Byproducts of Algorithmic Systems: Fairness, Usability, and Transparency Challenges

Algorithmic systems are pervasive in daily life, automatically recommending content to show on social media feeds as well as deciding mortgage approvals, in part due their efficiency gains and their purported objectivity. Yet, deployments of algorithmic systems have repeatedly faced scrutiny from the broader public and academics when found to be discriminatory, creepy, or illegal. The recurrence of these events, in spite of efforts to address these issues, suggests invariant attributes underlying the production and design of algorithmic systems. Specifically, this dissertation identifies origins for these issues within the shared norms of data science practitioners, then demonstrates how those norms are expressed through interface and algorithmic design. First, I conduct interviews with 20 data scientists to understand their perceptions of fairness interventions that modify data and what barriers exist to using such methods in practice. From these interviews, I describe a set of normative values which define the priorities and logics that prevent consideration of alternative fairness interventions and ultimately contribute to the maintenance of the status quo. Next, I investigate how those values and priorities present in user interactions with targeted advertising. Specifically, I characterize the types of information and controls that 22 large online platforms provide to users and conduct a user study with 198 participants and eight of those platforms to evaluate the usability and usefulness of the provided systems. Across all platforms I find vague controls and oversimplified information. The consequence for users is that many of the questions they hoped the system would answer, remained unanswered after exploration. Finally, I explore how user intention and design choices influence search and watch behavior on TikTok through a data donation user study with 86 participants. In particular, focusing on help-seeking and algorithmically prompted searches. In spite of TikTok's lack of transparency regarding prompted searches, I determine that 41\% of participants' searches were algorithmically prompted and find they are associated with longer search sessions participants described as ``rabbit holes.'' Taken together, I show that byproducts first originate as normative practices and values during the production of algorithmic systems, later metamorphizing for users as fairness, usability, and transparency challenges through system design. The tangible effect of these byproducts is a prevention of understanding, limitations of user agency, and manipulation of online behavior in ways that most users may not fully understand or expect.