Solving Social Problems with AI is Possible – But We Need Better Data Infrastructure First
August 13, 2018 | By Rachel Wilder, Program Assistant
Artificial intelligence for social good isn’t just hype. AI allows computer systems to perform tasks, like visual perception and decision making, that previously required human intelligence. Public and nonprofit sector leaders have an opportunity to increase their impact by applying AI to resource optimization and prediction problems outside the bounds of older methods: researchers and local government partners have used AI to better identify police officers at risk of adverse events like racial profiling or deadly use of force and to improve HIV awareness and testing rates among homeless youth. AI holds promise as a tool to approach a variety of social problems.
But in order for these benefits to be realized at scale, we need to overcome significant data infrastructure barriers. My brother, Bryan Wilder, is completing his PhD at the University of Southern California’s Center for AI in Society (CAIS). I found compelling overlap between the data needs that he and his advisor see and the Beeck Center’s work on data for social good. Three that stood out to me:
1. We need more high-quality data.
Despite the rapid expansion of public and private sector data collection, there often just isn’t enough data on the issues and people that AI for good can benefit most. For example, CAIS identifies many issues in public health (including outreach, disease tracking, and treatment decisions) that AI is well-positioned to address. However, public health data is especially scarce in the low-resource and developing country contexts where disease prevention could have the biggest impact. And as Gideon Rosenblatt and Abhishek Gupta recently commented in the Stanford Social Innovation Review, it isn’t enough that data is collected; datasets must be complete, accurate, and structured in order for machine learning systems to be developed.
2. We need to streamline data sharing across sectors.
Computer science researchers in academia have the energy and resources to apply AI to social problems, but they need access to data in order to do so. Even within interested social impact partner organizations, in-house data use restrictions can make the process of sharing with researchers prohibitively difficult.
A Beeck Center report published last year, “Accelerating the Sharing of Data Across Sectors to Advance the Common Good,” outlines a framework for governments and private companies to share data through a trusted intermediary with sensitivity to privacy and ethics concerns. This idea is echoed in discussion on the development of a “Data Commons” that would serve as a unified platform for data to be used in AI work. We should continue to push the conversation on getting data out of organizational silos and into the hands who can use it for good.
3. We need social sector leaders with data and technology literacy.
In order for governments and nonprofits to know that AI-driven solutions meet ethical considerations – including ensuring that racial and gender biases don’t influence results – there must be organizational leaders who can understand how algorithms arrived at recommendations or predictions.
Algorithmic bias is a serious and well-documented problem with AI, and it is especially important to uncover bias when working on social issues that affect groups already struggling with systemic bias. CAIS director Milind Tambe acknowledged that “[b]eing able to explain decisions an AI system has made to an end user is very important,” noting that “[i]n many cases, we are working with vulnerable communities and populations, and we need to ensure they will not be harmed.”
The Beeck Center and Deloitte’s Center for Government Insights have co-produced a playbook for Chief Data Officers in government that explicitly addresses this subject, giving data leaders a roadmap for understanding and managing algorithmic risks. Beeck Center researchers have also published a framework for ethical blockchain design that can serve as a template for ethical design in other technologies, including AI. This type of training for data officials in the public sector and implementers of new tech solutions will become increasingly important as AI becomes more common.
The practice of addressing social good questions with artificial intelligence is young, and it’s exciting to me to envision how AI tools could amplify the impact of public programs if they are successfully applied at scale. Enabling that future will require investing in data collection, sharing, and literacy. My colleagues at the Beeck Center are working at the heart of advocacy and education efforts to make those investments a reality – so stay tuned!