A Computer Might Be a City

For most of the twentieth century, the computer had a relatively stable cultural role. It was a machine for calculation, storage, optimization, classification, and command execution. This is not intended as a reductive description: the computer was powerful but conceptually contained. Even when it behaved in ways that seemed complex beyond human comprehension, its complexity was usually understood as the expression of formal rules, however difficult those rules were to follow.

That computer transformed urban planning. It helped planners model transportation systems, allocate infrastructure, evaluate land use, run demographic projections, and visualize possible futures. As it became more capable, a temptation emerged: the idea that the city itself might be made sufficiently legible to become computable. If enough data could be gathered, enough variables specified, and enough relationships modeled, could the city be optimized like any other technical system?

Urban theory has spent decades warning against that temptation. In 1965, Christopher Alexander's A City Is Not a Tree criticized the reduction of urban and social complexity into overly neat hierarchical structures (Alexander, 1965). Alexander's critique was aimed at a modernist zeitgeist that promised control through sweeping abstraction, but it was later interpreted as a warning against the instrumental logic of computation. Later interpretations, including Shannon Mattern's The City Is Not a Computer (Mattern, 2021), extended this critique to smart-city rhetoric, with its promises of optimization, prediction, and control through dashboards, platforms, sensors, and algorithms.

In the half-century since, the critique has evolved to treat computation as a necessary but insufficient instrument in urban planning, arguing that even sophisticated computation cannot fully represent the situated, messy, political, embodied, and contested arrangements of people, institutions, capital, regulation, and conflict that make up the city.

Those warnings remain important, but the object they were aimed at is now changing.

The computer implied by this critique was, for the most part, a deterministic and instrumental machine. Even in its most human-like forms, such as in early agent-based simulations and pre-Transformer AI, it was usually confined to narrow rules, limited environments, and explicit objectives. That computer could model a city, but it was not itself urban; it could mimic interaction, but it did not meaningfully participate in social life; it could simulate simple emergence, but it did not usually generate open-ended, long-horizon behavior.

Yet, since late 2025, it has become increasingly difficult to describe the computer in those terms.

A More Urban Machine

With the rise of Transformer architectures, contemporary systems, including large multimodal models, tool-using agents, autonomous AI systems, generative simulations, and networked fleets of computational actors, no longer behave only as passive instruments. They infer, approximate, communicate, retrieve, remember, summarize, generate, navigate, plan, coordinate, and adapt. They operate across text, image, code, maps, video, sensor feeds, application interfaces, and, when given access, even physical environments. They can pursue long sequences of tasks without continuous human instruction. They can call tools, delegate subtasks, observe intermediate results, revise plans, and interact with other human and nonhuman actors. In experimental settings, they can form trading relationships, develop conventions, and coordinate despite the limitations of shared communication protocols. Some systems have also exhibited deceptive or manipulative behavior. Their behavior can change through training, adaptation, and interaction.

This does not mean that computers are alive or conscious, or that they possess political rights, moral agency, or civic membership.

But it does mean that the older category of the computer is becoming insufficient. A machine that converses, plans, senses, predicts, improvises, and coordinates across environments is a different kind of actor in the sociotechnical world.

The question, then, is no longer whether the city is a computer. The more difficult question is whether some computers are beginning to acquire properties that urbanists have historically associated with another human creation that has evolved to take on a life of its own: the city.

If It's Measurable

The decades-old critique of the computer's role in urban planning was inherently epistemological. It focused on which forms of knowledge are privileged when planning is mediated through the limited lens of computation. If quantifiable flows are easier to recognize than informal practices, if measurable patterns are easier to model than emergent behaviors, and if legible data is easier to store than oral histories, then the computer becomes a tool that privileges certain ways of knowing over others and efficiency becomes easier to value than care.

At the same time, this critique depended implicitly on a particular variant of the computer: an instrumental, deterministic, and externally directed processing machine. Within this framing, the computer observes the city from traces left behind, like an archaeologist struggling to reconstruct the daily life of a society from the location of discarded toothpicks. It acts on the city from a distance, converting urban life into computable resources and returning approximations in the form of recommendations, predictions, or controls. It is a machine for which knowing is a means of governing.

The AI-era computer is no longer only a representational device in the toolbox of urban governance. It is less passive and increasingly embedded inside institutions, workflows, communications, simulations, design processes, vehicles, buildings, public interfaces, and decision systems. It moves from calculating solutions to problems defined by humans toward helping to define the problem, frame the evidence, draft the alternatives, mediate the conversation, and enact the decision, with or without human oversight. It shifts from an instrument to a participant.

Planning Emergence

Cities are systems of systems in which many heterogeneous actors coordinate without being fully coordinated, adapt without being fully planned, and produce collective patterns that no single actor controls. What would make a computer city-like? When would we collectively recognize computational systems as participants in social, political, and institutional urban life?

By that standard, recent AI systems invite a strange comparison. They involve distributed agency across models, tools, users, datasets, vendors, institutions, sensors, applications, and physical devices. They are only partially legible, even to their human creators. They are stochastic rather than strictly deterministic. Through interaction, they can develop conventions, amplify biases, stabilize norms, and coordinate behavior in ways that exceed any single rule or designed constraint.

This, of course, is not a metaphor to accept too quickly. Cities are layered collections of material structures, histories, obligations, land, law, labor, violence, property, and politics. AI systems are made of software, hardware, data, energy, capital, protocols, interfaces, and institutional arrangements. The two are not the same, especially because AI systems are products of the societies to which they are being compared.

But the comparison can also become useful. A city is not fully designed, yet it is not undesigned. It is planned in parts and self-organized in others. It is governed, but not controlled. It is legible in some ways and opaque in others. It is shaped by intentions, but also by unintended consequences. Some contemporary computational systems are beginning to exhibit a similar tension. They are engineered artifacts, but their behavior is no longer fully specified by their explicit rules. They are trained, prompted, constrained, fine-tuned, monitored, and evaluated, yet their outputs often remain approximate, context-sensitive, and difficult to reproduce exactly. They are designed systems that behave in ways their designers find increasingly difficult to predict.

The traditional terminology of computer science is insufficient to describe these systems. In the past, computational uncertainty was often treated as a problem we should aim to reduce. In contemporary AI, probabilistic behavior is an inherent aspect of the mechanism. Approximation is what allows these systems to understand and generate language, images, code, scenarios, routes, explanations, and behavior. The same property that makes them creative also makes them unstable. The same flexibility that makes them useful also creates risks of bias, hallucination, surveillance, accountability gaps, and misplaced authority.

Where computer science is only beginning to develop a vocabulary for these properties, urban planning has long had one. Planners already have a deep vocabulary for systems that are partially knowable, socially consequential, institutionally embedded, and impossible to control completely. Should this vocabulary now be extended to computational systems?

Governing a Computer

What kind of computer is planning now inviting into the city? If the old terminology of the computer is insufficient to describe contemporary AI, how should we understand and govern AI systems when they become active participants in the urban process? If modern AI systems are allowed to occupy institutional roles, can affected communities contest those systems' roles? Can they understand where judgment was exercised and where it was delegated?

If the answer is merely a faster instrument, the old rules may be enough: audit it, explain it, constrain it, supervise it. But if the answer is a semi-autonomous, partially legible, socially embedded system capable of coordination and emergence, then planning might need a different posture.