Lindsay Popowski

Complex machine learning algorithms are currently thought of as fundamentally unknowable to the general public. As these algorithms are deployed in user-facing contexts like commerce and social media, their opacity threatens human agency, leading to calls for simpler algorithms. I seek to demonstrate that the tradeoff between technical complexity and user understanding is not fixed. I hypothesize that people can understand and accurately predict algorithm behavior when it aligns with concepts they find easy to reason about, even if the underlying operation is complex. My work has focused on distilling this perspective into criteria for algorithm understandability based on high-level characteristics, rather than details of the implementation. Through experimental studies, I demonstrate that users can form and accurately predict even complex algorithms that satisfy these criteria. My work thus signals the possibility of useful and understandable algorithms, such as a social media feed algorithm that prioritizes democratic values.