Auditing Sentiment Filtering and Curation in Bluesky’s Feed Algorithm: Reinforcement and Readability

Abstract

Over the past decade, the issue of social media addiction has been rising to the forefront of public concern. Much of this concern has targeted feed algorithms: the decision-making forces that govern what we see on social media. Many people speculate that social media algorithms intentionally show users more negative content meant to spark outrage. However, scholarship is divided on the explicit role that feed algorithms play in the proliferation of negative/outrage content on social media platforms. While prior scholarship has examined the algorithmic proliferation of political “feed bubbles”—mechanisms that reinforce and radicalize beliefs—there has been limited research on the proliferation of sentiment-based feed bubbles. Bluesky, a Twitter/X alternative positioning itself as a decentralized and bias-free platform, emerged in 2024. Because of its smaller size, little research has been conducted on Bluesky’s default feed algorithm. This thesis studies how continued engagement on Bluesky affects feed sentiment, assessing whether a bot with predictable behavior affects changes in the sentiment on its feed. Furthermore, the study examines the relationship between a post’s sentiment and its readability as a way of understanding possible human biases towards sentiment. While results show very little sentiment drift and no engagement correlation on Bluesky over time, the study does find that sentiment is statistically correlated with post readability. Positive and negative sentiment posts are easier to read, while neutral posts are more difficult to read by approximately one standard deviation from the mean. Ultimately, this study provides supporting evidence for the hypothesis that Bluesky does not meaningfully consider sentiment in the curation of its feed algorithm. I hypothesize that instead of the algorithm driving sentiment, natural human tendencies toward negative content create a concurrent feedback loop that may work alongside explicit algorithmic bias and sentiment reinforcement. To address societal concerns of algorithm-based addiction, I propose a policy solution aimed at addressing proliferation of negativity while remaining mindful of legal constraints.