The Limits of Machine Unlearning as a Means of Political Protest

Abstract

A Machine Unlearning algorithm seeks to remove the influence of a given data point on a trained ML system. Analyses and implementations of Machine Unlearning have focused on its use for maintaining the data subjects’ privacy, yet other purposes prompt individuals and groups to request the so-called `forgetting’ or `unlearning’ of their data by machine learning models. In this paper, we focus on uses of unlearning motivated by a desire to fight three forms of oppression promoted by machine learning models: exploitation (as experienced by creatives whose work is used without consent), powerlessness (as experienced by a workforce undergoing the effects of a technology in which they have little say), and image-based sexual abuse (as experienced by victims of nudification apps). We first delimit two types of political actors who resort to Machine Unlearning: someone who has a desire not to be personally involved in the oppression, and someone who is motivated by a collective responsibility and seeks to diminish the capabilities or use of the oppressive machine learning model. We argue that Machine Unlearning, as it exists today, actually makes it harder to express political agency through data forgetting. On the engineering side, it's because the mathematical formalization of unlearning does not necessarily eliminate the influence of the forget data on the ML system. On the socio-technical side, the petition model of Machine Unlearning is an obstacle to effective collective action. We also examine other factors supporting our conclusion, including Machine Unlearning's focus on the costs to the system's owner, hard limits on performance degradation as a result of unlearning, and the reification of the data subject as a data point, which ignores the multifaceted ways in which a person shows up in a training dataset. We conclude with an examination of possible research avenues in machine unlearning to promote data-based protest of oppressive machine learning models.