FAT Flow: A data science ethics framework

The impact of data science in our society is undeniable, both in generating cost-efficiencies and in o ering better services and products. As data science often involves making decisions for humans, from deciding on whether to give credit or not, to a self-driving car deciding how to drive, the manner in which companies conduct data science will have large implications for humans (including their customers) too. The interest in the ethical aspects of data science is growing, and have become an increased focus point in both research and practice. Data science ethics looks at what is right and what is wrong when conducting data science. This goes beyond what is legal, and considers aspects as privacy, discrimination against sensitive groups, the ability to explain predictions, and accountability. This paper provides a framework in which concepts, techniques and cautionary tales related to data science ethics can be placed. Companies can use the framework to think about the ethical aspects of their own data science projects, be it at the start of a project or to review current data science practices. It provides guidance on what the important concepts are, how techniques can be used to improve on their data science, and what cautionary tales exist in domains that might be similar to their own. The FAT Flow framework looks at three dimensions: (1) the role of the humans involved in the project, being data subject, data scientist, manager and model applicant; (2) the stage of the data science project: from data gathering to model deployment; and (3) the FAT evaluation criteria: Fair, Accountable and Transparent.