b'ViewpointsSara Kassir is a recent Master of Public Policy graduate of the Harvard Kennedy School.for evaluating each aspect of a model for undue sources ofsocially constructed system of norms, values, beliefs, and influence. Further, because algorithmic audits encouragedefinitions is crucial to an entity being able to function systematic engagement with the issue of bias throughoutwithin society. 3the model-building process, they can also facilitate an organizations broader shift toward socially responsible dataThe degree of assurance that can actually be achieved collection and use. through an audit obviously varies depending on the industry. For example, in recent years, the field of safety engineering What is an algorithmic audit? has consciously attempted to signal that no audit can ever Algorithmic auditing is an effort to ensure that the contextdeem something like an airplane 100 percent safe. As one and purpose surrounding machine learning applicationsreport from the UC Berkeley School of Information points directly inform evaluations of their utility and fairness.out, In contrast to the ship it and fix it later ethos that Stephen Hawking wrote about the limitations of abstractionhas defined the tech industry, safety engineering requires in his A Brief History of Time: The usual approach of sciencethat the developer define what must be avoided (e.g., of constructing a mathematical model cannot answer theairplane crashes, patient death) and engineer backwards questions of why there should be a universe for the modelfrom there. 4Machine learning applications, with their to describe. Why does the universe go to all the bother ofdiverse consequences and potential for bias to emerge, are existing? 1Admittedly, the esteemed theoretical physicistsimilarly impossible to ever deem 100 percent risk-free, was not writing about machine learning applications in theand this spirit of imperfect assurance should inform how public sector, but his message on intentionality in analysisthey are tested. In particular, three tenets of general auditing is nonetheless salient. Data, models, algorithms, and othertheory map well to the complexity of auditing algorithms means of simplifying the world cannot be separated from thespecifically. These are: (1) the notion that an auditor must context in which they are produced. Through audits, machineexercise judgement to explore the relevant details of a case; learning tools are examined with the appropriate frame of(2) the need to assess the inner-workings of process, rather reference in mind. than only examining its outputs; and (3) the expectation that an organization, subject to auditing endeavors, document its What principles from auditing can be translated toactivities for the purposes of evaluation.machine learning?With the professional practice dating back to the IndustrialPrinciple 1: Marrying structureRevolution, an audit is defined as a formal examinationand judgmentof an organizations accounts, initially with the intentThe first auditing principle that is relevant of protecting a firms investors from fraud. 2Over the pastfor machine learning processes relates to the century, these examinations have diversified to encompasssteps an auditor is expected to follow in completing goals much broader than identifying financial risk. Today,their examination. Auditing, like any profession, is subject auditors may examine an organization in terms of itsto ongoing debates about best practices. Scholar Michael regulatory compliance, process efficiency, environmentalPower, who explores the field as a principle of social impacts, or ethical standards. But regardless of the preciseorganization, describes one of the industrys greatest tensions focus, the procedure is directed toward the establishmentas the structure-judgment problem, or the notion that a of legitimacy. According to sociologist Mark Suchman, thistradeoff exists between auditing procedures that rely on generalized perception or assumption that the actions ofprescribed techniques and those that give greater weight an entity are desirable, proper, or appropriate within someto individual judgement. Power uses the metaphors of WINTER 2019 / 2020 IBM Center for The Business of Government 89'