AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018 Mechanism Design for Social Good Rediet Abebe (Cornell University; [email protected]) Kira Goldner (University of Washington; [email protected]) DOI: 10.1145/3284751.3284761 Introduction fecting many individuals’ livelihoods. This research area falls at the interface of artifi- Across various domains—such as health, ed- cial intelligence, theoretical computer science, ucation, and housing—improving societal wel- and the social sciences. Since the fall of fare involves allocating resources, setting poli- 2016, the authors of this piece have been co- cies, targeting interventions, and regulating organizing the Mechanism Design for Social activities. These solutions have an immense Good research group, workshop series, and impact on the day-to-day lives of individuals, colloquium series Abebe and Goldner[2016, whether in the form of access to quality health- 2018]. The group comprises a large net- care, labor market outcomes, or how votes are work of researchers from various disciplines, accounted for in a democratic society. Prob- including computer science, economics, soci- lems that can have an outsized impact on ology, operations research, and public policy. individuals whose opportunities have histori- Members of the group partner with domain ex- cally been limited often pose conceptual and perts in non-government organizations, think technical challenges, requiring insights from tanks, companies, and other entities with a many disciplines. Conversely, the lack of inter- shared mission. The mission is to explore new disciplinary approach can leave these urgent frontiers, garner interest in directions in which needs unaddressed and can even exacerbate algorithmic and mechanism design insights underlying socioeconomic inequalities. have been under-utilized but have the po- To realize the opportunities in these domains, tential to inform innovative interventions, and we need to correctly set objectives and reason highlight exemplary work. about human behavior and actions. Doing so In this piece, we discuss three exciting re- requires a deep grounding in the field of in- search avenues within MD4SG. For each of terest and collaboration with domain experts these, we showcase ongoing work, underline who understand the societal implications and new directions, and discuss potential for im- feasibility of proposed solutions. These in- plementing existing work in practice. sights can play an instrumental role in propos- ing algorithmically-informed policies. In many cases, the input data for our algorithms may Access to Opportunity in the be generated by strategic and self-interested Developing World individuals who have a stake in the outcome of the algorithm. To get around this issue, we New technologies and data sources are fre- can deploy techniques from mechanism de- quently leveraged to understand, evaluate, sign, which uses game theory to align incen- and address societal concerns across the tives or analyze the strategic behavior of indi- world. In many developing nations, however, viduals who interact with the algorithms. there is a lack of information regarding un- derlying matters—whether that be the preva- The Mechanism Design for Social Good lence of diseases or accurate measurements (MD4SG) research agenda is to address prob- of economic welfare and poverty—due to the lems for which insights from algorithms, op- unavailability of high-quality, comprehensive, timization, and mechanism design have the and reliable dataUN[2014]. This limits the im- potential to improve access to opportunity. plementation of effective policies and interven- These include allocating affordable housing tions. An emerging solution, which has been services, designing efficient health insurance successfully demonstrated by the Information markets, setting subsidies to alleviate eco- Communication Technology for Development nomic inequality, and several other issues af- (ICT4D) research community, has been to Copyright c 2018 by the author(s). take advantage of high phone and Internet 27 AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018 penetration rates across developing nations to been realized through this system. design new technologies which enable collec- tion and sharing of high-quality data. There Availability of new technologies also presents has also been recent work from within the AI opportunities to tackle fundamental problems community to use new data sources to close related to poverty. Advances in last-mile pay- this information gap Abebe et al.[2018], Jean ment technologies, for example, enable large- et al.[2016]. Such AI-driven approaches sur- scale, secure cash transfers. GiveDirectly face new algorithmic, modeling, and mecha- leverages this and the popularity of mobile nism design questions to improve the lives of money across the world to create a system many under-served individuals. where donors can directly transfer cash to re- cipientsGD, Blattman and Niehaus[2014]. A prominent example is in agriculture, which GiveDirectly moves the decision about how to accounts for a large portion of the econ- use aid from policy-makers to recipients, giv- omy in many developing nations. Here, vi- ing recipients maximum flexibility. Such aid ral disease attacks on crops is a leading generates heterogeneity in outcomes—e.g., cause of food insecurity and poverty. Tradi- families may use aid to start a business, pay tional disease surveillance methods fail to pro- rent, cover health costs, and so on. Policy- vide adequate information to curtail the im- makers used to prioritizing specific outcomes pact of diseases Mwebaze and Biehl[2016], may be uncomfortable by such a model. A re- Mwebaze and Owomugisha[2016], Quinn search question then is: can we predict how a et al.[2011]. The Cassava Adhoc Surveil- given population will use aid? Likewise, how lance Project from Makerere University imple- can we target people for whom the interven- ments crowd-sourcing surveillance using pic- tions will make the largest difference? Aid has tures taken by mobile phones in order to ad- historically been targeted on the basis of find- dress this gap Mutembesa et al.[2018]. The ing the most deprived people. The ability to tool is set up as a game between farmers and model heterogeneous treatment effects opens other collaborators, and aims to collect truth- the door for designing more nuanced mecha- ful, high-value data (e.g., data from hard-to- nisms that fairly and efficiently allocate subsi- reach locations). This approach underlines dies in order to maximize a desired outcome. interesting challenges, such as how to opti- mally incentivize individuals to collect high- Problems in the developing world surface quality information and how to augment this unique challenges at the intersection of AI, information with existing methods. Similar ICT4D, and development economics. So- issues arise in other domains—e.g., in citi- lutions often have to be implemented in zen science and in computational sustainabil- resource-constrained environments (e.g., over ity Xue et al.[2016a,b]. Finding solutions in feature phones or with low network connectiv- the context of the developing world may there- ity) Brunette et al.[2013], Patel et al.[2010]. fore have a broader global impact. Key populations of interest (e.g., women, peo- ple living in rural parts, individuals with dis- Lack of information also leads to inefficien- abilities) may not be easily accessible Sultana cies in existing systems, presenting a possi- et al.[2018], Vashistha et al.[2015b,a]. In- bility to introduce solutions that abide by exist- dividuals may have low-literacy Sambasivan ing cultural and technological constraints. For et al.[2010]. Lack of understanding of socio- instance, large price discrepancies and major cultural norms and politics, furthermore, may arbitrage opportunities present in markets for inhibit proposed interventions Vashistha et al. agricultural products in Uganda suggest large [2018]. All of these highlight the need for market inefficiencies Ssekibuule et al.[2013]. a multi-stakeholder approach that leverages To alleviate this, Newman et al.[2018] intro- technological advances, innovative technical duce Kudu—a mobile technology that func- solutions, and partnerships with individuals tions over feature phones via SMS service. and organizations that will be impacted by the Kudu facilitates transactions between farmers solutions. MD4SG fosters one such environ- in rural areas and buyers at markets in cities ment in which insights from across these dis- by allowing sellers and buyers to post their ciplines inform the design of algorithms and asks and offers. Kudu has been adapted by mechanisms to improve the lives of individu- users across Uganda and many trades have als across the world. 28 AI MATTERS, VOLUME 4, ISSUE 3 4(3) 2018 Labor, Platforms, and Discrimination The aim is to choose hiring policies that will mitigate discrimination against protected cat- Online platforms are ubiquitous, providing a egories. Hu and Chen[2017] highlight addi- vast playground for algorithm design and arti- tional complexity that arises in dynamic set- ficial intelligence. Every policy decision, how- tings where workers are hired based on invest- ever, impacts and interacts with the platform’s ment decisions (e.g. college GPA) in an ini- strategic users. In this section, we will focus tial temporary labor market (e.g. internships) on online labor markets and how discrimina- and this job creates a worker’s initial produc- tion effects stem from a platform’s decisions.