Within Music
Can AI Learn From Copyrighted Music?
AI music systems raise disputes over whether copyrighted recordings and compositions can be used for training without consent.
On this page
- Training data and permission
- Copyright, consent and compensation
- Why the dispute affects human creators
Page outline Jump by section
Introduction
AI music systems challenge copyright because they need large quantities of existing music to learn how songs, recordings, voices, genres and production styles work. The dispute is not simply about whether a machine can make a song. It is about whether companies may copy copyrighted recordings and compositions into training datasets without permission, whether that copying is excused by law, and whether musicians should be paid when their work becomes part of the raw material for a competing system. The argument has moved quickly from theory to litigation: major record companies sued Suno and Udio in 2024 over alleged mass copying of sound recordings, while later settlements began pushing parts of the market towards licensed AI music models. RIAA [Pitchfork]pitchfork.comSource details in endnotes.
Why Training Is Different From Ordinary Listening
A human listener can learn from music without making a commercial copy of a catalogue. AI training is different because it typically involves collecting files, converting them into machine-readable form, extracting patterns and repeatedly processing them to adjust a model. In copyright terms, that process may involve reproductions of recordings, compositions or lyrics even before any AI-generated output reaches the public. The US Copyright Office’s 2025 report framed training as a chain of acts that can affect copyright markets through lost sales, market dilution and lost licensing opportunities, rather than as a single abstract act of “learning”. [U.S. Copyright Office]copyright.govSource details in endnotes.
Music makes this especially sensitive because a commercially released track usually contains more than one right. A sound recording may be owned or controlled by a label; the underlying composition may involve songwriters and publishers; lyrics may have separate licensing issues; performances may raise performer and voice-related concerns. A model trained on a recording may therefore touch multiple layers of rights at once, even if the final generated track does not reproduce a whole chorus or sample in the traditional sense.
The legal question is not settled everywhere. AI developers often argue that training is transformative because the system is not built to distribute copies of the original works but to generate new material. Music rightsholders answer that training requires unlicensed copying, that the resulting systems can compete directly with human-made songs, and that copyright already provides licensing markets for uses of recordings and musical works. The US Copyright Office noted that licensing markets for AI training were already being discussed or developed in sectors including music, and that such markets need not be old or universal to matter in a copyright analysis. [U.S. Copyright Office]copyright.govSource details in endnotes.
Training Data and Permission
The central permission problem is practical as well as legal. A modern music model may need huge volumes of audio to learn rhythm, harmony, vocal timbre, mixing choices, genre conventions and how words sit against melody. If that material is drawn from commercial catalogues, the model developer may need permission from parties who control both recordings and songs. That is manageable for a narrow licensed catalogue, but far harder for internet-scale scraping, old recordings with unclear ownership, remixes, samples, covers and user-uploaded tracks.
The Suno and Udio cases made this issue concrete. The Recording Industry Association of America announced lawsuits in June 2024 on behalf of major labels, alleging that the companies copied copyrighted sound recordings without permission to train music-generation services. Reuters reported that the labels sought statutory damages of up to US$150,000 per copied song and alleged copying of hundreds or thousands of works across the two cases. [RIAA]riaa.comOpen source on riaa.com.
The defendants’ broad answer was that training on copyrighted works can be lawful fair use under US law. That matters because fair use is not a blanket permission rule; it is a context-specific defence that weighs factors such as purpose, amount used and market effect. In music, the market-effect question is unusually sharp because AI songs can occupy the same attention economy as the recordings that may have helped train the model. The US Copyright Office stressed that courts consider not only harm from one defendant’s act but also the effect of unrestricted and widespread conduct of the same kind. [U.S. Copyright Office]copyright.govSource details in endnotes.
Permission also becomes harder when AI firms treat training datasets as trade secrets. Developers may argue that dataset disclosure would reveal competitive information, while rightsholders argue that they cannot enforce rights, negotiate licences or verify opt-outs without knowing what was used. This is why transparency has become a governance issue rather than a mere paperwork issue: without some form of dataset disclosure, audit trail or rights-reservation mechanism, consent can be almost impossible to check after the fact.
Copyright, Consent and Compensation
The policy fight is often described as “licensing versus innovation”, but the real choice is more precise: who bears the transaction cost of permission? An opt-in system requires AI developers to secure licences before using copyrighted music. An opt-out system allows some uses unless rightsholders reserve their rights in a recognised way. A broad exception allows training with little or no case-by-case permission. Each approach favours different actors.
The UK’s 2026 report on copyright and AI shows how divided the field is. In the UK consultation, a broad data-mining exception with rights reservation was supported by only 3% of respondents, while the report recorded strong opposition from creative industries, individual creators and performers. A separate option to strengthen copyright so that licensing is required for AI development was supported by 81% of respondents. [GOV.UK]GOV.UKReport on Copyright and Artificial IntelligenceReport on Copyright and Artificial Intelligence
The opt-out model sounds tidy until it meets the realities of music. A global hit may be controlled by sophisticated companies that can deploy technical rights reservations, but a session musician, independent songwriter, small label or self-releasing artist may not know where their work has travelled or how to mark it machine-readably. The UK report recorded concerns that opt-outs could impose a heavy administrative burden on rightsholders, especially individuals and small businesses, while AI developers themselves disagreed over what form of rights reservation should count. [GOV.UK]GOV.UKReport on Copyright and Artificial IntelligenceReport on Copyright and Artificial Intelligence
Licensing is not a magic fix either. Direct deals can bring money and control, but they may favour large catalogues that have the leverage to negotiate. Collective licensing can reduce transaction costs and help smaller rightsholders bargain together, but it may not suit every right, repertoire or market. The UK government described AI training licensing as a new and growing market, noted that collective licensing is often useful where direct licensing is impractical, and said it would keep market-led licensing under review rather than impose a licensing structure immediately. [GOV.UK]GOV.UKReport on Copyright and Artificial IntelligenceReport on Copyright and Artificial Intelligence
Recent music deals show the market moving towards controlled licensing, but not uniformly. Universal Music Group reached an agreement with Udio that includes compensation for participating UMG artists and songwriters for training and outputs, with a new platform planned under licensed terms. Warner Music Group later settled with Suno, with Reuters reporting that Suno would replace current models with licensed AI models in 2026 and introduce download restrictions. [Pitchfork]pitchfork.comSource details in endnotes.
Why Outputs Do Not Solve the Training Problem
A common misunderstanding is that copyright is only implicated if an AI-generated song sounds too much like a protected work. Output similarity matters, but it is not the whole dispute. Rightsholders argue that the unauthorised copying happens at the training stage, even if many outputs are new. Developers answer that training is analogous to analysis and that liability should depend on whether the system produces infringing material. This difference explains why the same case can involve both technical questions about training data and familiar questions about substantial similarity.
Music also has a distinctive “style” problem. Copyright generally does not protect style in the abstract: no one owns “1980s synth-pop”, “trap hi-hats” or “Beatles-like harmony” as a general idea. Yet a model that can produce tracks strongly associated with a living artist’s sound may still affect that artist’s market, reputation and negotiating power. The US Copyright Office noted that even where outputs are not substantially similar to a specific work, stylistic imitation made possible by training can affect a creator’s market, while also recognising that copyright protection for style itself is legally limited. [U.S. Copyright Office]copyright.govSource details in endnotes.
This is why the music industry often links training disputes to voice, likeness and artist identity. A synthetic song may not copy a full recording, but it can imitate a singer’s timbre, a producer’s sonic signature or a genre niche built by real performers. That may fall partly outside traditional copyright and into publicity rights, passing off, unfair competition, contract or platform policy, depending on jurisdiction. Still, the training question remains the foundation: whether the model should have been allowed to absorb the relevant recordings and compositions in the first place.
How Regulation Is Trying to Catch Up
Governments are experimenting with transparency, rights reservation and licensing support rather than converging on one global rule. The EU AI Act requires general-purpose AI providers to address transparency and copyright compliance obligations, and the European Commission’s General-Purpose AI Code of Practice includes separate transparency and copyright chapters to help providers comply with Article 53 obligations. [Digital Strategy]digital-strategy.ec.europa.euSource details in endnotes.
For music, EU-style transparency matters because it can make hidden training practices more visible. If a model provider must document training processes or summarise training content, rightsholders gain at least some basis for asking whether their catalogues were used. However, high-level summaries may still be too vague for a songwriter or small label trying to prove that a particular work was copied. That tension is why creator organisations continue to argue that transparency must be detailed enough to support enforcement, not merely broad enough to satisfy regulators.
The UK has taken a more cautious path after strong opposition to an opt-out exception. Its 2026 report proposed monitoring market-led licensing and developing best practice on input transparency and technical standards, rather than immediately imposing a new licensing mechanism. That approach avoids rushing a flawed framework, but it also leaves uncertainty for musicians and developers while courts, licences and platform policies evolve. [GOV.UK]GOV.UKReport on Copyright and Artificial IntelligenceReport on Copyright and Artificial Intelligence
The United States remains especially important because many leading AI companies and major music rights disputes are centred there. The Copyright Office did not recommend an immediate compulsory licensing scheme for generative AI training, but its analysis rejected the idea that all training is automatically fair use or automatically infringing. For music, that means case-specific litigation and negotiated licensing will continue to shape the practical rules before a single legislative answer emerges. [McDermott]mcdermottlaw.comSource details in endnotes.
Why the Dispute Affects Human Creators
For listeners, AI music may look like another production tool. For working musicians, the training dispute is about bargaining power. If existing recordings can be used without permission to build systems that generate low-cost substitutes, the value created by past human work may be transferred into products that compete with future human work. That risk is clearest in production music, jingles, demos, library music, background tracks and other areas where clients may choose speed and price over a named artist.
The effect will not be the same for every creator. Some established artists and catalogues may benefit from licensing deals, official remix tools or fan-interaction products. Some disabled, independent or experimental musicians may use AI tools to sketch ideas, arrange demos or overcome technical barriers. But those benefits do not erase the copyright question: a tool can be creatively useful and still be built on contested training data. The governance challenge is to make space for legitimate AI-assisted creativity without treating the existing music ecosystem as free infrastructure.
Compensation is also about distribution, not just whether money changes hands. A label-level licence may pay a rightsholder, but performers, songwriters and producers may not automatically share equally unless contracts, collective agreements or policy rules require it. The UK report recorded creator-group concerns that AI licensing benefits may not flow to individual or small rightsholders, and that consent should be explicit, specific and meaningful rather than buried in old or generic contracts. [GOV.UK]GOV.UKReport on Copyright and Artificial IntelligenceReport on Copyright and Artificial Intelligence
The deepest concern is cultural as well as economic. Music scenes depend on people taking risks, developing voices, building audiences and making a living long enough to keep creating. If AI systems can absorb that labour without permission, the incentive to invest in new human artists may weaken. If permission and payment systems are too rigid or expensive, smaller AI developers and independent musicians may be locked out while only large technology firms and major catalogues can participate. A sustainable settlement has to solve both sides of that problem.
The Governance Choice Ahead
The future of AI music copyright is unlikely to be decided by one court case or one law. It will be shaped by a mix of litigation, licensing deals, technical standards, platform rules, dataset transparency and public expectations about consent. The emerging direction is already visible: the most durable AI music services will probably need clearer rights pathways, auditable training practices, output controls and payment models that recognise both catalogue owners and human creators.
The hard question is not whether AI can learn from music. It can. The hard question is whether that learning should happen by unlicensed copying, by negotiated permission, by statutory exception, or by a hybrid system that distinguishes research, commercial deployment, public-domain works, licensed catalogues and creator opt-outs. Music copyright is being forced to define what “learning” means when the learner is not a person, the training material is a valuable cultural catalogue, and the outputs can enter the same market as the humans whose work made the system possible.
Amazon book picks
Further Reading
Books and field guides related to Can AI Learn From Copyrighted Music?. Use these as the next step if you want deeper reading beyond the article.
Platform Revolution
Directly useful for understanding moderation and governance tradeoffs.
All You Need to Know About the Music Business
Explains rights ownership and compensation structures.
The copyright handbook
First published 1992. Subjects: Popular works, Copyright, International Copyright, Copyright, united states, Copyright, international.
eBay marketplace picks
Marketplace Samples
Example marketplace items related to this page. Use the search link to explore similar finds on eBay.
Endnotes
-
Source: riaa.com
Link: https://www.riaa.com/record-companies-bring-landmark-cases-for-responsible-ai-againstsuno-and-udio-in-boston-and-new-york-federal-courts-respectively/ -
Source: pitchfork.com
Link: https://pitchfork.com/news/universal-music-group-and-ai-music-company-udio-reach-agreement-in-lawsuit -
Source: reuters.com
Link: https://www.reuters.com/legal/litigation/warner-music-group-settles-copyright-case-with-suno-licensed-ai-music-2025-11-25/ -
Source: copyright.gov
Link: https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-3-Generative-AI-Training-Report-Pre-Publication-Version.pdf -
Source: reuters.com
Link: https://www.reuters.com/technology/artificial-intelligence/music-labels-sue-ai-companies-suno-udio-us-copyright-infringement-2024-06-24/ -
Source: GOV.UK
Title: Report on Copyright and Artificial Intelligence
Link: https://www.gov.uk/government/publications/report-and-impact-assessment-on-copyright-and-artificial-intelligence/report-on-copyright-and-artificial-intelligence -
Source: copyright.gov
Link: https://www.copyright.gov/ai/ -
Source: copyright.gov
Title: and Artificial Intelligence Part 2 Copyrightability Report
Link: https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-2-Copyrightability-Report.pdf -
Source: engage.pc.gov.au
Link: https://engage.pc.gov.au/document/2640 -
Source: riaa.com
Title: Udio Complaint 6.24.241
Link: https://www.riaa.com/wp-content/uploads/2024/06/Udio-Complaint-6.24.241.pdf -
Source: digital-strategy.ec.europa.eu
Link: https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai -
Source: digital-strategy.ec.europa.eu
Title: ai code practice
Link: https://digital-strategy.ec.europa.eu/en/policies/ai-code-practice -
Source: mcdermottlaw.com
Link: https://www.mcdermottlaw.com/insights/us-copyright-office-issues-report-addressing-use-of-copyrighted-material-to-train-generative-ai-systems/ -
Source: digital-strategy.ec.europa.eu
Title: eu A I Act | Shaping Europe’s digital future
Link: https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai -
Source: mishcon.com
Title: us copyright office report part 3 generative ai training
Link: https://www.mishcon.com/news/us-copyright-office-report-part-3-generative-ai-training -
Source: completemusicupdate.com
Link: https://completemusicupdate.com/us-copyright-office-report-on-whether-ai-training-is-fair-use-concludes-it-depends-but-generally-favours-copyright-owners/
Additional References
-
Source: youtube.com
Title: How To Copyright AI
Link: https://www.youtube.com/watch?v=lCIFGCJDj90Source snippet
AI music training copyright issues explained How To Copyright AI (Step By Step Guide) | Lawyer Explains Top Music Attorney...
-
Source: linkedin.com
Link: https://www.linkedin.com/posts/meera-nair-11baa735_understanding-cc-licenses-and-ai-training-activity-7329142745225199616-f8E0 -
Source: taylorwessing.com
Link: https://www.taylorwessing.com/en/campaigns/de/2025/ai-and-copyright-tracker -
Source: linkedin.com
Link: https://www.linkedin.com/pulse/ai-training-uks-proposed-opt-out-copyright-model-shift-irving-david-un41e -
Source: facebook.com
Link: https://www.facebook.com/musicradartech/posts/ai-powered-music-generation-platforms-suno-and-udio-have-come-under-fire-in-rece/1118267656996650/ -
Source: humanartistrycampaign.com
Link: https://www.humanartistrycampaign.com/ -
Source: reddit.com
Link: https://www.reddit.com/r/creativecommons/comments/10c2iw2/is_there_a_creative_commons_license_that/ -
Source: reddit.com
Link: https://www.reddit.com/r/OpenAI/comments/1dni56e/record_labels_sue_suno_and_udio_over_aigenerated/ -
Source: instagram.com
Link: https://www.instagram.com/reel/DV1bH9SEfes/ -
Source: musicbusinessworldwide.com
Link: https://www.musicbusinessworldwide.com/suno-moves-to-keep-size-of-its-ai-training-data-sealed-in-umg-and-sonys-copyright-case-citing-competitive-harm/
Topic Tree