Algoritmo Auditável: o Papel dos Sistemas de Recomendação na Criação de Filtros Bolha de Desinformação sobre as Urnas Eletrônicas no YouTube

Abstract

Recommender Systems are a tool conceived out of the need of effective navigation through the vast amount of data available online. Its current implementation, abusing of user’s personal data and inherent subjectivities for better recommendations has been the subject of studies and criticism on several fronts. With its proven influence on popular decision-making processes such as USA’s 2016 Elections and the Brexit referendum, these systems end up reinforcing pre-established views of the world, allowing the creation of filter bubbles through which users consume information and shape their opinions with the influence of a confirmation bias. In this sense, users that consume minsiformative contents end up being directioned to the consumption of even more misinformative contents. When we consider brazilian Electronic Ballots – one of the topics with the greatest information fluxes over the last years and target of constant misinformative campaigns – the creation of misinformation filter bubbles poses an even greater threat to democracy. Considering that YouTube is the second most used social media in Brazil, with a total of 142 million users (almost 70% of the country’s population, according to the 2022 Census), this study sought to audit how the platform’s Recommender Systems contribute to the creation of misinformation filter bubbles, both in search results and recommended videos, simulating diverse content consumption profiles related to Electronic Ballots. Through the author’s developed methodology, it was possible to confirm the widespread presence of misinformation in popular videos on the platform regarding this topic, as well as a general increase in recommended misinformation across all profiles. For the misinformation consumption profile, the creation of a misinformation filter bubble was proven, with the Recommender System significantly reducing the recommended informative videos and substantially increasing the recommended misinformative videos — to the point where one out of every three recommended videos presents potential misinformation.