Data-Driven Analysis of Form and Harmonic Structure in AI-Generated Popular Music

We present a large-scale analysis of the harmonic and formal structures of music generated by users of two AI music generation platforms, Suno and Udio, compared with human-composed popular music. We address three questions: 1) What considerations arise when applying music information retrieval (MIR) tools to AI-generated music? 2) How does AI-generated music's harmony relate to human-composed popular music? 3) What structural similarities or differences exist? We analyze 20,002 Suno tracks, 20,002 Udio tracks, and 20,010 human-composed recordings. We perform harmonic analysis using a state-of-the-art chord estimation model and structural analysis using a joint beat-tracking and semantic section-labeling model. We model harmonic progression and musical form using trigram and tetragram analysis of detected chords and labeled sections, respectively, and thereby quantify their diversity in our three music collections. Our analysis of chord progressions reveals that the music in our Suno collection exhibits a stronger concentration on Western pop harmonic patterns, such as I-V-vi-IV, than the music in either our Udio or human collection. Our analysis of section labels shows that the music in our Suno collection tends toward simple verse-chorus repetitions, whereas that of Udio exhibits a wider structural variety. These findings suggest that the music generated by Suno (up to version 3.5) sits in a narrower harmonic and formal space than that of Udio and human-composed music. Fully automated, open-source MIR tools now support corpus-scale harmonic and structural analysis; chord estimation remains an open challenge, particularly for inversions, extensions, and substituted dominants, yet it supports the analysis of harmonic patterns across the three collections. We release our beat-level chord annotations, metadata, section boundaries, self-similarity matrices, audio links, and code under an MIT license. Dataset The dataset comprises 60,014 tracks across three collections: Collection Tracks Avg Duration Total Duration Lastfm (human) 20,010 261.3 ± 182.1 s 1,452 h Suno (AI) 20,002 175.2 ± 54.6 s 1,085 h Udio (AI) 20,002 143.2 ± 98.6 s 802 h