A Systems Approach to Mitigating Harms of Content Recommender Systems

Abstract

Content recommender systems play a critical role in the dissemination and discussion of information. Therefore, it is crucial that there are structures and processes to support the development of safe, trustworthy recommender systems. To responsibly deploy these models, Machine Learning (ML) practitioners must go out of their way to navigate a complex ecosystem of regulation, organizational structures, and resources. In this paper, I take an approach rooted in systems theory to detail a holistic overview of the content recommendation ecosystem and the challenges that arise from its structure. Through a survey and semi-structured interviews, my research highlights how these issues stem from the absence of a shared language around harms. Furthermore, practitioners lack a framework to evaluate the ethical and societal implications of their recommender systems. I detail how the current approach to identifying and mitigating harms does not adequately prevent these harms. I also identify challenges that arise from the current organizational arrangements of knowledge and expertise. Based on these findings, I recommend how we can use oversight mechanisms to establish a standard for harm mitigation approaches. I further suggest we modify the existing ecosystem to distribute responsibility and equip practitioners to confidently deploy safe content recommender systems.