UN smart city AI toolbox puts people in charge of public decisions
Setting a positive example: Helsinki's public AI register describes systems, their purposes and the people responsible for supervising them.
When a city uses artificial intelligence (AI) to flag a welfare application, recommend a road repair or answer a resident’s question, its output can influence how public money is spent and how people are treated. Someone must remain responsible for checking that output, explaining the resulting decision and correcting mistakes.
Keeping that responsibility with people is the central aim of the United Nations’ new Toolbox for People-Centred Smart Cities in the AI Era. Announced in Barcelona on 4 November 2026, it gives mayors, city managers and policymakers practical guidance for maintaining accountability as AI becomes part of everyday government.
The resource is issued by the International Telecommunication Union (ITU), the UN agency for digital technologies, and UN-Habitat, the UN programme responsible for sustainable urbanisation. It was developed through United for Smart Sustainable Cities, or U4SSC, a UN initiative marking its tenth anniversary. Its first publication, Guiding Principles and Frameworks for People-Centred Smart Cities in the AI Era, translates that aim into an implementation process and practical safeguards.
Human accountability for AI in city services
The guidance asks cities to define the public problem before choosing a technology, identify who could benefit or be harmed, and assign responsibility throughout the project. Its eight-step implementation cycle connects those early choices with design, deployment, measurement and adaptation.
Consider software that recommends which roads should be repaired first. Processing inspection data can help a municipality identify damaged surfaces, but deciding how to distribute a limited maintenance budget also involves public priorities. Officials must be able to assess the recommendation and explain the choices they make.
To exercise meaningful control, officials need sufficient understanding to question a system’s outputs, time to consider them and authority to depart from them. An approval step offers little protection if the employee cannot assess the recommendation.
The appropriate arrangement depends on the task. A planning officer could review an AI-generated application summary before deciding; a traffic controller could supervise adaptive signals and intervene when necessary.
AI bias in Amsterdam’s welfare screening
Amsterdam’s failed pilot with the "Slimme Check" (Smart Check) system underscores why algorithmic scrutiny must persist after an AI tool goes live. Designed to flag welfare applications for fraud investigations, the system trained on historical data to identify risk patterns. However, AI is fundamentally limited by its training data. Because machine learning models replicate and amplify historical patterns, biased inputs or unrepresentative data inevitably lead to discriminatory predictions and systemic unequal treatment.
This was precisely what happened in Amsterdam. Although city officials adjusted the training data to mitigate bias after initial tests exposed disparities, a subsequent pilot run between June and August 2023 merely shifted the inequality. This time, women and Dutch nationals were significantly more likely to be falsely flagged for investigation than men and non-Dutch applicants.
Consequently, Amsterdam shelved the pilot in November 2023. Ultimately, the project yielded a vital lesson: an apparent improvement in fairness during a controlled testing phase does not guarantee equitable outcomes in practice. This failure underscores the indispensable role of human oversight in continuously evaluating AI performance and deciding whether a system remains fit for service.
New York’s MyCity chatbot gave incorrect advice
While Amsterdam’s system influenced the scrutiny of individual applications, New York’s experience demonstrates how errors can directly mislead the public through advice issued in a city’s name. In March 2024, testing by the investigative newsroom The Markup revealed that New York’s MyCity business chatbot was giving incorrect advice, falsely informing users that they could withhold employees’ tips and refuse cash payments.
Although the city defended the service as a pilot and highlighted disclaimers warning of inaccurate answers, spotting these errors required the very knowledge that business owners used the chatbot to find.
The UN toolbox addresses this friction between automated services and public accountability. It recommends providing human points of contact alongside non-digital or assisted channels for essential services, ensuring that user complaints reach someone with the authority to act.
The toolbox recommends a "traffic light" approach to governance, applying lighter safeguards for routine tasks like drafting and transcription, and far more stringent scrutiny for applications that impact legal rights, public safety, or essential services.
Higher-impact use cases include allocating housing, managing welfare benefits, and operating critical infrastructure. For these high-risk deployments, recommended protections include human rights impact assessments, independent reviews, and mandatory consultation with affected groups.
Ultimately, this risk-based framework means a chatbot providing simple library opening hours requires a completely different level of oversight from one determining eligibility for financial assistance, even if both systems rely on similar underlying technology.
Helsinki’s public AI register makes oversight visible
For residents to challenge or question an AI-supported service, they must first know exactly where automation is being deployed. Helsinki, the capital of Finland, provides a leading example of making this role visible through its public AI register, which details the systems in use, their specific purposes, and the officials responsible for overseeing them. Current entries include a parking and streets bot, a rental housing chatbot, and a library virtual assistant.
In official guidance published on 27 October 2026, Helsinki stated that its services must explicitly disclose AI usage whenever it impacts the service a customer receives, the processing of their case, or content published by the city. Under these rules, digital services indicate AI involvement directly within the user interface, and human operators verify information before it is published.
By clearly documenting specific services and their lines of accountability, the register provides residents with a vital starting point for directing questions to the appropriate authorities.
Better information, accountable choices
These safeguards can accompany practical improvements in municipal services. In South Africa’s eThekwini municipality, which includes Durban, the Building and Establishment Automated Mapper (BEAM) has helped planners obtain more complete information about buildings.
Developed by the United Nations Innovation Technology Accelerator for Cities, or UNITAC Hamburg, with municipal partners, BEAM detects rooftops in aerial or satellite imagery. UNITAC reports that it mapped more than 1.5 million building footprints across eThekwini and reduced mapping time from months to around 72 hours.
This information can support housing and infrastructure planning. Officials remain responsible for deciding how to use it, which communities receive investment and how those priorities are explained.
The guide recognises that declining or withdrawing an AI application can be a responsible outcome. Equally, a clearly defined task such as mapping gives authorities a basis for measuring benefits and deciding where to extend its use.
For a resident questioning a welfare investigation or seeking reliable advice, accountability has a practical meaning: an identifiable authority can explain what happened, review the evidence and put an error right. The toolbox aims to preserve that relationship as more public services incorporate AI.
Liked this article? You can support our independent journalism via our page on Buy Me a Coffee. It helps keep MoveTheNeedle.news focused on depth, not clicks.