Algorithmic transparency and accountability

Abstract

Computer algorithms are being deployed in ever more areas of our economic, political and social lives. The decisions these algorithms make have profound effects in sectors such as healthcare, education, employment, and banking. Their application in the anti-corruption field is also becoming increasingly evident, notably in the domain of anti-money laundering. The expansion of algorithms into public decision making processes calls for a concomitant focus on the potential challenges and pitfalls associated with the development and use of algorithms, notably the concerns around potential bias. This issue is made all the more urgent by the accumulating evidence that algorithmic systems can produce outputs that are flawed or discriminatory in nature. The two main sources of bias that can distort the accuracy of algorithms are the developers themselves and the input data with which the algorithms are provided. Equally troublingly, the analytical processes that algorithms rely on to produce their outputs are often too complex and opaque for humans to comprehend, which can make it extremely difficult to detect erroneous outputs. The Association for Computing Machinery (2017) points to three potential causes of opacity in algorithmic decision making processes. First, there are technical factors that can mean that the algorithm’s outcomes may not lend themselves to human explanation, a problem particularly acute in machine-learning systems that can resemble a “black box.” Second, economic factors such as commercial secrets and other costs associated with disclosing information can inhibit algorithmic transparency. Finally, socio-political challenges, such as data privacy legislation may complicate efforts to disclose information, particularly with regards to the training data used. This paper considers these challenges to algorithmic transparency, and seeks to shed some light on what could constitute meaningful transparency in these circumstances, as well as how this can be used to leverage effective accountability in realm of algorithmic decision-making.