Are AI-driven cities optimizing for the wrong outcomes?

Artificial intelligence is no longer a future layer of urban life. It is already embedded in the systems that make cities function, from transport networks to energy grids, improving their efficiency and reducing environmental impact in real time.

This shift is often presented as a pathway to more sustainable cities. But do smarter systems necessarily produce more socially equitable ones?

Across cities, AI applications are already visible. In Barcelona and Madrid, digital twins simulate urban systems such as traffic, pollution and mobility, allowing planners to test interventions before implementing them. AI systems are optimizing traffic flows in Dubai. Companies such as Google are using AI to map urban heat islands through satellite data, helping cities identify where to plant trees or deploy cooling measures.

At the core of these developments is what is often referred to as “physical AI”: traffic sensors, pollution monitors, smart meters and cameras continuously collect data, feeding AI systems that can respond in real time. This combination of sensing and decision-making allows cities to move beyond analysis toward continuous management of infrastructure. Traffic lights adapt to changing flows, energy systems balance supply and demand, and logistics networks reroute dynamically.

These developments are changing how cities operate. AI is moving from analyzing urban systems to actively shaping them in real time. In some cases, the impact is already visible. IoT-based traffic systems in Barcelona, for example, have reduced travel times by up to 30%.

These developments are often presented as a clear technological progression toward sustainability: better data, smarter systems, more efficient cities. But they also reveal a structural challenge. While they may help cities move toward net zero targets, it is far less clear whether they make them more equitable and inclusive.

Urban systems have long been designed and governed by a limited set of actors, creating persistent blind spots in how cities function and whom they serve. There is growing recognition that many large cities are structured by “clusters of disadvantage”, where spatial, social and environmental inequalities overlap and reinforce each other. Peripheral neighbourhoods often face lower access to services, fewer economic opportunities and greater exposure to environmental risks. Even environmental performance is unevenly experienced. Poorer areas are significantly hotter due to urban design and infrastructure choices, making them less livable and more vulnerable to climate-related health risks.

AI systems do not operate independently of these patterns. They are designed to optimize measurable outcomes based on data, predefined objectives and mathematical models. Trained on existing urban data, they risk inheriting and amplifying underlying inequalities. Optimization models may prioritize areas with higher demand or better data coverage, unintentionally directing resources away from already underserved communities. Unless these underlying inequalities are explicitly accounted for, AI may reinforce and stabilize existing clusters of disadvantage rather than addressing them, further marginalizing those already most affected.

This dynamic is not limited to older cities. It is also visible in new, purpose-built AI-driven urban developments such as NEOM, Masdar City, Saad Al-Abdullah and Yiti Citi, where abundant capital, available land and highly centralized decision-making have enabled large-scale experimentation. However, the absence of legacy constraints does not automatically produce more inclusive outcomes. It can just as easily reproduce old blind spots in a more technologically advanced form.

Research shows a clear mismatch between AI capabilities and urban challenges. Even when studies claim to address these issues, explicit grounding in urban theory remains rare. A recent large-scale review estimates that just over 1% of studies directly engage with established frameworks, with nearly half of research driven primarily by technological possibilities rather than real-world problems. In effect, both research and practice tend to focus on what AI can do, rather than what cities need.

Cities risk becoming great for some people but difficult for the majority when technology amplifies existing divides rather than address them.

This challenge is compounded by uneven capacity. Discussions at the OECD’s 5th Roundtable on Smart Cities and Inclusive Growth point to significant differences in how cities adopt and govern AI, driven by disparities in skills, infrastructure and institutional resources. In practice, this often means that cities do not develop these systems themselves. They procure them.

Procurement therefore becomes a central, but often underexamined, point of control. Accountability in urban AI is not simply a technical requirement, but a social relationship that must specify who is responsible, to whom and for what. When systems are designed and deployed by a small number of private actors who also control data and infrastructure, key decisions about how cities function are effectively externalized. Once implemented, these systems can be difficult to scrutinize or contest.

This has implications for public trust. Even in highly technical domains such as autonomous systems, the challenge is not only safety, but explainability. Decisions made through complex systems are often difficult to interpret, even for those responsible for their deployment. When outcomes cannot be clearly explained or meaningfully challenged, trust becomes harder to sustain.

At the same time, these systems could be used differently. AI can support decision-making by helping policymakers explore trade-offs and assess how different groups are affected by urban interventions. But this depends on how systems are designed, and whether they incorporate the perspectives of those typically excluded from planning processes.

While the risks of "black-box" urbanism are real, some cities are already demonstrating that AI can be reclaimed as a tool for civic agency.

Helsinki and Amsterdam, for instance, have pioneered public AI registries which serve as "transparency portals" that allow citizens to see exactly which algorithms are managing their services, what data they use, and how human oversight is maintained. This shifts the role of the resident from a passive data point to an informed stakeholder.

Moreover, Amsterdam has experimented with participatory AI design, involving citizen councils, including disability advocates, to co-create "accessible route planners." Rather than optimising for the fastest path for a "standard" user, the AI is trained on the lived experience of those with mobility impairments. These examples suggest that AI’s value in the city need not be confined to narrow efficiency.

When governed through frameworks of digital sovereignty and inclusive procurement, these technologies can support more accountable decision-making. As Professor Laura Ruotsalainen notes, such tools allow us to "test" the social consequences of urban policies in virtual simulations – before they become irreversible realities on our streets.

AI can support better urban decisions if its design makes visible who benefits, who is excluded and who remains able to challenge its outcomes.