration: closing the gap between AI capability and decision support, which requires uncertainty quantification, scenario comparison, counterfactual reasoning, and the gradual incorporation of causal inference techniques. The third is participatory framing: treating affected communities as co-designers rather than as data subjects or end users, and embedding algorithmic accountability mechanisms such as public registers, impact assessments, and procurement requirements within municipal AI pipelines.

We close with two methodological caveats. The descriptive distributions and correspondence-analysis geometry reported here are direct features of the corpus, but the interpretations advanced about institutional incentives, publisher preferences, and funding effects are abductive readings drawing on prior literature; they are hypotheses for further empirical work rather than tests of those mechanisms. The language and indexing constraints of the search also mean that conclusions about regional governance practices and locally calibrated AI implementations should be read with a clear sense of what the corpus does and does not see.