A 2019 survey found that 73 of 84 prominent AI strategy documents referenced transparency or explainability [81]. Influential intergovernmental bodies such as United Nations agencies and the Organization for Economic Cooperation and Development (OECD) have put forth transparency and explainability as key mechanisms for ensuring that algorithmic systems produce beneficial outcomes and uphold "democratic values" [121, 143]. Algorithmic transparency and explainability can serve many purposes, but some of the most important are legal in nature: allowing lawmakers to understand and craft effective rules for algorithmic systems, enabling a broader set of stakeholders to be aware of (and obtain redress from) algorithmic harms, and assisting regulators in exercising meaningful oversight over the use of algorithms [81, 109]. To serve these objectives, transparency measures and explanation techniques must be developed with an understanding of the specific goals, constraints, and incentives of policymakers. This paper aims to help bridge the gap between policymakers and the explanation research community, helping researchers to better understand and respond to the needs of policymakers. To this end, it provides case studies illustrating five uses for algorithmic transparency and explanation in policy settings. These case studies (Table 1) were selected to span four axes: the spectrum from explanation to transparency (including both requirements for specific explanation techniques, like those developed by the machine learning research community, and broader forms of transparency requirements); different jurisdictions (including U.S. federal regulators, U.S. states, and the EU); policy actors with differing technical and financial capacities; and a diverse array of policy approaches (including prescriptive technical rules, process-oriented rules, nonbinding guidelines, and modifications to legal procedures). Building on these case studies, this paper argues that explanation techniques developed by the research community can be too complex, too uncertain, or too restricted to satisfy the constraints that policymakers and the law operate under in practice. As a result, explanation is often limited in its ability to enable meaningful public policy solutions to algorithmic harms.

The legal relationship between government and citizens is mediated by documents. This paper identifies four classes of cognitive assistants that could improve the experience of citizens and government officials in using and understanding government documents: self-filling forms; error-detecting forms; proactive information search; and deductive document synthesis. Each of these classes of cognitive assistants has the potential to significantly improve access to justice and delivery of information, services, and other benefits to citizens by improving the ability of citizens to understand and correctly fill out forms and to comprehend informational documents.