Before a city agency deploys an algorithm to manage traffic signals, allocate social services, or prioritize infrastructure repairs, the technical team must step back and answer a fundamental question: how does this system impact the people it serves? An AI impact assessment is not just a compliance checkbox; it is a structured, plain-language exercise designed to identify risks, clarify responsibilities, and ensure that public technology serves the public interest. This worksheet provides a framework for civic leaders and developers to evaluate their systems before they go live, ensuring that accountability is built into the process from the start.
Who Is Affected
The first step in any assessment is to map the human landscape. It is not enough to say "the public" is affected. You must identify specific groups of residents who will interact with the system, either directly or indirectly. Consider the demographic, socioeconomic, and geographic diversity of your jurisdiction. Who is likely to be on the receiving end of the system's output? For example, if a system predicts crime hotspots for police deployment, consider which neighborhoods or communities will see increased police presence. If a system automates eligibility checks for housing assistance, identify the populations most likely to be denied or delayed. Think about vulnerable populations, including individuals with disabilities, non-native speakers, and those with limited digital access. The goal is to move beyond abstract users to specific, named groups whose daily lives may be altered by the algorithm's operation. This identification process helps prevent unintended bias and ensures that the system does not disproportionately burden marginalized communities.
What Decisions the System Influences
Next, define the specific decisions the AI system supports or makes. Is the system making the final call, or is it providing a recommendation to a human officer? Clearly distinguish between decisions that are purely informational and those that carry legal or significant administrative consequences. If a system flags a resident for a benefits review, that is a high-impact decision with tangible consequences for their livelihood. In contrast, a system that suggests maintenance schedules for streetlights has a lower individual impact but a higher collective one. For each decision point, ask: What is the authority behind this decision? Is there a clear legal basis for using an automated tool in this context? Defining the decision authority helps clarify who is responsible when the outcome is controversial or harmful. It also helps residents understand their rights in the process, knowing whether they are dealing with a human judgment or an automated calculation.
What Can Go Wrong and for Whom
This section requires honest risk analysis. Identify potential failure modes, such as data errors, model drift, or algorithmic bias. Consider the worst-case scenarios: What happens if the system consistently misclassifies a specific group of residents? What are the immediate and long-term consequences for those individuals? For instance, if a predictive maintenance system incorrectly identifies a bridge as safe, the risk is public safety. If a social services algorithm incorrectly denies aid, the risk is financial hardship and loss of essential support. You must also consider the "second-order" effects, such as how public trust in the agency might erode if errors are discovered. By explicitly linking potential failures to specific groups of people, you create a clear mandate for mitigation. This is not about fearing technology, but about preparing for it. It forces the team to ask: If this fails, who pays the price, and do we have the resources to fix it quickly?
What Data It Uses
Transparency about data inputs is critical for public trust. List the specific datasets used to train and run the system. Where does this data come from? Is it historical administrative data, sensor data, or third-party commercial data? Historical data often contains biases from past human decisions, such as policing patterns or historical housing segregation. If the system learns from biased data, it will likely reproduce or amplify those biases. Explain the data quality controls in place. How do you ensure the data is accurate, complete, and up-to-date? Also, address data privacy. What personal information is processed? Is it anonymized or pseudonymized? Do residents know their data is being used for this specific purpose? Clear data documentation helps residents understand the "black box" and provides a basis for challenging decisions if they believe the underlying data is flawed or unfair.
How Residents Learn About It and Appeal
Accountability requires accessible communication. How will residents be informed that an AI system is being used in a service they interact with? Avoid technical jargon. Use plain language to explain the system's role, its limitations, and its purpose. Where will this information be posted? Is it on the agency's main website, on the specific application form, or in community centers? Accessibility is key; ensure the information is available in multiple languages and formats. Equally important is the appeal process. If a resident disagrees with a decision influenced by the AI, what is the path to challenge it? Is there a human review option? How long does the review take? What evidence can the resident provide? The appeal process must be clear, fair, and free of excessive bureaucratic hurdles. Residents should never feel trapped by an algorithm they cannot understand or challenge.
Publish and Review
Finally, commit to ongoing transparency and evaluation. Publish a summary of this impact assessment on the city's public portal. Do not hide it in a technical appendix. Make it easy to find and read. Establish a schedule for regular reviews. AI systems are not static; they change as data changes and as models are updated. Schedule annual or bi-annual reviews to reassess the impact. Include community feedback in this review process. Hold public meetings or comment periods where residents can share their experiences with the system. Monitor performance metrics not just for efficiency, but for equity. Are certain groups facing higher error rates? Are appeal rates higher in specific demographics? Use this data to refine the system or, if necessary, to decommission it. The goal is to create a culture of continuous learning and accountability, where the city's use of AI is a living, responsive process rather than a one-time implementation. By publishing these assessments and committing to regular reviews, the city demonstrates its commitment to serving all residents equitably and transparently.
- Which specific groups of residents are most likely to be impacted by this system, and how will they be affected?
- Does the system make final decisions or only provide recommendations to humans, and what are the consequences of those decisions?
- What are the potential biases in the historical data used to train the model, and how will they be mitigated?
- How will residents be notified that an AI system is involved in their service interaction?
- What is the clear, accessible process for a resident to appeal a decision made or influenced by this system?
- How often will the system's performance and impact be reviewed, and who is responsible for acting on the findings?
Put it into practice
The Accessibility Compliance Quickstart from CivicWarden collects the templates and checklists behind this guide in ready-to-use form, so your team can apply the framework without drafting from scratch.