Disentangling the Components of Ethical Research in Machine Learning

Abstract

While practical applications of machine learning have been the target of considerable normative scrutiny over the past decade, there is growing concern with machine learning research as well. Debates are currently unfolding about how the research community should develop its research agendas, conduct its research, evaluate its research contributions, and handle the publication and dissemination of its findings, among other matters. At times, these debates have been quite heated, with different actors adopting different positions on what it means to do machine learning research ethically. In this paper, we show that some of the disagreement owes to a lack of clarity about what ethical issues are at stake in machine learning research, how these issues - in particular, the concerns with research integrity, research process harms, and downstream consequences - relate to (or, more often, differ from) one another. We then explore which mechanisms are most appropriate for dealing with the different types of ethical issues, and highlight which ethical issues require more attention than they are currently receiving. Ultimately, we hope to foster more productive discussions about the responsibilities that the community bears in addressing the ethical challenges tied to machine learning research and how to best fulfil these responsibilities.