Smart cities began as an engineering idea about sensors and data. A new UN toolbox recasts them around residents, drawing on 15 cities that used AI in public services, and one that switched it off.
A smart city uses digital technology to run itself better. Sensors count traffic and adjust signals, meters flag water leaks, and apps let residents report a broken streetlight or book a council service. Software then turns that data into decisions about where to send crews, buses or money. Councils adopt these systems to respond faster and spend less.
The ITU, the UN’s agency for digital technology, agreed an official definition in 2016. It describes a smart sustainable city as one that uses information and communication technologies “and other means to improve quality of life”. The same definition adds efficient services, economic competitiveness and the needs of future generations. In 2023 the UN-Habitat Assembly adopted its own guidelines. Under them, a “people-centred” smart city is developed with residents, gives everyone fair access to digital services and protects human rights. Technology should serve sustainable development and stay a means to it.
From sensors to services
Early smart city schemes focused on computing and modelling urban systems, then on connectivity and data. A new UN report says that phase often left investment poorly connected to local priorities. It also raised risks of surveillance, cyberattack and algorithmic discrimination.
In Toronto, Alphabet’s Sidewalk Labs spent two and a half years planning a data-rich waterfront district at Quayside. It walked away in May 2020, citing economic uncertainty, after months of public concern over data collection. The project became shorthand for a city designed around a technology company’s ambitions.
Chatbots now answer residents, algorithms rank roads for repair and software helps decide who gets investigated for benefit fraud. Cities already house 45% of the world’s 8.2 billion people, and the report expects about two thirds of population growth to 2050 to happen in them. More than 1.1 billion people live in informal settlements or slums, which official data often barely records.
A traffic light for algorithms
The ITU and UN-Habitat built that report as the first part of a toolbox for city governments. They launched it at the UN’s first pavilion at the Smart City Expo World Congress in Barcelona. It tells mayors to start from a problem residents face and keep decisions with named officials. It also treats switching an AI system off as a legitimate result.
Every AI use gets a traffic-light rating. Green covers office tasks such as drafting and transcription, which need a staff policy and a human check. Amber covers chatbots and maintenance forecasting, which also need local testing and an easy route to a person. Red covers decisions on benefits, housing, emergency dispatch and facial recognition in public space. These require a human rights assessment before launch, plus independent audit. Qatar’s AI building-permit checker, meant to cut waits from up to 30 days to 120 minutes, sits in red because permits carry legal and safety consequences.
Each city should also put five basics in place within 100 days. They include an inventory of the AI it already runs and contract clauses that stop suppliers training models on city data. Doreen Bogdan-Martin, ITU secretary-general, said smart cities should be judged on “how well people benefit from technology designed to improve everyday living”.
Three cities, three lessons
Helsinki and Amsterdam launched public AI registers in 2020. Each entry lists an algorithm’s purpose, data, human oversight and a contact point. Eurocities later turned the format into a shared standard that other cities can reuse. The report holds it up as proof that transparency can start small and cheaply.
In eThekwini, the municipality around Durban, more than a quarter of 4.2 million residents live in informal areas. UNITAC’s BEAM tool mapped 1,530,546 building footprints across the city in about 72 hours. Conventional surveys took months or years to do the same job. Planners now use the data to target housing and services at settlements that records used to miss.
In Rotterdam, the income department used a machine-learning model from 2017 to score about 30,000 welfare recipients for fraud checks. In 2021 the city’s Court of Audit found a risk of biased outcomes. Rotterdam then stopped using the model and returned to random selection. Lighthouse Reports and Wired later reconstructed the system and showed that gender, parenthood and Dutch language skills pushed scores up. The toolbox lists the city’s decision to stop as a responsible outcome.
Who the data leaves out
Informal work, unregistered homes and missing civil records leave whole groups out of the data that AI systems learn from. As a result, the report says, a model can overlook entire neighbourhoods when a council decides where to invest. Anacláudia Rossbach, UN-Habitat’s executive director, said AI must serve communities, “especially those living in informal settlements and other underserved areas”.
The guide is voluntary and defers to national law, such as the EU AI Act’s rights assessments for high-risk public uses. Councils can adopt its principles by resolution and publish a register inside the 100 days the toolbox suggests. The harder step is the one Rotterdam took when evidence turned against its own system.
Subscribe
Sign-up to receive our newsletter
