Unpacking Algorithmic Bias in YouTube Shorts by Analyzing Thumbnails

Abstract

As digital platforms increasingly shape our online experiences, the influence of recommendation algorithms on user behavior becomes ever more significant. This research delves into the biases inherent in YouTube Shorts' recommendation algorithms by analyzing the topical content of thumbnails through captions generated by advanced generative AI models, specifically GPT and Llama. Employing topic modeling and clustering techniques, we scrutinized a substantial dataset of YouTube Shorts to uncover patterns of bias within the recommendation process. Our findings reveal a significant drift in recommended content from serious geopolitical topics to broader, entertainment-focused themes, underscoring the impact of algorithmic preferences on user engagement. This study highlights the necessity for greater transparency and fairness in content recommendation systems, offering valuable insights into the ethical implications of algorithmic bias in digital media.