When No One Is to Blame: How AI Is Breaking Urban Responsibility

When a tree fell on a major power line in Switzerland in September 2003, few could have predicted what followed. The disruption rippled down through the interconnected European grid into southern Italy, leaving roughly 56 million people without electricity for hours and causing around 1.2 billion euros in damage. No single actor had failed; the catastrophe emerged from the structure of the system itself. That event has become a textbook illustration of what risk analysts call systemic risk, and it is precisely the kind of cascading failure that a new theoretical study argues will become harder and harder to assign blame for as artificial intelligence spreads through the nervous system of modern cities.

The study, published in the journal AI & Society by Daniele Chiffi and Stefano Moroni of the Politecnico di Milano, takes on one of the most uncomfortable questions raised by the smart city: when something goes badly wrong in a tangle of algorithms, sensors, contractors, agencies and citizens, who is actually responsible? The authors argue that the answer is increasingly nobody, and that this responsibility vacuum is itself a danger. Their analysis, grounded in an extensive interdisciplinary literature review, identifies three structural problems that disperse responsibility in distinct ways, and proposes a new form of planning designed to close the gaps before harm occurs rather than after.

The first and oldest of these is the problem of many hands, a concept introduced by political theorist Dennis Thompson in 1980. It describes situations in which numerous people contribute to an outcome, making it nearly impossible to say who is answerable for what. A bridge collapse is the classic case: designers, contractors, subcontractors, material suppliers, inspectors, municipal technical offices and maintenance staff may all have played a part, and any one of them may have behaved perfectly correctly. The disaster can be caused not by any individual failure but by the interaction of many individually sound behaviours. Urban traffic congestion works the same way. No single motorist is responsible for a citywide gridlock at rush hour, yet the collective behaviour of thousands of drivers leaving work simultaneously produces exactly that.

The second problem, many things, was articulated by philosopher Mark Coeckelbergh and shifts attention from people to artefacts. Modern technological action involves not just many human hands but many interconnected technologies: software, sensors, controllers, machine-learning systems and physical infrastructure, all of which causally contribute to outcomes and may interact in unexpected ways. An autonomous vehicle in mixed traffic may fail to yield to a pedestrian at an unmarked crossing because its sensors misread the scene. In automated energy grids or urban water systems, failures may stem from no single device or operator but from the complex interplay among multiple sensors, algorithms and human supervisors. Disasters, the authors argue, are manifestations of these broader dynamics rather than the result of isolated individual errors.

The third problem is the paper’s genuine novelty: the problem of many tasks. It arises specifically from general-purpose AI, the category of systems that includes large language models and multimodal foundation models. Unlike traditional fixed-purpose systems trained for a single defined function, general-purpose AI can perform, or be adapted to perform, a wide range of tasks, including some for which it was never intentionally trained. The authors analyse a task as having three components: a problem, a goal and an action. With general-purpose AI, the normative component, the goal, becomes unstable, because the system’s purposes cannot be fully specified in advance. This creates gaps in both retrospective blame and prospective prevention that fixed-purpose technologies simply do not generate.

The consequences are concrete. A general-purpose AI system deployed in a city might simultaneously analyse traffic flows, energy consumption, public health data and social media signals. Each task may look safe in isolation, but their combination, or the system’s autonomous generation of new tasks, can produce unintended consequences such as discriminatory routing or privacy violations. Where fixed-purpose surveillance AI is bounded to facial recognition or crowd detection and is therefore easier to regulate, general-purpose systems can integrate previously separate data sources and scale surveillance in ways that affect far larger populations. Harms may emerge in contexts far removed from the original deployment, making accountability extraordinarily difficult to pin down.

Crucially, the three problems are not independent. The authors describe them as progressively more complicated versions of a single dynamic: many hands concerns who acts, many things concerns how outcomes are produced, and many tasks concerns what a system does, or comes to do, over time. A general-purpose AI in a smart city may be used by many actors, embedded in a dense network of sensors and data infrastructures, and gradually drift beyond the purposes for which it was configured. Responsibility gaps are most severe where all three forms of dispersion overlap, which is why the authors insist they cannot be solved one at a time.

Their proposed remedy is a rethinking of urban planning itself. Traditional risk planning focuses on identifiable hazards, discrete technologies and clearly bounded projects, an approach the authors find increasingly insufficient. Instead, they call for a systemic, participatory and trajectory-based approach. For the many hands problem, this means making responsibility allocations explicit in advance: clarifying which institutional level, national, regional or municipal, handles what, which duties belong to elected councils versus designated agencies, and which responsibilities remain in public hands versus being assigned to private actors. The COVID-19 pandemic offered a sobering lesson here, as countries such as Italy and Spain suffered serious decision-making failures rooted in unclear divisions of responsibility between central governments and regional administrations.

For the many things problem, the authors advocate meaningful human control over autonomous and semi-autonomous systems. Drawing on the work of engineer W. D. Holford, they argue that responsibility must rest squarely on human decision-makers who are given the full means to exercise it. In practice this means humans must be able to supervise system behaviour in real time through monitoring dashboards, intervene or override decisions, contest outcomes such as AI-driven loan decisions, and continuously redesign systems to eliminate biases or vulnerabilities. Regulatory oversight, independent audits, accountability frameworks and ethical review boards are the institutional machinery that keeps human authority central. The 2019 Ethiopian Airlines Flight 302 disaster, in which an automated system repeatedly overrode pilot inputs based on faulty sensor data, illustrates what happens when that control fails.

The many tasks problem demands the most radical shift: from task-based authorisation to what the authors call trajectory-based governance. Because the relevant tasks cannot be exhaustively specified in advance, planners must monitor how a system’s role evolves over time, reassessing it not only when the technology is modified but whenever its function within the urban network changes, for instance through the integration of new data categories, extension into new decision-making domains, or a significant increase in the affected population. Planning instruments, they argue, should be conditionally open rather than indeterminately adaptive, establishing clear framework conditions while preserving room for revision under democratic control. Genuine in-person participatory processes, the authors suggest, can act as early-warning mechanisms that detect task drift before it accumulates into systemic risk. The alternative, a city governed by opaque systems whose purposes no one fully specified and whose failures no one can attribute, is not a hypothetical dystopia but a structural tendency already built into the technology now being deployed.

Subject of Research: Responsibility attribution and systemic risk in AI-enabled urban planning

Article Title: Responsibility and planning amid growing urban complexity and AI proliferation: many hands, many things, many tasks

Article References: Responsibility and planning amid growing urban complexity and AI proliferation: many hands, many things, many tasks. (n.d.). https://doi.org/10.1007/s00146-026-03285-2

Image Credits: AI Generated

DOI: 10.1007/s00146-026-03285-2

Keywords: artificial intelligence, urban planning, systemic risk, responsibility gaps, general-purpose AI, smart cities, many hands problem, meaningful human control, urban governance, AI ethics, cascading failures, trajectory-based governance