Artificial intelligence

"AI in cities" is rarely one technology. In the documented record it is four distinct jobs: watching, sorting, forecasting and scheduling. Separating them matters, because they differ completely in maturity, in evidence quality and in how much trouble they can cause.

This guide covers more than 40 documented entries across more than 30 cities in over 20 countries. The oldest is not recent at all: Los Angeles has run algorithmic signal control since 1984. Cities have automated decisions for forty years. What changed is the label, the scope — and the number of cameras.

What does AI actually run in cities today?

Four distinct jobs recur across the documented record:

How well is any of this actually known?

Of the AI-tagged entries in the atlas that carry an evidence grade, the overwhelming majority are graded official — the operator's or the city's own account. Exactly one is peer-reviewed. That is the single most important fact on this page, and it is not an accusation: vendors and agencies publish what they measure, and nobody has funded the independent evaluations.

It does mean the well-known numbers should be read with their source attached. Hangzhou's City Brain reports roughly 15% higher travel speeds and halved emergency-response times — Alibaba's figures, not independently audited. Tianjin's automated terminal reports 60% fewer staff and 17%+ less energy — the operator's and Huawei's. A State Grid pilot cut blackout repair time from six-to-ten hours to about three seconds — state media, self-reported. Every one of those entries says so on its own page.

Which urban AI actually works day to day?

The sorting layer — routing citizen reports — is the boring, working part. Amsterdam's Signalen is the most instructive AI deployment in the atlas precisely because it is unglamorous. It auto-categorises public-space reports at a macro F1-score of 0.88 — and anything the classifier is less than 40% confident about goes to a human. The design states its own error rate and builds the fallback into the workflow.

Buenos Aires' Boti makes the complementary point about channels rather than models: launched on WhatsApp in 2019 rather than as a municipal app, it went from ~600,000 monthly conversations in 2020 to over 26 million in a single quarter by 2022. Its documented lesson is that meeting residents in the app they already use beat building a better one.

Where has city AI failed — and why?

On trust, not accuracy. The largest AI-adjacent failure in the atlas was never switched on. Sidewalk Toronto was cancelled before construction, and the independently documented reason was unresolved data-privacy governance, not technical feasibility. Technology-first urbanism without public trust is dead on arrival.

Kuala Lumpur's record shows the live version of the same problem. Officials report street crime down sharply and credit the camera-and-AI network — self-reported figures — while the rollout outran the country's data-protection safeguards, from which government bodies are exempt, leaving facial-recognition oversight contested. And Hangzhou's own recorded lesson is the blunt one: big gains are available when a single platform gets all the data, at a privacy price most democracies will not pay.

The counter-move exists and is also in the atlas. Amsterdam's algorithm register — launched with Helsinki in 2020, absorbed into the Dutch national register in 2025 — publishes which systems the administration runs, for what purpose, on what data. Sensor registers and privacy registers extend the same idea. Cities that deploy AI without one are asking for trust they have not offered a way to check.

Questions to ask before procurement

• Which of the four jobs is this? A classifier that routes complaints and a camera network that identifies people are not the same purchase.
• What is the error rate, and who catches errors? Amsterdam's 40% confidence threshold is a model: state the number, name the human fallback.
• Who verified the vendor's figures? "Not independently audited" should appear in the procurement file, not only in a footnote.
• Is it registered? If the city would not publish the system in an algorithm register, that is worth understanding before signing.
• What happens to the data? Retention, access and deletion rules decide more of the public reaction than accuracy does.
• What is the non-AI baseline? Several documented gains — clearer processes, better sensors, more staff — did not require a model.

The practical takeaway: the AI that works in cities today is narrow, scheduled and boring — routing reports, timing signals, predicting leaks and load. The ambitious versions fail on governance rather than mathematics. Buy the narrow one, publish what you run, and treat any unaudited percentage as a claim until somebody outside the contract checks it.

Frequently asked