Can artificial intelligence deliver more efficient, and fair, city public services? It depends.
Artificial intelligence (AI) boosterism has woven itself into almost every facet of our lives. From targeted advertising to calculating credit scores, from medical triage to automated vehicle controls, AI is rapidly shifting how we make decisions and who, or what, influences our behavior. The integration of AI into the management and planning of cities has been a slower process, first emerging with the “smart city” narrative of the large tech companies and urban developers that began almost 20 years ago.
Technology offers an opportunity to improve the delivery of and access to urban public services, but it also raises real concerns about data and algorithmic bias and the fairness of how and what decisions are made. At its core, city government provides the basic necessities that make dense agglomerations of people livable. This includes protecting residents; providing water, energy and transportation infrastructure; and making sure that garbage is removed, the streetlights stay on and our homes are safe. Emerging sensor technologies, such as the New York City Department of Environmental Protection’s smart water meters that track near-real-time usage or air quality monitors that identify particulate concentrations, provide new streams of data for city agencies to process, integrate and analyze. Add these to the administrative and other records that agencies routinely collect, like building permit applications, traffic tickets and building violations, and cities have access to vast — and diverse — data resources.
Artificial intelligence is a blanket term that covers a range of algorithms and models, from now-familiar large language models such as Claude or ChatGPT to a host of other methods in machine learning, network science and spatial analytics. One thing that current AI models do well is to quickly identify patterns in complex data and then find ways to optimize those patterns toward some simple objective. A disconcerting example of this is online gambling: Machine learning algorithms are able to identify the gamblers who are most likely to lose money, and direct behavioral prompts and game designs to keep them betting.
From the city perspective, these same capabilities seem well suited to a common service-delivery problem: How do you allocate limited resources to deliver the best outcomes? These resources could be housing inspectors or road repair crews, or new bikeshare infrastructure or a mobile police kiosk. But it is in defining the “best” outcomes that urban problems are inherently different from commercial uses of AI; the city is not a single thing; rather, it is made up of a cacophony of different groups, interests and communities, each with their own priorities, needs and values. This is where the smart city folks failed: They approached the city as just another system to be optimized toward some universal goal of efficiency. In the urban context, embedded biases and systemic discrimination challenge utilitarian measures of efficiency, and thus make handing over urban public-sector decisions to machines problematic.
To provide city services fairly, city agencies first need to hear from residents. The proliferation of 311 resident reporting systems across North American cities is an important example of how information technologies facilitate this type of engagement. By providing a mechanism for residents to request specific services and complain about problems in their community, city agencies can get a near-real-time view of conditions and needs across different neighborhoods. New York City’s 311 system, run through the City’s Office of Technology and Innovation, receives about 37 million contacts per year. This level of engagement creates a trove of useful data that opens opportunities to complement human know-how with machine intelligence.
But these data also create a number of fairness and equity concerns. There has been a lot of interest in algorithmic bias — the bias that emerges from “black box” AI models that can reinforce existing disparities. Research in my lab has focused on another aspect of the AI bias problem: biased data.
It is in defining the “best” outcomes that urban problems are inherently different from commercial uses of AI; the city is not a single thing; rather, it is made up of a cacophony of different groups, interests and communities, each with their own priorities, needs and values.
Cities have more data than ever. Requests to 311 are a good example, but so are GPS traces from vehicles and connected devices, smart water meter data and bikeshare usage. Each type of data comes with its own potential biases, whether from reporting bias in 311 requests, or coverage issues in GPS traces, or representativeness bias in bikeshare or other infrastructure-dependent systems. Data bias is a central aspect of algorithmic bias, which makes understanding data such an important step. Having more data does not mean having the right data. Understanding how data were collected, for what purpose and by whom influences how bias propagates through a decision-making process. Collectively, we can refer to this as data provenance.
Let’s consider a hypothetical example: Suppose there are 1,000 pothole complaints in a city, and five repair crews to address them. One way to allocate those crews would be to simply send a crew to each complaint as it comes in, a first-come, first-served model. Another approach would be to use the data to identify spatial hotspots, and position crews to repair potholes in those areas with the highest density of complaints. And a baseline method could simply be to let the crews drive around and repair the potholes they see, basically ignoring the complaints altogether.
All of these approaches have trade-offs along the spectrum of efficiency and equity. And when we consider data bias, the picture becomes more complicated: What we intuitively know, and what our research has demonstrated, is that people do not complain at the same rate. These differences among groups, neighborhoods and individuals stem from a range of causes, such as differing levels of trust in government, doubts around the government’s ability (or willingness) to respond, and a cultural or personal aversion to complaining about something. When we drill down further, reporting propensity differentials on particular issues (like potholes) can also be a function of residents having bigger problems to worry about. Households without heat or hot water in the middle of winter are probably not going to complain about a pothole in the street.
The trade-off is weighing the likelihood an algorithm will fail, versus the likelihood of the status quo failing.
Artificial intelligence algorithms can propagate bias from data, to models, and then to outputs. For example, using a predictive model on 311 complaint data without accounting for reporting propensities will typically result in an unfair outcome because the model does not know what lies outside the training datasets. My team’s research found this for potholes in Kansas City, where predictive algorithms trained on 311 data overestimated repair needs in wealthier neighborhoods with objectively better street conditions.
Therefore, the challenges of using AI for predictive city service delivery do not end with the data and algorithms. The model outputs have to be used to decide how to allocate scarce resources, and how to do so in a way that prioritizes those most in need. This kind of decision-making cannot be easily replaced by machines. For one, setting priorities, balancing competing needs and understanding residents’ stressors are tasks that collective, engagement-based decision-making is designed to do. Second, and as importantly, public-sector decisions require accountability. It is not sufficient to say that a building was left without heat for weeks because an algorithm did not see it as a problem. People need to make decisions and to be held accountable for the process.
This accountability cuts both ways. The potential for a model to fail, resulting in a resident not getting the help they need, has often been cited as a reason not to use data analytics in city agencies. On the other hand, predictive modeling can be effective in identifying risk, and by not using AI tools, a city agency may be missing an opportunity to improve city services. The trade-off is weighing the likelihood an algorithm will fail, versus the likelihood of the status quo failing. The risk here is compounded by the reality that vulnerable and underrepresented groups often bear a disproportionate burden when a failure occurs.
Artificial intelligence can certainly add efficiency to city service resource allocations. Even baseline data analytics, including structuring and storing data so that they are accessible and easily machine-readable, can be useful in making more informed decisions. But without (1) expertise in how the models work, (2) an understanding and appreciation of data provenance, and (3) guardrails on how decisions are ultimately made by ensuring human knowledge is prioritized, cities seeking greater efficiency through AI risk making existing inequalities worse. City governments are there to balance, respect and incorporate the differing needs, priorities and values of their residents. Artificial intelligence can help deliver better services, but not without recognizing and accounting for data bias and embedding fairness into a data-driven decision-making process.
