The Algorithmic Flattening of Sound: Computational Evidence and Justice Implications of AI Music Homogenization

AI summary

Overview

Research area: AI safety and ethics; computational music information retrieval (MIR); algorithmic fairness and cultural justice.

Technical level: Intermediate. The paper's argument is accessible to a general reader, but the audit relies on a 72-feature MIR feature space, dispersion and separability diagnostics, mixed-effects models, and classifier cross-validation.

Scope: A black-box audit of two commercially deployed text-to-music systems (Suno and Lyria 3) across four genres, testing whether AI-generated tracks exhibit measurable acoustic homogenization relative to matched human corpora, paired with a justice-centered account of why the pattern matters.

What This Paper Is About

Generative music systems can now produce complete tracks from short text prompts, and existing scholarship has mainly debated copyright, labor displacement, and authorship. This paper asks a different question: whether these systems systematically flatten musical variation at the point of production, rather than merely filtering an existing repertoire downstream. The authors audit Suno and Lyria 3 against human baselines across Afrobeats, K-pop, Dance Pop, and Heavy Metal, and argue that the measured acoustic patterns are a justice-relevant condition affecting which musical styles become legible, valued, and economically rewarded.

Key Contributions

1. The authors develop what they describe as the first AI audit approach for testing within-genre musical homogenization across commercially deployed text-to-music systems.
2. They provide empirical evidence that Suno and Lyria 3 exhibit structurally distinct acoustic tendencies across four genres.
3. They introduce a null-prompt baseline condition (genre name only, no additional instructions) that isolates system-level learned priors from prompt engineering effects.
4. They develop a justice-centered analysis of how production-level homogenization can reshape cultural authority, economic value, and epistemic legitimacy in music.

Main Findings

• Prompt fidelity is low for both systems. Suno shows near-zero prompt fidelity (max |r| = 0.26); tempo and onset density reach only |r| = 0.15 and |r| = 0.08 respectively. Lyria tracks tempo moderately (r = 0.30–0.57) with limited fidelity for harmonic ratio and self-similarity. The authors conclude the homogenization patterns reflect learned priors rather than their prompt design.

• Lyria homogenizes within genre. In three of four genres, Lyria outputs cluster more tightly than human reference tracks. The aggregate variance ratio (AI/Human) is 0.839, meaning AI output variance averages 84% of human variance, with 58% of feature × genre cells showing AI variance below human. A mixed-effects model confirms the reduction (Cohen's d = −0.265, p < 0.0001, BH-corrected). Mean pairwise distances fall below the human baseline in three of four genres (range 0.773–1.119).

• Suno disperses within genre but collapses genre boundaries. Suno shows higher within-genre spread than human tracks (variance ratio 1.667; 83% of feature × genre cells show AI > human; Cohen's d = +0.646), with normalized pairwise distances exceeding 1.0 in three of four genres (range 1.004–1.582), and Afrobeats showing the largest deviation (1.58×).

• Lyria preserves genre boundaries; Suno does not. Human tracks maintain a genre separation ratio of 0.662. Lyria is nearly identical at 0.676 (+1%), while Suno falls to 0.429 (−36% relative to human), indicating that its tracks bleed across genre boundaries.

• The two systems do not converge on a common sound. All convergence ratios exceed 1.0, ranging from 2.70 (Heavy Metal) to 4.64 (Afrobeats), with Dance Pop at 4.38 and K-pop at 4.47. The two systems are more acoustically distant from each other than two random human subsets would typically be, across every genre.

• AI and human outputs are near-perfectly discriminable. A classifier on the 72-dimensional MIR feature set, evaluated with 5-fold stratified cross-validation, achieves mean AUC = 0.991 ± 0.003 and accuracy = 0.967 ± 0.021. The three highest-importance features are MFCC Δ² mean (0.144), MFCC 0 (0.117), and IOI mean (0.051).

• Discriminability is not driven by vocals. Restricting the classifier to Afrobeats, the only genre where human tracks are also instrumental, AUC remained 0.980 ± 0.011, with vocal-sensitive MFCC features dropping from 64.5% to 38.0% of top-10 importance, confirming rhythmic regularity as the driver.

• Null-prompt results mirror the steered condition. For Suno, null-prompt and MIR-steered ratios are nearly identical (differences of −0.302 to +0.094). For Lyria, null-prompt tracks are more homogeneous than MIR-steered tracks in three of four genres (differences −0.178 to −0.048). Null-prompt Lyria (0.595–1.071) remains more homogeneous than null-prompt Suno (0.981–1.366).

• Suno's generated titles collapse to a narrow prior. Across 100 null-prompt tracks per genre, only 13–16% of titles are unique within each genre, and a single title can account for up to 20% of all tracks in a genre. "Palm Wine Drift" was assigned to 20 of 100 Afrobeats tracks; "Anvil Breaker" was the top Heavy Metal title (18 occurrences) and also appeared 14 times in Dance Pop. Lyria 3 does not produce titles.

Methodology in Plain English

The authors treat Suno (Pro tier; v5.5) and Lyria 3 as black boxes with no access to training data, architecture, or objectives, collecting all generations in April–May 2026 with platform defaults held constant and no post-processing beyond excluding silent or corrupted outputs.

Four genres were chosen for contrastive coverage along two dimensions—geographic origin (Western / non-Western) and training-data representation (high / low): Afrobeats (non-Western, low), K-pop (non-Western, low), Dance Pop (Western, high), and Heavy Metal (Western, high).

Human baselines consist of 100 tracks per genre, drawn from openly available Spotify playlists filtered to release years 2000–2026. From a candidate pool of 400 tracks per genre, the authors deduplicated by capping at five tracks per artist and removing duplicate titles, then extracted MIR features and applied k-means clustering (k=10), sampling proportionally across clusters to reach 100 tracks per genre. Because Spotify preview clips were unavailable at sampling time, corresponding 30-second audio was retrieved via Deezer.

Two experiments were run. In Experiment 1, prompts were algorithmically derived from the MIR features of human tracks, following each platform's documented prompt guidance, so each AI track had an explicit human counterpart (n = 100 per genre per system; 800 total). In Experiment 2, tracks were generated from genre-name-only prompts with a no-vocals instruction, again n = 100 per genre per system (800 total), using the same human corpora as baselines.

All audio was analyzed as 30-second excerpts from the middle of the track using a single librosa pipeline applied identically to AI and human excerpts, with features z-scored over the combined AI+human set within each experiment, and scaling performed within cross-validation folds for classification. Homogenization was measured with five diagnostics (D1–D5) covering global dispersion with a one-sided Mann–Whitney U test and a 10,000-iteration label-shuffle permutation test, feature-level variance, distributional concentration, feature-space region size under dimensionality reduction, and system-to-system convergence and AI-vs-human discriminability.

The authors note interpretive limits: findings do not estimate population prevalence across all generative music models, genres, or time periods, nor identify internal causes of reduced variation; human baselines are constrained by 30-second preview availability; and standard MIR features may underrepresent musically salient but hard-to-formalize distinctions such as groove feel and expressive microtiming.

Why This Matters

Impact on research. The paper reframes musical homogenization as a property of production rather than only of distribution. Prior work on radio consolidation, the "loudness war," and recommender-driven concentration described homogenization as operating through selection and circulation; this work argues generative systems relocate it upstream, into creation itself. It also supplies a methodological template—paired MIR-derived prompting plus a null-prompt baseline—for auditing commercial generative systems as deployed rather than as documented.

Real-world applications:

• Platform content governance. The finding that generated audio is not merely entering streaming systems but may be especially easy for them to sort, recommend, and amplify has direct implications for playlist assembly and genre labeling.
• Genre definition and metadata practice. Suno's cross-genre title leakage ("Anvil Breaker" in both Heavy Metal and Dance Pop) and its collapsed genre separation suggest metadata conventions built on genre fidelity may not hold for generated catalogs.
• Cultural heritage and representation work. With Afrobeats—the genre the authors selected for its internal diversity and its underrepresentation in training data (roughly 0.3% African music in reported corpora)—showing the largest deviation from the human acoustic center under Suno, the paper points to targeted monitoring for low-resource musical traditions.
• Artist and labor advocacy. The redistribution analysis frames near-zero marginal-cost substitutes as a shift in bargaining power and licensing incentives, especially for musicians outside mainstream visibility.

Industry relevance. The paper cites Deezer reporting that 28% of new uploads in 2025 were AI-generated, and Suno publicly reporting approximately 12M users in 2024. Combined with reported training-data imbalance—approximately 94% Western traditions, African music around 0.3%, the Middle East 0.4%, South Asia 0.9%, and over 93% of surveyed researchers focusing on Global North music—the authors argue that statistical dominance in training data can become difficult to disentangle from normative centrality, with low-frequency musical practices less likely to be learned with stability.

Future Directions

• Testing whether the observed patterns generalize beyond two systems, four genres, and the April–May 2026 collection window, since the authors explicitly state their findings do not estimate population prevalence.
• Investigating internal causes of reduced variation, which the black-box design cannot identify, by connecting model architecture, training distribution, and optimization objectives to output-level acoustic contraction.
• Developing perceptual and community-grounded evaluations to complement MIR features, which may underrepresent musically salient distinctions such as groove feel, swing, regional inflection, and microtiming.
• Studying the hypothesized platform feedback loop empirically—whether genre reference distributions shift as AI outputs grow as a share of what recommendation systems surface—and whether repeated generation narrows the space of prompts users write.

Target Audience

Researchers in AI ethics, fairness, and accountability; MIR and music technology researchers; cultural policy scholars and cultural economists; platform governance and content moderation teams; and musicians, advocacy organizations, and genre communities concerned with how their musical traditions are represented, defined, and compensated as generated music scales.

Authors’ abstract

This paper audits whether large-scale generative music systems exhibit measurable musical homogenization relative to human-produced music, and develops a justice-centered account of why this matters. We audit two commercially deployed systems (Suno and Lyria 3) across four genres (Afrobeats, K-pop, Dance Pop, and Heavy Metal). For each system and genre, we generate 100 tracks and compare them against human corpora of equal size, using 72 music information retrieval (MIR) features and multiple diagnostics of dispersion, redundancy, and separability. We define homogenization as reduced acoustic variation in standard computational audio features including rhythm and timing, timbre/spectral shape, and dynamics, both within genres and across genre boundaries. We also generate tracks using only a genre name as the prompt, with no additional instructions, to reveal each system's default musical tendencies. The results show two structurally distinct homogenizing tendencies. Lyria reduces within-genre acoustic diversity, while Suno collapses the acoustic distinctions between genres without compressing within-genre spread. Neither system follows user prompts faithfully, indicating that the observed patterns reflect learned priors rather than prompt constraints. The two systems do not converge on a common acoustic profile and are more acoustically distant from each other than two random human subsamples would typically be. Nevertheless, a standard classifier distinguishes AI from human tracks near-perfectly on MIR features alone. We argue that these patterns matter not as an aesthetic curiosity but as a justice-relevant condition, shaping which musical styles become legible, valued, and economically rewarded as generated outputs increasingly circulate at scale.