Storytelling as a Bridge Between Impacted Communities and Data-Driven Policymakers

Technical systems designed to inform data-driven policy are often built by technical experts disconnected from the communities in which their technology will be deployed. This process tends to neglect the perspectives of people impacted by those policies. As a result, technical systems and the policies they inform can lead to unanticipated dynamics—some wonderful, others unwanted or harmful—that impact people, communities, and society. In our research, funded by SSRC’s Just Tech program, we develop and test social impact scenarios—a novel method for (a) identifying and articulating unanticipated dynamics from the perspective of impacted communities and (b) effectively communicating those unanticipated dynamics and their potential impacts to technical teams and local policymakers.

What are social impact scenarios?

Social impact scenarios are concise narrative vignettes (four to seven paragraphs) rooted in empirical data. Their purpose is to mitigate the potential for unwanted or harmful impacts of technical systems by making local concerns, tensions, and tradeoffs visible and actionable to policymakers, system designers, and machine learning (ML) practitioners. The scenarios can serve as positive stories about how technical systems are used in beneficial ways (e.g., minimizing hunger), as well as precautionary tales about how things could go awry or cause harm (e.g., reinforcing existing ethnic tensions). Social impact scenarios are not abstract or generic; they are situated in a particular social, political, and cultural context and depend on the specific design of the technical system.

Social impact scenarios are developed by directly engaging with impacted communities to surface salient concerns, tensions, and tradeoffs. Once developed, social impact scenarios can help policymakers, system designers, ML practitioners, and others understand a technical system’s unintended impacts on people, communities, and society. The effective use of social impact scenarios requires seeding the scenarios with local knowledge, exploring salient themes in further research, and leveraging the scenarios as part of a reflexive technical practice.

The Togo case study

We develop and evaluate social impact scenarios in the context of a data-driven system to allocate humanitarian aid in Togo. The Togolese government used this system to provide subsistence cash transfers to the country’s poorest individuals to help them survive the worst phase of the Covid-19 pandemic. The government did not have a poverty registry to determine eligibility for the program, so it relied on nontraditional data to identify the people most in need. Specifically, it used satellite images to identify the poorest regions of the country, then used mobile phone metadata to identify the poorest individuals within those regions.1Emily Aiken et al., “Machine Learning and Mobile Phone Data Can Improve the Targeting of Humanitarian Assistance” (working paper, National Bureau of Economic Research, Cambridge, MA, 2021).

While this technical system holds great promise and can systematically reduce errors of exclusion,2Aiken et al., “Machine Learning and Mobile Phone Data.” it also raises important societal and ethical concerns. For instance, how might this technical system impact privacy, economic justice, or how money is distributed between and among people? Even when the data are carefully protected, whose privacy might be at risk and in what ways? These questions require thoughtful attention, integrating the technical system with Togo’s particular social, cultural, and political contexts. In this ongoing research, we develop and evaluate a new method—social impact scenarios—to act as a bridge for communication between impacted communities and local policymakers, as well as system designers, ML practitioners, and policymakers in other countries.

To develop the first social impact scenario, we conducted five pilot interviews with people in Togo to understand (a) their concerns about data privacy and (b) their understanding of the criteria used to determine program eligibility: in other words, their mental model of a program that uses algorithmic decision-making to allocate resources. Pilot interviews were semi-structured one-on-one interviews that lasted between one and three hours. The semi-structured format allowed us to explore key themes while remaining open to novel topics initiated by participants. In recent months, we have conducted an additional 15 interviews using the same format with rural people in Togo.

Preliminary findings: Technical design versus on-the-ground practices

Drawing on pilot interviews, we developed the first social impact scenario that explores how a single feature of the technical system can lead to several unanticipated social dynamics. Specifically, the technical system uses mobile phone metadata (such as the mobile money balance, the towns visited by the subscriber, etc.) to identify individuals most in need of cash aid. Embedded in this feature is the assumption that each person owns their own mobile phone. Yet, mobile phone sharing emerged as a common practice in our data and is supported by prior research.3→Jenna Burrell, “Evaluating Shared Access: Social Equality and the Circulation of Mobile Phones in Rural Uganda,” Journal of Computer-Mediated Communication 15, no. 2 (January 2010): 230–50.
→Syed Ishtiaque Ahmed, Nusrat Jahan Mim, and Steven J. Jackson, “Residual Mobilities: Infrastructural Displacement and Post-Colonial Computing in Bangladesh,” in Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems (Seoul: Republic of Korea, 2015), 437–46.
→Emily L. Aiken, Viraj Thakur, and Joshua E. Blumenstock, “Phone Sharing and Cash Transfers in Togo: Quantitative Evidence from Mobile Phone Data,” arXiv (October 2021). This observation underpins three unanticipated dynamics raised by participants in pilot interviews.

Unanticipated dynamic 1: Individual inferences are not always individual
Participants often noted that because phone sharing is common, data from a single mobile phone’s mobile money account may not accurately represent how much money belongs to an individual. We asked participants if it would be “alright” or “not alright” to use the amount of money in their mobile money account to determine their eligibility for humanitarian aid. Several participants said this was “not alright” because it would lead to inaccurate conclusions about how much money belongs to them. For example, one participant described temporarily storing money on her mobile money account for an elder who did not have a phone. In that moment, the total amount of money in the participant’s mobile money account did not belong to her alone.

Unanticipated dynamic 2: Taking the long view
Beyond individual eligibility and accuracy, one participant raised community-level concerns that might arise over time if mobile money is used to determine aid eligibility. She explained that people in her community often pool money into a single mobile money account to help ill community members pay for healthcare and related expenses. She said that if her mobile money account balance was used to determine aid eligibility, she might think twice—or perhaps refuse altogether—before supporting the sick member of her community, for fear that contributing to such an account would make her ineligible for aid. This unanticipated dynamic points to the need for technologists and policymakers to consider how technical design decisions may impact people who do not directly interact with a system and that those impacts may not be immediate but rather evolve over time.

Unanticipated dynamic 3: Mental models and ownership
When humanitarian aid is targeted using individual mobile phone data in contexts where mobile phone sharing is common, beneficiaries’ mental models about eligibility underpin their understanding about which user of a shared mobile phone is the intended beneficiary. In the Togolese context, the technical system requires beneficiaries to have both a voter ID and a SIM card. In pilot interviews, however, we found examples where a younger person (with a SIM card but no voter ID) partnered with an elder (with a voter ID but no SIM card) to receive a payment. While both parties assumed that the money was intended for the elder—based on a mental model that humanitarian aid eligibility was tied to a person’s voter ID—this meant the younger person effectively missed out on the opportunity to receive aid.

Conclusion

In this ongoing research, we are developing and evaluating a new method to serve as a bridge for communication between impacted communities and policymakers, system designers, and ML practitioners. We are also exploring how a similar approach might work in an entirely different context: a San Francisco Bay Area county implementing a data-driven approach to allocating housing and services to people experiencing homelessness. We hope that social impact scenarios will be taken up by other technologists, practitioners, and policymakers who seek to proactively identify unanticipated dynamics that might arise when the technical systems they design are integrated into the complex, messy, and nuanced social world. We hope it will be useful to people working with governments to allocate resources, as well as more generally in product development processes.

Please reach out to the authors if you are interested in learning more or applying the method to your own context.