Involving Communities in Model Design Could Reduce Bias in AI

From routing citizen nonemergency calls to operating affordable housing, state and local governments are increasingly relying on artificial intelligence (AI) to manage bureaucratic processes. While these tools can improve efficiency, they can also reinforce bias. Data used to train AI often reflects decades of disparities tied to gender, race, ethnicity, and age — patterns that can carry through to the very communities these systems are meant to serve.

“Organizations need to shift the way they treat bias in AI,” said Reeham Mohammed, a postdoctoral fellow in Georgia Tech’s Jimmy and Rosalynn Carter School of Public Policy. “Bias isn’t a problem that can be solved once and for all, and it affects everyone. We are all stakeholders in how this technology is implemented.”

Most efforts to address AI bias focus on fixing systems after they are built. Reeham proposes starting earlier. In a recent book chapter, she outlines an approach called participatory modeling, where stakeholders help map how AI systems function before they are deployed. The process reveals where bias can enter along the way.

Her framework draws on an unexpected comparison: the human immune system. When the body detects a virus, multiple systems respond and remain active until the threat is contained. Addressing AI bias, she argues, requires a similar approach — one that is continuous, adaptive, and systemwide.

“It takes the whole body to react to a virus, and the immune system stays active until it contains it,” Reeham said. “Stopping AI bias also requires ongoing intervention. AI is rapidly evolving, and our response should evolve with it.”

To test this idea, Reeham and her co-author, Erik Johnston of Arizona State University, wrote the recent chapter reporting on three empirical studies examining how bias can emerge across complex systems.

In Peoria, Illinois, the research mapped how health services, social services, and other support systems used AI and how it affected their operations. Three main issues stood out: Access barriers to resources kept some stakeholders out, AI systems reflected their designers more than the intended users, and the complexity of public institutions made public trust challenging.

Johnston analyzed 311 service request data, showing how citizen input, usage patterns, and response systems can reflect — and sometimes reinforce — existing disparities between neighborhoods. For example, low-income, ethnically diverse communities often don’t use 311 due to language barriers or low government trust. In a third study, Reeham conducted focus groups with 57 stakeholders at a higher education institution, including students, instructors, and administrators. These interviews explored perceptions of AI use and oversight within the higher education environment. Many of the conversations addressed faculty policies on student AI usage, with students reflecting frustration that they were not consulted in the making of these policies. 

Together, the findings show AI systems cannot be designed effectively from the top down without excluding important groups that actually use the AI services. Instead, participatory modeling calls for stakeholders to be involved early to help identify where bias begins. Whether through town halls or focus groups, stakeholders should be part of discussions when designing a new system. Stakeholders also need to remain engaged after deployment to ensure that systems stay fair; this could include everything from community advisory boards to third-party consultants.

“People should know that if they see bias in AI, they need to speak up,” Reeham said. “We often assume machines are more objective than humans, but that’s not always true.”

CITATION: Johnston, E.W., Mohammed, R.R. (2026). How Participatory Modeling Can Enable Collective Bias Mitigation when AI Is Used across Systems and Institutions. In: Ahrweiler, P., Gilbert, N. (eds) Participatory Modelling and Simulation to Improve AI-based Public Social Services. Artificial Intelligence, Simulation and Society. Springer, Cham. https://doi.org/10.1007/978-3-032-15283-1_13