What counts as seeing a city

A camera can show a bus passing an intersection. Deciding what that means for a neighborhood requires more than recognizing the bus.

That gap is a useful way into How AI Sees the City: Urban Visual Intelligence. In a September 24 report, MIT News describes the book, published that month by Routledge, by Fábio Duarte, Martina Mazzarello, Carlo Ratti and Fan Zhang. Its account presents visual AI as a way to investigate traffic, public-space use and street-level greenery, while warning about surveillance and bias. [1]

For residents, the potential value lies in better questions: where might an intersection need attention, or which spaces appear underused? Images could help target further investigation. They cannot, on their own, explain why people avoid a place or decide what should change.

From vehicles to estimated emissions

An earlier study, published on April 2 in Nature Sustainability, offers a concrete example. Songhua Hu and colleagues combined traffic-camera footage with mobile-phone data to estimate vehicle emissions across Manhattan. MIT News explains that vehicle recognition and movement information were combined with emissions rates. [2][3]

These are different steps. Recognition assigns a vehicle to a category. An emissions model uses traffic information and rates to calculate expected releases. Direct air-quality measurement instead measures pollutant concentrations in the air. A camera-based emissions estimate does not establish what a pedestrian inhaled at that corner.

The paper’s abstract reports that, eight weeks after congestion pricing began, traffic volume had fallen about 10%, alongside a modeled emissions decline of 16–22%. MIT News dates the policy’s introduction to January 2025. These are the researchers’ reported findings, not a fresh September result or a direct measurement of pollution exposure. The full paper was not accessible for this article; the abstract alone is insufficient to assess the causal analysis. [2][3]

The questions a map cannot settle

Privacy deserves the same separation between a stated practice and a demonstrated safeguard. MIT News says the traffic study did not compile license-plate numbers. The paper’s public data statement says Cuebiq supplied mobile-location data under agreements that prevent public sharing. Neither statement, by itself, establishes that every privacy risk has been resolved. [2][3]

The book’s authors also warn, in MIT News’s account, that training data can reinforce social biases. [1] For a city considering these tools, that suggests questions about coverage as well as recognition: which streets, times and people are represented? If some neighborhoods are poorly captured, treating the resulting map as equally informative everywhere could misdirect attention. This is a deployment risk to examine, not a bias finding established here about the Manhattan study.

Our editorial view is that residents should ask what decision the system will inform, which outputs are observations and which are estimates, and how errors will be checked locally. Who can access the underlying images and location records, how long are they retained, and who can challenge a conclusion? Officials should explain these choices before using a map to justify a policy.

Sources & context

  1. The promise and peril of using visual AI to study citiesPeter Dizikes | MIT News | September 24, 2026

    [1] Full institutional article consulted; existing package material reused. Supports the book hook, authors, suggested urban uses and their concerns about surveillance and bias. The book itself was not read. This and [2] are institutional communication, not independent corroboration.

  2. MIT researchers measure traffic emissions, to the block, in real-timePeter Dizikes | MIT News | April 2, 2026

    [2] Full institutional article consulted. Supports the account of camera, mobility and emissions-rate inputs, January 2025 policy timing, and the stated absence of license-plate compilation. The article reports these as MIT findings; privacy protections were not independently audited.

  3. Ubiquitous data-driven framework for traffic emission estimation and policy evaluationSonghua Hu and colleagues | Nature Sustainability | April 2, 2026 | DOI: 10.1038/s41893-026-01797-9

    [3] Original publisher abstract and public data-availability statement consulted. Supports Manhattan application, eight-week comparison, 10% traffic decline and modeled emissions decline of 16–22%; Cuebiq mobility data have access restrictions. Full paper was subscription-restricted; supplementary PDF could not be retrieved. No methods replication, causal audit or independent validation was performed.

AI assisted with preparing and checking this article. The responsible editor accepted this version for publication following an assisted review of facts, sources and attribution.

Related reading