Smart Cities and AI: How Urban Centers Are Using Machine Learning to Transform Infrastructure

Where AI delivers documented results in city operations, from Surtrac's Pittsburgh pilot to Google's Project Green Light, the limits on surveillance, public AI registers, and a checklist for evaluating smart-city proposals.

PUBLISHED // UPDATED
DURATION // 08 MIN READ
Smart Cities and AI: How Urban Centers Are Using Machine Learning to Transform Infrastructure
FIG.01 // VISUAL DOCUMENTATION Smart Cities and AI: How Urban Centers Are Using Machine Learning to Transform Infrastructure

“Smart city” projects have promised for years that sensors, data, and algorithms would make urban infrastructure run itself. The record is mixed. Projects that target a narrow, measurable problem, such as traffic-signal timing or water leaks, often produce clear results. Projects sold as a citywide “AI brain” tend to stall on procurement, data sharing between departments, and public trust.

This guide reviews where AI has documented results in city operations, where the main risks lie, and how municipal leaders can evaluate proposals.

Traffic signals: the best-documented use case

Traffic-signal optimization is the most mature AI application in cities, partly because results are easy to measure: stops, wait times, travel times, and emissions at specific intersections.

Surtrac (Carnegie Mellon University). Surtrac uses decentralized control: each intersection optimizes its own timing in real time while sharing information with its neighbors. In a 2012 pilot at nine intersections in Pittsburgh’s East Liberty neighborhood, Carnegie Mellon reported reductions of 40% in vehicle wait time, nearly 26% in travel time, and 21% in projected vehicle emissions. The pilot was small, but the decentralized design is important: there is no single central system whose failure stops every intersection.

Google’s Project Green Light. Green Light takes a different approach. Rather than installing new hardware, it analyzes aggregated driving data from Google Maps to recommend signal-timing changes that city engineers can apply to existing controllers. Google says early numbers indicate a potential for up to 30% fewer stops and 10% lower greenhouse-gas emissions at affected intersections, and that the program, launched in 2023, is available in more than 100 cities, including Haifa, Kolkata, Hamburg, and Boston. These are Google’s own figures and describe potential improvements at intersections, not citywide traffic.

The two approaches illustrate a choice cities face:

ApproachExampleHardware neededMain trade-off
Real-time adaptive controlSurtracSensors and controllers at each intersectionHigher cost; responds to live conditions
Data-driven retiming recommendationsGreen LightNone beyond existing signalsLow cost; depends on engineers applying changes and on a vendor’s data

Water and energy networks

Utilities were among the first public services to use machine learning, because their networks already produce large amounts of sensor data. Common applications include:

  • Leak and burst detection in water networks, by spotting unusual flow and pressure patterns
  • Demand forecasting for electricity and water at neighborhood level
  • Managing distributed energy, such as rooftop solar and battery storage
  • Energy management in public buildings, adjusting heating, cooling, and lighting to occupancy

Vendors in these areas often publish percentage savings. Cities evaluating such claims should ask for the baseline period, the network size, and whether the results come from an independent audit or the vendor’s own analysis.

City digital twins

A digital twin is a detailed virtual model of a city used for planning and simulation. Singapore’s Virtual Singapore program, described by the Singapore government, built a 3D model of the city to support planning and analysis. Digital twins are useful for testing scenarios (new transit routes, flood risk, building placement) before committing money, but they are expensive to build and maintain, and their value depends on how current the underlying data is.

Public safety: where the controversy is

AI in public safety is the most contested area of smart-city work. Predictive policing, which uses historical data to forecast where crime will occur, has faced sustained criticism for reinforcing bias: if police are sent to predicted areas, they record more incidents there, which then feeds back into the model as confirmation.

Facial recognition has drawn even stronger responses. In 2019, San Francisco became the first major U.S. city to ban facial recognition use by city agencies, and other cities followed. In Europe, Article 5 of the EU AI Act largely prohibits real-time remote biometric identification in publicly accessible spaces for law enforcement, with narrow exceptions.

Less controversial safety applications focus on response rather than prediction: optimizing where ambulances and fire engines are stationed based on demand patterns, or detecting incidents such as fires or flooding and alerting dispatchers, with a human confirming before action.

Transparency: public AI registers

Some cities now publish registers of the algorithms they use. Amsterdam maintains an algorithm register and Helsinki publishes an AI register, describing each system’s purpose, data, and oversight. Registers do not guarantee good outcomes, but they give residents and journalists a starting point for questions, and they force departments to document what systems do.

The equity problem

A recurring criticism of smart-city projects is that benefits go first to neighborhoods that already have good services. Systems optimized on usage data favor areas that generate more data: busier transit routes, commercial districts, and neighborhoods with more connected devices. Over time, that can widen service gaps.

Cities can address this directly:

  • Require vendors to evaluate systems for disparate impact before and after deployment
  • Set targets so underserved areas receive a proportional share of improvements
  • Include community representatives in oversight of high-impact systems
  • Publish performance data by neighborhood, not only citywide averages

Lessons from Toronto’s Quayside project

The most prominent smart-city project of the last decade ended before construction began. Sidewalk Labs, an Alphabet company, had proposed redeveloping part of Toronto’s waterfront as a technology-driven neighborhood. In May 2020 it abandoned the 12-acre Quayside project. Chief executive Dan Doctoroff said that “unprecedented economic uncertainty” during the pandemic had made it “too difficult to make the 12-acre project financially viable without sacrificing core parts of the plan.” The Guardian noted that the project had also become the target of fierce criticism from privacy advocates, including the resignation of a privacy expert who had advised it.

The official reason was economic, but the long public dispute leading up to it offers lessons for any city working with a technology company:

  • Settle data governance first. Who collects data in public spaces, who owns it, who can use it, and for what purposes should be decided before design work, not negotiated later.
  • Keep the public partner in charge. Projects led by a private company’s vision struggle to earn public trust, even when the company promises safeguards.
  • Separate infrastructure from data business models. Residents are more willing to accept sensors that serve a clear public purpose, such as traffic signals or leak detection, than open-ended data collection.
  • Plan for exit. Contracts should specify what happens to data, hardware, and software if a partner withdraws.

High-risk rules for city infrastructure

In Europe, some smart-city systems now fall under specific legal obligations. Annex III of the EU AI Act lists as high-risk AI systems “intended to be used as safety components in the management and operation of critical digital infrastructure, road traffic, or in the supply of water, gas, heating or electricity.” Systems in these categories face requirements for risk management, data governance, documentation, human oversight, accuracy, and robustness.

For city officials, that means traffic-control and utility systems with AI safety functions need the same disciplined documentation and testing as other regulated systems, and procurement contracts should require vendors to supply the technical documentation needed for compliance. Even outside the EU, these requirements are a useful checklist for systems whose failure could put people at risk.

How to evaluate a smart-city proposal

QuestionWhat a good answer looks like
What specific problem does it solve?A named metric, such as stops per intersection or water lost per day
What is the baseline?Current measurements taken before deployment
Who verifies the results?City staff or an independent evaluator, not only the vendor
What data is collected, and how long is it kept?A published data inventory and retention period
Can the city switch vendors?Data and configurations can be exported in usable formats
What happens if the system fails?Safe fallback behavior, such as fixed signal timing
Is it in the public register?Yes, with purpose and oversight described

Starting with a pilot on a small number of intersections, pipes, or buildings, as Pittsburgh did with Surtrac, allows a city to measure results before committing to a larger contract.

Bottom line

The smart-city projects that deliver are usually specific: retiming signals, finding leaks, maintaining buses, and managing building energy, with results measured against a baseline. Surtrac’s pilot and Google’s Green Light figures show what targeted traffic work can achieve. The harder problems are governance: limits on surveillance, transparency through public registers, and making sure benefits reach every neighborhood. Cities that treat those as part of the project, not an afterthought, are the ones most likely to see lasting value.

Sources