![Overgrown vacation lot in front of city skyline](/sites/g/files/omnuum10826/files/styles/hwp_21_9__1920x825/public/2026-09/AdobeStock_374505400%20blight%20overgrown.jpg?itok=JiMPGPpl) # New Tools to Improve the Quality of City Spaces Two cities are using cameras and AI-powered vision language to more quickly track neighborhood nuisances, from abandoned cars to code violations. A system that finds problems earlier produces better property conditions at a lower enforcement cost. September 09, 2026 [ Stephen Goldsmith ](/stephen-goldsmith) *This article originally appeared in Governing Magazine.* As a prosecutor, I read James Q. Wilson and George L. Kelling’s influential [“broken windows” article](https://www.theatlantic.com/magazine/archive/1982/03/broken-windows/304465/) through a criminal-justice lens. Later, as a mayor, I found their observation — that one unrepaired broken window signals that no one cares, so breaking more costs nothing — meant far more. The quality of streets, parks and sidewalks shapes social capital, and with it a community‘s hope, outlook and confidence in the future. As Yuhan Yao and colleagues observe in [a recent study](https://www.mdpi.com/2073-445X/15/2/244) of vision language models for urban perception, a growing body of research confirms that more attractive streetscapes are linked to residents‘ mental and physical health. A built environment‘s visual quality, safety, vitality and social infrastructure determine how people feel about a place and become measures of a neighborhood‘s success. Enforcement of local laws dealing with abandoned cars, graffiti and land use are fundamental determinants of urban livability. Wilson and Kelling stressed not just the harm of neighborhood nuisances but the speed of remediation. Artificial intelligence generally, and [vision language models](https://www.nvidia.com/en-us/glossary/vision-language-models/) in particular, now give cities new tools to improve the quality of city spaces. Cleveland‘s [pilot deployment of such a system](https://www.clevelandohio.gov/311/citizen-support-vehicle) in partnership with City Detect shows what this looks like in practice: a 311-branded, camera-equipped “city support vehicle” to perform a citywide property survey to detect emerging issues, enable validation of inspectors’ notes and, in one neighborhood, a tire-dumping review. A neighbor who calls about a collapsing porch, a pile of dumped tires or grass grown four feet high wants compliance, not a penalty. Worse, when a problem persists, residents stop calling altogether and the city loses the reports it relies on to find conditions at all. For every week the city takes to find and address the problem, conditions worsen — the tires attract more tires, the vacant lot becomes a dumping ground. Slow responses invite the next violation. Cities cannot inspect their way out of this with headcount. Cleveland has some 18,000 vacant lots. Elizabeth Crowe, Cleveland‘s chief innovation and technology officer, describes the moment the math became undeniable. After a 90-minute meeting with IT, public works, and building and housing to map how the city would track high grass, issue a citation, send a crew to cut it and bill the owner, the group concluded, in her words, that “there has to be a better way.” Dallas, which covers nearly 400 square miles, faced the same constraint. Brita Andercheck, the city‘s chief data officer, recalls the objection her team raised when the code department first proposed camera-based detection: “If we use this technology and increase the identification of infractions, how will we ever have enough code officers to examine all these visual notices?” More detection without more capacity would produce a longer queue, not a solution. Escaping that trap requires changing the objective from writing tickets to improving neighborhood conditions. According to Andercheck, a letter from the city produces roughly an 80 percent voluntary compliance rate. If 4 out of 5 property owners fix the problem on notice, then the scarce resource is not the citation — it is knowing that a condition exists and how quickly someone in authority notifies the owner. A system that finds problems earlier and notifies faster produces better property conditions at a lower enforcement cost than one that finds fewer conditions and litigates them harder. The measure of success becomes the porch repaired, not the ticket written. ## What the Machines Actually Do When the scarce resource is knowing that a condition exists, the cheapest place to produce it is a city vehicle already driving past it. Dallas mounted its cameras on sanitation brush trucks. “We chose those trucks because they had a good overview of the city,” Andercheck says, and City Detect’s vision language technology is used “to identify high weeds, code violations, properties that are in blighted condition, graffiti, etc.” Cleveland began its City Detect partnership in August 2025 and by December had mounted two side-facing cameras on a city car. Crowe says that by this May the car’s survey had covered roughly 158,000 parcels — work that in the past would have required 40 officers over six months. The machine does not write tickets. It nominates conditions for remediation. Cleveland officials review the pictures before anyone drives out to inspect building violations or illegal dumping. Cleveland is also using the technology to close complaints: Weeks after sending a letter, staff can check a subsequent City Detect image to see whether the condition had been addressed, which minimizes the need to send an inspector out a second time. In Dallas, the initial rollout identified nearly 29,000 potential property maintenance concerns across 49 issue types and roughly 12,900 right-of-way encroachment issues. “All AI-generated detections are reviewed and validated by city staff before any enforcement action is taken," says Code Compliance Director Chris Christian. Andercheck describes how the camera feeds inverted Dallas’ old process, in which code officers visited a site only after a neighbor’s report. The city no longer must ration discovery to the number of officers it can put in cars. The cameras change what officers spend their hours on: Instead of driving routes to find conditions, inspectors work a ranked queue of conditions already found. Two measures are worth watching. The first is confirmation rate: What share of visual nominations do inspectors validate as real violations, and how has that share moved as the cities tuned their models? Cleveland’s confirmation against inspectors’ notes should reveal it. The second is time to resolution — how many days now elapse between a condition appearing and an owner receiving a notice, how that compares with the pre-camera baseline, and whether more cases now close through voluntary compliance rather than enforcement. Detection volume proves the camera works. Confirmation rate proves the model is trustworthy. ## How Vision Language Models Change the Economics What makes that volume cheap has less to do with the cameras than with what the models can now be asked. In earlier generations of municipal computer vision, pavement scoring systems rated roads and license plate readers matched plates. Each answered one question, and a second question meant buying a second system. Vision language models, by contrast, describe and reason over an image in open vocabulary, which means a city can put new questions to imagery it has already collected without retraining a bespoke model. The same pass down a street can support code enforcement, 311, public works and storm-damage assessment. That turns a blight camera into general civic infrastructure. For Crowe, step zero was to get clear on the problem and to create forums where people can come together to map a process. Only after Cleveland mapped the grass-cutting process end to end, for example, did it start looking for technology to address it: “We didn’t go tech first. We went problem first. And that’s, I think, helped guide how we prioritize the work.” ## Guarding Against Abuses Residents also deserve protection and privacy: Any powerful tool brings risks, beginning with where to draw the line. The defensible line separates property-facing imaging — parcels, pavement, structures, lots — from person-facing imaging: faces and individual behavior. Dallas frames its program that way: The cameras capture conditions only from the public right of way during routine city operations. Similarly, Cleveland defers to Ohio sunshine and code enforcement law: The city may photograph only from the right of way. Other limits help too, such as data retention. For example, Dallas requires the vendor to delete images that show no violation rather than store them. Keeping a human between detection and consequence offers another layer of protection. Both cities have made this structural rather than discretionary: Cleveland writes no citation from a camera image alone, and Dallas validates every AI detection with staff before enforcement. Fixed-purpose sensors gave cities the answer to one question. Vision language models paired with vehicles already making their rounds give cities an actionable visual record of the public realm. The detection layer largely works; the remaining problems are institutional: validating machine findings against human inspectors, routing nominations into work that gets done, and setting durable rules about which questions a city may ask of the imagery. Dallas and Cleveland answer all three in public, which provides a reassuring and critical layer of trust. ## About the Author ### Stephen Goldsmith ![Headshot of Stephen Goldsmith](/sites/g/files/omnuum10826/files/styles/hwp_1_1__100x100_scale/public/datasmart/files/goldsmith_headshot_2018.jpg?itok=_stVEJro) Stephen Goldsmith is the Derek Bok Professor of the Practice of Urban Policy at the Harvard Kennedy School and the director of Data-Smart City Solutions at the Bloomberg Center for Cities at Harvard University. He previously served as the mayor of Indianapolis and deputy major of New York City. Share on:- [ Facebook ](#) - [ Twitter ](#) - [ Linkedin ](#)