A few months ago researchers at MIT’s Senseable City Lab published a study that used machine learning to identify vehicle types from 331 traffic cameras across New York City and to estimate each vehicle’s emissions. With enough cameras, visual AI can monitor emissions at a precision and scale never before possible.
The team treats every city photograph as a dataset, applying computer‑vision techniques to turn images into quantifiable urban features. As MIT researcher Fábio Duarte puts it, “We can treat digital images as data and quantify city features.”
Their findings appear in the new book How AI Sees the City: Urban Visual Intelligence, co‑authored by Duarte, Martina Mazzarello, Carlo Ratti, and Fan Zhang of Peking University. The book surveys how visual AI can answer urban‑planning questions such as why traffic jams form, which intersection designs are most hazardous, and which parts of plazas attract the most people.
The authors trace a long visual tradition—from Roman marble maps and early photography to Kevin Lynch’s and William Whyte’s seminal works—arguing that visual AI extends this lineage, allowing observation of cities at unprecedented scale and detail.
Beyond emissions and traffic flow, visual AI can map street‑level activity, sidewalk usage, and urban greenery. Satellite images show tree cover, while ubiquitous phone photos reveal how often residents actually see green spaces, a factor linked to reported well‑being.
Zhang emphasizes that the real promise of visual AI is not just processing millions of pictures, but linking visible elements—streets, buildings, traffic, public spaces—to larger questions about how cities function and how people experience them.
A striking example comes from a study that analyzed 400,000 Airbnb listings worldwide. Contrary to claims of a homogenizing global interior style, the images uncovered strong geographic differences in design.
The book also warns of serious pitfalls. Ubiquitous cameras raise privacy concerns; London has about 210 cameras per square mile, while eight of the world’s ten most camera‑dense cities are in China, with Shanghai exceeding 5,000 per square mile. The authors argue that safety benefits must be weighed against the erosion of personal freedom and the risk of abuse inherent in constant monitoring.
AI systems can also reinforce social biases. Models trained primarily on majority‑group data may evaluate minority groups unfairly. Duarte notes, “AI is not neutral; it sees what we teach it.” Mazzarello adds, “Our eyes are not neutral either. Every tool must be guided and trained properly.”
The book has received praise; Michael Batty of UCL called it a “fascinating book” that shows how urban analytics, AI, and large language models can improve design.
In conclusion, the authors urge a wise, critical, and creative approach to deploying visual AI, believing that with caution and independent thinking, progress in understanding and improving our cities is possible.
Review