AI and Music at a Crossroads: The Role of the Music Information Retrieval (MIR) Community in Shaping Trustworthy, Open, and Culturally Inclusive Music‑AI

2.1 The MIR community

The International Society for Music Information Retrieval (ISMIR) is a central institutional hub for the MIR field, supporting research and education through its conferences, publications, and community initiatives (Dixon et al., 2018). These structures have encouraged interdisciplinary exchange, open knowledge‑sharing, methodological reflection, and collaboration across technical and music‑related disciplines. However, interdisciplinary work remains difficult because of differences in methods, terminology, and research practices across fields, limiting the uptake of MIR research in areas such as musicology and highlighting the need for stronger methodological and communicative bridges (Borsan et al., 2023; McKay and Cuenca Rodríguez, 2026).

2.2 Methodological evolution in MIR

MIR is a research field concerned with developing computational methods for processing, analyzing, searching, organizing, and accessing music‑related data, as well as supporting interaction with musical materials. It has developed as an interdisciplinary field combining musicology, information science, cognitive science, signal processing, and statistical modeling (Müller et al., 2011). Its methods have evolved from hand‑crafted features and relatively interpretable models toward data‑intensive deep‑learning approaches centered on learned representations, training protocols, and large datasets (Müller et al., 2019; Peeters et al., 2025). This transition has advanced many established tasks while intensifying questions about generalization, interpretability, scale, and evaluation validity (Gómez et al., 2024).

2.3 Evaluation practices and structural constraints

A distinctive feature of MIR has been its collective approach to comparative evaluation and reproducibility. MIREX established shared tasks, benchmarks, and evaluation procedures (Downie, 2008), while tools such as mir_eval and mirdata have supported standardized evaluation and reproducible dataset use (Raffel et al., 2014; Bittner et al., 2019). These practices operate under persistent structural constraints, including copyright restrictions; annotation costs; and limited access to large, diverse, and multimodal datasets (Gotham et al., 2025). Such constraints distinguish many academic research settings, which have often emphasized openness and shared evaluation, from industrial development based on proprietary resources and deployment‑oriented benchmarks (McFee et al., 2019).

MIR has also increasingly questioned the adequacy of accuracy‑based evaluation alone. Recent work emphasizes the validity of experimental conclusions and the need for perceptual, cultural, repertoire‑aware, and context‑sensitive assessment (Sturm and Flexer, 2023; Weck et al., 2024). These developments support a broader understanding of evaluation aligned with trustworthy and culturally inclusive music‑AI.

2.4 Data and software infrastructures

MIR research relies on shared datasets, repositories, and software spanning audio, symbolic music, metadata, annotations, and multimodal materials. Widely used resources include the RWC Music Database, GTZAN, the Million Song Dataset, MedleyDB, FMA, Freesound¹, and Dunya² (Goto et al., 2002; Tzanetakis and Cook, 2002; Bertin‑Mahieux et al., 2011; Bittner et al., 2014; Defferrard et al., 2017). Other important infrastructures include IMSLP³, MusicBrainz⁴, Discogs⁵, Wikipedia⁶, and OpenScore⁷, which support symbolic analysis, metadata enrichment, entity linking, and corpus construction.

Because many influential datasets remain copyrighted or non‑redistributable, including resources used for chord recognition and structural analysis, openness in MIR has always involved legal and practical compromises (Harte et al., 2005; Burgoyne et al., 2011; Smith et al., 2011). Concerns about dataset flaws, bias, representativeness, and benchmark overfitting have consequently motivated stronger attention to documentation, cultural diversity, and multimodal coverage (Morreale et al., 2023; Born, 2020; Gotham et al., 2025).

Open‑source software has complemented these resources by enabling transparent and reproducible workflows. Examples include MIRtoolbox, Sonic Visualiser, Essentia, librosa, Humdrum, music21, and Partitura (Lartillot et al., 2008; Cannam et al., 2010; Bogdanov et al., 2013; McFee et al., 2015; Huron, 2002; Cuthbert and Ariza, 2010; Cancino‑Chacón et al., 2022). By contrast, the open release and governance of music‑AI models remains less mature, particularly for large generative systems (Batlle‑Roca et al., 2025).

The methodological evolution and supporting infrastructures discussed in this section, together with the principal application areas introduced in the following section, are summarized in Table 1.

3.1 Music analysis and representation

A central contribution of MIR has been the development of computational methods for analyzing and representing music across audio, symbolic, textual, visual, and multimodal forms. These include automatic transcription, harmonic and structural analysis, source separation, tempo and beat tracking, and music similarity, among other tasks (Benetos et al., 2019; Pauwels et al., 2019; Nieto et al., 2020; Rouard et al., 2023; Foscarin et al., 2024; Volk et al., 2016). Such methods produce structured representations that support scientific inquiry, retrieval, musicological analysis, education, recommendation, and creative applications. These representations are necessarily selective, however, and their validity depends on how appropriately they relate computational outputs to the musical phenomena and contexts being studied.

3.2 MIR systems for access, interaction, and musical practice

MIR has also developed systems that support access to musical collections; interaction with musical materials; and participation in learning, listening, performance, and creative practice. Recommendation, retrieval, categorization, visualization, and browsing systems enable users to navigate large collections and discover relationships among recordings, scores, metadata, and other musical resources (Schedl et al., 2022; Knees and Schedl, 2016; Knees et al., 2020; Goto and Dannenberg, 2019). Related research addresses performance analysis, educational feedback, score‑following, accompaniment, and interactive machine learning, often emphasizing human control and collaboration rather than autonomous content production (Lerch et al., 2020; Dannenberg and Raphael, 2006; Nistal et al., 2024). MIR methods have also supported the study, documentation, and preservation of Western classical, popular, and culturally diverse musical traditions (Weiß et al., 2019; Oramas et al., 2025; Serra, 2011; Duan et al., 2023).

3.3 Music generation

Music generation is connected to MIR research through symbolic modelling, statistical learning, analysis, and interactive systems. Generative and discriminative approaches frequently share representations and model architectures, with related methods being applied to transcription, beat‑tracking, representation learning, source separation, transformation, and generation (Hawthorne et al., 2018; Zhao et al., 2022; Flores Garcia et al., 2023; Li et al., 2024; Rouard et al., 2023). Contemporary generation spans audio, symbolic, and multimodal settings, including score and event‑sequence generation; harmonization; accompaniment; performance rendering; and systems connecting language, notation, and sound.

What distinguishes recent generative music‑AI is therefore less the appearance of entirely new modeling principles than a shift in scale, modality, accessibility, and degree of autonomy. Among recent audio‑generation systems, common approaches use either continuous latent representations combined with diffusion models or discrete representations derived from neural audio coding and autoregressive models (Pasini et al., 2024; Agostinelli et al., 2023; Défossez et al., 2023). Although these systems offer substantial generative capabilities, their limited transparency, controllability, accessibility, and reproducibility create important challenges for scientific scrutiny and meaningful human–AI interaction.

4.1 Ethical and legal frameworks

The rapid expansion of AI applications in music both intensifies existing ethical and legal questions and introduces new ones that extend beyond technical performance or usability. Ethical frameworks articulate values, responsibilities, and societal expectations, whereas legal frameworks define enforceable rules and compliance requirements (Floridi et al., 2018). Both are essential for guiding responsible innovation.

Across policy, standards, and research communities, these concerns have converged around the concept of trustworthy AI. International and regional frameworks developed by UNESCO (2021), the European Commission (2019), the Organisation for Economic Co‑operation and Development (OECD, 2019), IEEE (2019), and the National Institute of Standards and Technology (NIST, 2023) emphasize principles such as human rights, transparency, fairness, accountability, human oversight, technical robustness, and environmental sustainability. UNESCO’s Recommendation on the Ethics of Artificial Intelligence is particularly relevant as a globally adopted framework that connects ethical principles with concrete areas of policy action. In the music domain, these principles acquire specific meaning through data practices, evaluation methodologies, cultural representation, and deployment contexts that MIR research has long engaged with.

4.2 Ethical challenges specific to music‑AI

Ethical challenges in music‑AI stem from music’s role as a creative, cultural, and social practice. Respect for creative integrity is central: AI systems should support musicians and musical communities rather than obscure authorship or replace human agency. Closely related are issues of attribution and credit, particularly when systems are trained on musical corpora whose provenance and labor histories are not always visible.

These concerns extend beyond the status of musical works and datasets to the people whose activity makes those materials possible. Recordings and other musical resources embody the labor and decisions of composers, performers, producers, engineers, annotators, and other cultural workers. When such materials are treated primarily as training data, these contributions may become obscured, particularly when consent, attribution, compensation, and control over reuse are unclear. Ethical analysis should therefore consider not only what data a system uses but also whose work and agency are represented in that data, who benefits from its use, and who bears the resulting risks (Hubbles et al., 2026).

Cultural sensitivity constitutes a further ethical concern. Music is shaped by culturally specific meanings, practices, and structural conventions, and computational approaches risk misrepresentation when context is ignored (Holzapfel et al., 2018; Huang et al., 2023). Bias awareness therefore extends beyond technical fairness to questions of representation, visibility, and access across repertoires (Porcaro et al., 2021).

A further challenge is the opacity of many music‑AI systems. Limited information about training data, model design, evaluation procedures, and system limitations can make it difficult for researchers, musicians, and users to understand how outputs are produced, assess their reliability, or contest harmful or misleading results. Transparency is therefore not only a technical requirement but also an ethical condition for informed use, meaningful scrutiny, and accountability (Bryan‑Kinns et al., 2025).

Ethical responsibility also includes how humans interact with music‑AI systems. Systems should support informed use by communicating limitations, uncertainty, and appropriate contexts, thereby enabling meaningful human–machine collaboration. Finally, ethical reflection must address power and governance: the concentration of data, models, and infrastructures raises questions about who controls musical technologies and whose interests they serve (Morreale, 2021).

4.3 Research and development contexts

Public debate often blurs a key distinction between exploratory research and the commercial deployment of music‑AI systems. Although the boundary is not always clear and legal treatment varies across jurisdictions, the two contexts differ in their typical purposes, scale, responsibilities, and consequences, as summarized in Table 2. In some jurisdictions, scientific research benefits from specific exceptions or limitations, particularly those related to text and data mining. In the European Union, such exceptions permit the analysis of copyrighted works for research under defined conditions (EU, 2019, 2024), while the AI Act also excludes certain research, testing, and development activities from its scope. In particular, Article 2(8) provides that the Act does not apply to activities concerning AI systems or models before they are placed on the market or put into service, although the exclusion does not cover testing in real‑world conditions and remains subject to other applicable EU law (EU, 2024; Meszaros et al., 2026). The Act separately excludes systems or models specifically developed and put into service solely for scientific research and development under Article 2(6). These provisions nevertheless leave important ambiguities around mixed academic–commercial projects, evolving intended uses, and the transition from laboratory research to real‑world deployment (Meszaros et al., 2026). In the United States, related research practices have been enabled through interpretations of ‘fair use’ (Sag, 2019).

By contrast, commercial deployment is subject to stricter legal and ethical obligations, including licensing, accountability, and consumer protection (Oğul, 2024). Research prototypes typically prioritize exploration and validation rather than scale or market impact. Treating exploratory research as equivalent to deployed products risks constraining legitimate research without addressing the responsibilities associated with large‑scale commercial use.

4.4 Copyright and generative music‑AI

Copyright law has become one of the most contested areas in relation to generative music‑AI (e.g., District Court of Massachusetts, 2024). Designed for pre‑digital modes of creation and distribution, its assumptions about authorship, originality, and human intent are increasingly challenged by data‑driven generative systems (Sobel, 2017). Legal questions surrounding training data, model outputs, attribution, and rights allocation remain unsettled and vary across jurisdictions (Sturm et al., 2019).

These tensions have prompted continuing debate about whether and how existing legal approaches should be clarified or adapted to distinguish research from commercial use, improve consistency, and balance the interests of creators, users, and technology developers (Lucchi, 2025).

Copyright provides only a partial account of these concerns. Even where the use of musical material may be legally permitted, questions remain about consent; recognition; compensation; creative labor; and the distribution of benefits and harms among musicians, producers, platforms, and technology developers.

4.5 Regulation as a research driver

Recent regulatory developments are actively shaping music‑AI research agendas. The AI Act (EU, 2024), for example, introduces transparency obligations for generative AI and explicitly calls for mechanisms to mark and identify AI‑generated audio while also acknowledging the absence of mature technical solutions. Similar priorities appear in international frameworks such as the Organisation for Economic Co‑operation and Development (OECD) AI Principles (Oğul, 2024) and the National Institute of Standards and Technology (NIST) AI Risk Management Framework (NIST, 2023).

For MIR, these developments represent research opportunities rather than mere compliance challenges (Serra et al., 2026). Topics such as content marking, watermarking, fingerprinting, metadata management, provenance tracking, and forensic analysis overlap with established areas of MIR expertise and create opportunities for the field to contribute methods for more auditable and trustworthy music‑AI systems.

4.6 Scientific principles

Based on the ethical and legal frameworks reviewed above, the music‑specific challenges identified in Sections 4.2 to 4.5, and the scientific practices discussed earlier in relation to MIR evaluation and research infrastructures, we identify four principles that are particularly relevant to current debates on music‑AI: transparency, accountability, provenance, and sustainability, as summarized in Table 3. These principles were selected through discussion among the authors as a concise synthesis of recurring concerns across trustworthy AI frameworks and MIR research. They are not intended as an exhaustive or entirely novel framework, nor as independent or mutually exclusive categories. Rather, they provide an organizing structure for connecting broad ethical and regulatory expectations with areas in which the MIR community has relevant methodological experience, while also making visible important limitations and gaps in current practice. Fairness, diversity, and bias awareness are not presented as a separate fifth principle but instead as cross‑cutting concerns that should inform the interpretation and application of all four principles.

These principles are not unique to AI ethics or recent regulatory frameworks; they are also long‑standing aspirations of responsible scientific practice across disciplines. Within MIR, they acquire a particularly concrete meaning because of the field’s engagement with complex musical data, subjective and culturally situated evaluation criteria, copyrighted materials, and computationally demanding methods. Elements of these principles can be found in existing MIR research practices, infrastructures, and community initiatives, but their adoption remains uneven and is often constrained by technical, legal, institutional, and economic factors. We therefore present them not as a description of established community‑wide practice but as criteria against which current practices can be critically assessed and future work can be guided.

Transparency in music‑AI refers to making the assumptions, inputs, and limitations of systems explicit. This includes clear documentation of model architectures and design choices; evaluation protocols; and the communication of known biases, uncertainties, uneven performance across repertoires and communities, and appropriate use contexts. Such transparency is particularly important in benchmark‑driven research, where improvements in aggregate metrics may obscure design complexity or fail to reflect the underlying musical problem. Transparency enables meaningful scrutiny of system behavior and supports reproducibility and informed interpretation of results.

Accountability concerns the ability to assign responsibility for the development, evaluation, and deployment of music‑AI systems. It can be supported through documented datasets; reproducible experimental pipelines; appropriate and culturally diverse benchmarks; and explicit reporting of failure cases and limitations across repertoires, populations, and contexts. In practice, however, benchmark‑driven and publication‑oriented incentives may discourage methodological simplicity, replication, negative results, and the reporting of unsuccessful projects. Accountability therefore depends not only on research procedures, but also on publication and institutional structures that make claims, errors, and failures open to scrutiny and correction.

Provenance addresses the traceability of musical data and model outputs: where they originate, how they were curated or annotated, and which processing steps they have undergone. From an MIR perspective, provenance encompasses licensing and data origin; annotation procedures; preprocessing operations such as segmentation, augmentation, cleaning, and deduplication; model lineage; and output traceability. Metadata standards, fingerprinting, watermarking, and identification techniques can contribute to verification, attribution, and auditing, but they do not guarantee complete provenance. Large‑scale web collection, automated curation, incomplete metadata, and repeated dataset reuse make traceability an increasingly difficult and unresolved challenge.

Sustainability has both environmental and infrastructural dimensions. Environmentally, it involves awareness of the computational and energy costs of training and deploying music‑AI systems, including transparent reporting of model scale and resource use. Infrastructurally, it concerns the long‑term viability of the scientific ecosystem—the maintenance; governance; and accessibility of datasets, tools, benchmarks, and educational resources beyond individual projects or funding cycles.

Taken together, these principles provide criteria for critically assessing current MIR practices and identifying priorities for future work. Progress will require more than individual commitments: it will depend on community‑level incentives, sustained infrastructures, appropriate governance mechanisms, and institutional support. This is particularly important for early career researchers, who may face strong pressure to prioritize short‑term publication outcomes over activities such as documentation, replication, infrastructure maintenance, and the reporting of negative results. Recognizing and rewarding these contributions is therefore necessary to make transparent, accountable, traceable, and sustainable research feasible across different institutions, resource levels, and career stages.

5.1 Collective and mission‑oriented collaboration

Many of the key challenges in music‑AI—including transparency and provenance, cultural representation, copyright ambiguity, evaluation complexity, and environmental sustainability—cannot be addressed through isolated projects or short‑term initiatives. They call instead for mission‑oriented efforts that bring together machine learning, signal processing, musicology, human–computer interaction, ethics, law, policy, and artistic practice. Such initiatives require durable collaborations and shared agendas that extend beyond individual projects or funding cycles. From an MIR perspective, an important contribution lies in helping ensure that technological development remains connected to musical knowledge, ethical reflection, and engagement with diverse musical communities while also strengthening the mutual intelligibility needed for sustained exchange with adjacent fields.

Within the ISMIR community, such shared agendas can be shaped through conference themes and calls for papers, special sessions and community initiatives, funding priorities, peer‑review criteria, and recognition or incentives for forms of work that support collective goals.

5.2 Sustaining open infrastructures

Open datasets, models, and tools have been central to music‑AI research, but their long‑term viability increasingly depends on stable funding, institutional stewardship, and collaborative governance. Key challenges include long‑term hosting and curation; clear documentation of provenance and licensing; and governance structures that support maintenance, inclusivity, and accountability. As proprietary models and closed datasets become more prevalent, the continued availability of robust public infrastructures is essential to preserve reproducibility, transparency, and scientific independence—while also acknowledging the practical limits and responsibilities associated with openness. These infrastructures should also be guided by FAIR principles—making research resources Findable, Accessible, Interoperable, and Reusable (Wilkinson et al., 2016)—as a basis for improving their accessibility, interoperability, documentation, and long‑term reuse.

In the context of ISMIR, dedicated grants and collective funding mechanisms could support shared datasets, software, benchmarks, and repositories, while distributing maintenance and governance across institutions rather than individual laboratories.

5.3 Toward rigorous and context‑ sensitive evaluation

Evaluation plays a decisive role in shaping research incentives and deployment decisions. In music‑AI, meaningful evaluation extends beyond accuracy metrics to include perceptual relevance, cultural validity, fairness, robustness across repertoires, and ecological validity. Future work should therefore prioritize transparent and auditable evaluation pipelines; benchmarks that reflect musical diversity and context; and explicit documentation of assumptions, limitations, uncertainty, and failure cases. At the same time, it is important to acknowledge that evaluation—especially for generative systems—remains an open and difficult problem, requiring continued methodological research and critical scrutiny rather than definitive solutions.

Within ISMIR, community‑supported open‑source tools, benchmarks, repositories, and reporting templates could facilitate maintenance, auditing, replication, and the gradual inclusion of new repertoires and evaluation criteria. Conference and peer‑review guidelines could encourage their use and contribution.

5.4 Cultural diversity, inclusion, and responsibility

Music‑AI systems influence how musical cultures are represented, accessed, and valued. Ensuring that these systems do not reinforce existing forms of inequity or distort cultural practices is both a scientific and ethical challenge. Work in this area must emphasize the inclusion of underrepresented repertoires, sustained collaboration with musicians, cultural practitioners and community experts, culturally informed annotation and evaluation practices, and safeguards against extractive or exploitative data use. Rather than treating diversity as an auxiliary concern, music‑AI research must integrate cultural responsibility into its core methodological choices.

Within the ISMIR community, this could be supported through dedicated conference topics and calls for papers, special sessions, workshops, and community‑led initiatives focused on underrepresented musical traditions and research perspectives. Travel grants, fee waivers, mentoring schemes, and targeted support for researchers and practitioners from underrepresented regions could lower barriers to participation.

5.5 Bridging research, industry, policy, and society

As music‑AI becomes increasingly entangled with legal, economic, and cultural structures, sustained dialogue across sectors is essential. Research alone cannot determine acceptable uses, resolve questions of attribution and rights, or establish appropriate governance arrangements, while policy and industry discussions risk becoming technically unrealistic when disconnected from research evidence and musical practice. MIR can contribute by clarifying the capabilities and limitations of current technologies; developing shared vocabularies around transparency, provenance, attribution, and robustness; and providing empirical evidence for standards and policy discussions. Cross‑sector engagement should also examine how music‑AI systems redistribute visibility, income, labor, control, and bargaining power, since recommendation, rights‑identification, source‑separation, and generative systems can affect which artists are discovered, whose work is monetized, and who controls the conditions of musical production and circulation. Shared approaches to documentation, metadata, licensing, content identification, and traceability are therefore important for supporting attribution, auditing, regulatory compliance, and responsible reuse across research and commercial settings.

Within the ISMIR community, this engagement could be strengthened by establishing a dedicated policy and governance strand in the conference program, comparable to the existing industry program. It could include invited panels, policy‑oriented papers, workshops, and structured discussions involving researchers, musicians, industry representatives, policymakers, legal experts, and civil society organizations.

5.6 Education and capacity‑building

The long‑term impact of music‑AI will depend on the capacity of musicians, researchers, educators, and policymakers to critically engage with these technologies. Education should combine MIR and machine learning with musical knowledge, ethics, law, cultural studies, and reflection on the social and economic contexts in which AI systems operate. Capacity‑building should be reciprocal: musicians and cultural practitioners need opportunities to understand and influence AI systems, while technical researchers need to recognize forms of musical, cultural, and professional knowledge that cannot be reduced to computational data. This requires accessible tools and documentation, interdisciplinary curricula, participatory design activities, and opportunities for exchange between technical and non‑technical communities. It must also address unequal access to technical and financial resources so that a wider range of institutions and communities can participate in shaping datasets, models, standards, and research agendas.

Within the ISMIR community, these goals could be supported through tutorials, summer schools, doctoral consortia, mentoring programs, and conference activities that bring together technical researchers, musicians, educators, legal scholars, and policymakers. Dedicated grants, fee waivers, travel support, open educational resources, and partnerships with institutions in underrepresented regions could broaden access to training and participation.

These priorities form a connected agenda. Open infrastructures support transparent and reproducible evaluation; meaningful evaluation depends on culturally appropriate data, concepts, and expertise; responsible governance requires collaboration across research, industry, policy, and musical communities; and education is a factor determining who can participate in these processes. Advancing trustworthy music‑AI therefore requires coordinated investment in people, infrastructures, methods, and institutional arrangements, together with incentives that recognize transparency, cultural responsibility, interdisciplinary collaboration, and long‑term stewardship. These efforts should ultimately be assessed by whether they contribute to music‑AI that is scientifically robust, ethically grounded, culturally inclusive, human‑centered, openly governed, and sustainable.