Who Gets Heard? Rethinking Fairness in AI for Music Systems

AI summary

Overview

• Research area: AI safety and ethics, specifically fairness, cultural representation, and genre bias in AI systems for music (music-AI), framed as a computer-society (cs.CY) position paper presented at a workshop on AI for Music.
• Technical level: Beginner-Friendly. The paper contains no experiments, no models trained, and no quantitative benchmarks; it is an argument-and-framing paper built around stakeholder analysis, technical gap analysis, and recommendations.
• Scope (one sentence): The paper maps who the stakeholders in the music-AI ecosystem are, how representational bias harms each of them, which design choices in datasets, representations, and interfaces cause those harms, and what interventions could make music-AI systems more culturally fair.

What This Paper Is About

Existing work on risks in generative music AI has focused on copyright infringement, deepfakes, market saturation, and transparency, with mitigation efforts centered on AI-generated music detection, deepfake detection, and audio watermarking. This paper argues that a different risk is underexamined: cultural and genre bias in music-AI systems, which skews representation heavily toward certain regions and genres and away from Global South traditions (South Asia, South East Asia, Oceania, the Middle East) and genres such as country and folk music. The goal is to trace how that bias plays out for each group of people in the music ecosystem and to propose dataset-, model-, interface-, and governance-level responses.

Key Contributions

1. A stakeholder map of the music-AI ecosystem. The paper identifies five core roles — creators (composers and performers), distributors and marketing professionals, listeners, teachers and experts, and students — and describes how AI enters each role's work, while noting that in practice one person may perform multiple roles and that the analysis assumes one role per person.
2. A per-stakeholder account of representational bias harms. It organizes the consequences of bias into four named categories (misrepresentation harms, homogenization and exposure bias, cultural erosion and widening economic disparities, and opaque training processes and datasets) and links each to the specific stakeholders it affects.
3. An analysis of the design choices and technical gaps that produce bias. Four concrete causes are named: flattened genre labels such as "world" or "ethnic," missing regional and cultural metadata in training data, Western defaults baked into symbolic formats like MIDI, and prompt interfaces that steer users with culturally narrow defaults such as mood and genre tags like "cinematic" or "relaxing," compounded by overreliance on English metadata and tags.
4. Recommendations across four levels — dataset, model, system/interface, and governance — aimed at improving transparency and reducing socio-cultural inequity in music-AI, alongside a set of open research questions for the music and AI community.

Main Findings

• Bias is structural, not accidental. The paper states that representational skew arises from long-standing patterns of cultural dominance, market-focused choices, and platforms that favor certain music while leaving others out, building on prior work by Mehta et al. showing a large skew in global representation with decreasing focus on Global South cultures and genres such as country and folk.
• Equal representation is treated as infeasible. Because genres are fluid, hybrid, and constantly emerging in local, diaspora, and transnational contexts, the authors argue it is practically and epistemologically impossible for any AI system to equally represent all global musical expression. The stated goal is therefore fairness strategies under inevitable inequality, not parity.
• Misrepresentation harms trust and authenticity. Distorted outputs (the paper's example is distorted ragas) mischaracterize traditions, undermine creators' authenticity and trust, mislead listeners who intend to explore diversity, and risk being propagated by teachers and students who use these tools to build educational content.
• Homogenization and exposure bias narrow the whole ecosystem. Probabilistic models produce uniform outputs shaped by dominant datasets, and recommendation systems reinforce exposure to familiar music. The paper attributes to this: constrained innovation for creators, marginalization of small-scale distributors and labels working with underrepresented music, narrowed listener preferences, and limited creative discovery for students.
• Cultural erosion feeds an economic snowball. Generative models homogenizing traditional music reduce access to distinctive regional forms; because under-represented genres and their practitioners receive limited listener attention and distributors prioritize popular genres, the economic gap between creators of popular and under-represented genres widens, which the paper says leads to fewer creators from those communities pursuing such genres as a career and eventually to genre erosion.
• Opaque data and models block attribution. The paper states that closed-source models such as Udio and many open-source models, including MusicLM, are trained on closed datasets often not accessible to the community, leading to training on uncurated datasets with unclear composition and intent, making it impossible to trace where music originates and causing human creators to lose recognition. Listeners, teachers, and students may then treat AI-generated content as authentic.
• Four technical causes of homogenization are named. First, genre labels flattened into ambiguous catch-alls such as "world" or "ethnic" — the paper contrasts hyperspecific Western microgenres like melodic house with the treatment of non-Western forms as undifferentiated. Second, training data that rarely retains regional or cultural metadata such as place of origin, instruments, performance context, or linguistic association. Third, symbolic formats like MIDI that encode Western tonal and rhythmic defaults, marginalizing microtonality, heterophony, and non-metric rhythms. Fourth, prompt interfaces built on culturally narrow defaults and English metadata and tags, restricting exploratory access to underrepresented traditions.
• No quantitative results are reported. The paper reports no dataset sizes, no benchmark scores, and no model evaluations; its evidence base is a literature-grounded argument plus unstructured interviews.

Methodology in Plain English

The authors did not run experiments. They reviewed prior work on bias in AI (including earlier findings from NLP on dataset diversity and language inclusion, and prior music-specific studies on global representation skew), then identified the core roles in the music-AI ecosystem. To ground the analysis, they conducted unstructured interviews with researchers, practitioners, and professionals working in music and AI, focusing on risks tied to representational bias and fairness — particularly cultural inclusion, misrepresentation, and consequences for stakeholders. Based on those discussions, they worked through the implications for each stakeholder role, analyzed the design and technical choices inside music-AI systems that give rise to those implications, and derived recommendations at the dataset, model, system/interface, and governance levels. The paper frames itself around three research questions: who the stakeholders are and how bias affects them, which design choices and technical challenges shape how bias emerges, and what strategies can ensure fairness when equal representation is impractical. The specific number of interviewees, their identities, and the interview protocol details are not reported in the paper content.

Why This Matters

Impact on research. The paper shifts the fairness conversation in music AI away from purely legal and safety framings (copyright, deepfakes, watermarking, detection) toward representational and cultural fairness, and argues that music-AI development must move beyond efficiency to support diverse and equitable creative expression. It introduces the music-AI ecosystem as a site of fairness research and lays out open problems for the music and AI community.

Real-world applications.

• Generative composition tools inside DAWs: creators using AI-assisted composition could receive outputs that are culturally faithful rather than distorted pastiches, with attribution information attached to what influenced an output.
• Streaming and distribution: recommendation and marketing systems (the paper names Spotify, Spotify's Discover Weekly, and SoundOn by TikTok) could avoid reinforcing exposure bias and could give under-represented genres and their small-scale distributors and labels a fairer share of listener attention.
• Music education: AI tools used by teachers and students — for feedback, accompaniment, and learning environments, with Tonara named as an example — could avoid transmitting culturally homogenized material, and institutions such as The Royal Academy of Music are cited as actors who curate curricula and archives.
• Interfaces for non-expert users: systems offering "vibe music composition" without technical skill could surface specific tradition labels (the paper's examples: Gnawa, Taiko) rather than vague tags, with contextual information about style, region, and instruments.

Industry relevance. Record labels, streaming platforms, promotion networks, and model developers all sit inside the system the paper critiques. The governance recommendations — involving musicians and cultural experts in system design, defining guidelines for sensitive music, enabling opt-out mechanisms for creators, supporting compensation or licensing models, and extending AI regulation to cultural expression and creative labor — are directed at these commercial actors and at policymakers.

Future Directions

• Evaluating fairness and cultural representation in generated music, and building evaluation that goes beyond signal quality to account for genre diversity, style preservation, and cultural variation, which the paper acknowledges is still complex and emerging.
• Ensuring credit and consent for creators and communities through appropriate policies and governance, including traceability so users can see what data influenced an output.
• Supporting cultural fidelity without reinforcing stereotypes, including reducing language-driven biases in prompts and metadata.
• Extending symbolic infrastructures like MIDI to non-Western traditions so that microtonality, heterophony, and non-metric rhythms are representable.

Target Audience

This paper is most useful for music-AI researchers and model and dataset builders; AI ethics and fairness researchers looking at domains beyond text and images; music educators and institutions; music industry professionals in distribution, curation, and label work; and policymakers and governance bodies considering how AI regulation should cover cultural expression and creative labor. Because it is non-technical and contains no experiments, it is also accessible to musicians, cultural practitioners, and students who want to understand how music-AI systems may affect their traditions and work.

Authors’ abstract

In recent years, the music research community has examined risks of AI models for music, with generative AI models in particular, raised concerns about copyright, deepfakes, and transparency. In our work, we raise concerns about cultural and genre biases in AI for music systems (music-AI systems) which affect stakeholders including creators, distributors, and listeners shaping representation in AI for music. These biases can misrepresent marginalized traditions, especially from the Global South, producing inauthentic outputs (e.g., distorted ragas) that reduces creators' trust on these systems. Such harms risk reinforcing biases, limiting creativity, and contributing to cultural erasure. To address this, we offer recommendations at dataset, model and interface level in music-AI systems.