Auditoria de sistemas de inteligência artificial em plataformas de Streaming: um estudo experimental sobre vieses e recomendação musical no Spotify

Abstract

Algorithmic recommender systems have become the main mediators between musical works and listeners, operating as black boxes whose selection criteria remain opaque. This study audits Spotify’s recommendation system from an Information Management perspec tive, investigating whether automated curation fosters cultural diversity or the homoge nization of musical taste. A black-box, sock-puppet auditing methodology was adopted, using four synthetic personas of contrasting archetypes, mainstream, functional/lo-fi, nos talgic, and niche, comparing the declared stimuli (input) with the recommendations gen erated in the Daily Mixes (output). Given the progressive removal of fields from the Spotify Web API during the research, treated here as meta-evidence of the platform’s opacity, the data were enriched with the external sources Last.fm and MusicBrainz. Di versity and concentration were measured through Shannon entropy, Pielou’s evenness, the Gini coefficient, the Herfindahl-Hirschman Index, and the Jaccard index, with inferential statistical treatment (bootstrap confidence intervals, the Mann-Whitney test, rarefaction, and a permutation test). The results reveal a three-level phenomenon: at the content level, the personas’ artist repertoires remain entirely disjoint (Jaccard = 0, more segre gated than chance, p < 0.001); at the thematic level, there is partial genre convergence; and at the magnitude level, diversity converges through catalog richness expansion rather than entropy homogenization. A popularity bias (Daniel, +131% listeners per artist) and a hit bias within the long tail (Sofia, +405% listeners per track) are also confirmed. We conclude that, for the four synthetic personas and within the collection window analysed, the algorithm did not homogenize recommendations across profiles: it expands the inter nal richness of each profile and brings them thematically closer, while preserving distinct content silos.