Research Question

How do users experience, perceive, or relate to the workings and effects of AI systems that mediate their everyday lives? In this work, I explore this broader question in the context of civic AI systems, due to their dominant yet highly invisible nature. Specifically, I select predictive policing as a case study. Predictive policing systems have been in existence for over a decade now making it an ideal case to answer this question.

Methods

To answer this question, I employed a ‘research through design’ methodology and conducted eight participatory mapping workshops with diverse publics such as— police reform groups, city planners, neighborhood and community leaders, civic development agency, educators, and students.

The workshop call can be found here. The workshops were conducted in person in Atlanta, Georgia located in the Southeast United States. They were scheduled to be 90 mins long. Before each workshop, a short survey was shared with the participants. The survey served two primary purposes: (1) it confirmed participation through an informal registration process, and (2) it provided workshop moderators with background information that was helpful for unique workshops.

In the workshop, participants were asked to make spatial predictions as an AI system would. They identified places on the map that they were familiar with and believed a predictive tool may identify as a crime hotspot. They were then asked to reflect on their predictions by considering their own reasoning and the effects of their predictions. The workshops followed a loose protocol that guided such reflection. Participants contemplated an AI system’s prediction goal, data type, data source and amount, spatial aggregation of data, and prediction effects in relation to their own speculations. In doing so, moderators prompted participants to consider the socio-political qualities of places they identified in relation to these technical aspects of the AI system. A focus on specific locations revealed spatial, social, political, and historical characteristics relevant to the technical workings of the predictive tool.

The protocol was documented as a toolkit and shared as a take-away artifact with the workshop participants to support their questioning of AI systems beyond the workshops. It is essential to note that this structure of the toolkit was not rigid and merely acted as a guide to the conversation while allowing the group to deviate into explaining and questioning components of AI systems that were most relevant.

Findings