AI Models and Copyrighted Music: What to Know

Yes, AI models can be trained on copyrighted music in a technical sense, but whether a particular training use is lawful depends on the jurisdiction, the source and terms of the recordings, how copies are made, and how the model is used. Creators and developers should not assume that publicly available music is free to use, or that training automatically is—or is not—infringement.

What music-model training involves

A music-generation system may learn from audio recordings, sheet music, lyrics, or descriptive labels such as genre and instrumentation. Training commonly involves collecting files, converting them into a form the software can process, and repeatedly analyzing patterns in the data. Those steps may require making copies, which can raise copyright questions even if the finished model does not contain a conventional audio library.

Copyright can apply separately to different parts of a track. A sound recording and the underlying musical composition are distinct works; lyrics may also be protected. A single recording can therefore involve multiple rights holders, such as a performer or label for the recording and songwriters or publishers for the composition. Rights vary by country and by contract.

Not every training system uses complete commercial recordings. Some are trained on licensed catalogs, commissioned music, public-domain works, or audio made available under specified terms. Others use mixed datasets. The details of the dataset and the process matter more than the broad label “AI music.”

Why the legal answer is unsettled

Copyright laws differ across jurisdictions, and courts and lawmakers are still addressing how existing rules apply to generative AI. Some legal arguments focus on whether copying for analysis is permitted under doctrines such as fair use in the United States, or text-and-data-mining exceptions in some other legal systems. These are not universal permissions: their conditions, scope, and application can differ, and a court’s assessment is fact-specific.

Relevant facts may include whether the source was lawfully obtained, whether copies were retained, the purpose of the use, any market effects, and whether the system can reproduce protected expression. A developer’s location, the location of data processing, and where a model is offered may also matter. A claim that training is “transformative” or “noncommercial” does not by itself settle the question.

There is also a difference between training and outputs. A model’s training may involve one set of legal issues, while generating, distributing, or performing an output can raise others. If an output reproduces recognizable protected music, that can create risk even when the developer believes the training was lawful. Conversely, a musical resemblance or shared genre convention does not automatically prove copying; evidence and legal standards matter.

What creators should know

Creators should distinguish between their preferences and the rights they can actually control. Owning a master recording does not necessarily mean owning the composition, and a label, publisher, co-writer, or distributor may hold relevant rights. Contracts and licenses can also grant permissions that affect what a creator may authorize or restrict.

Check ownership and agreements: Review recording, publishing, collaboration, distribution, and work-for-hire terms before offering music for training or making claims about its use.
Keep useful records: Preserve dated source files, release details, registrations where applicable, licenses, and communications about permissions. These records can help establish provenance and ownership.
Read platform terms: Uploading music to a service may grant it specified rights. Look for language about machine learning, sublicensing, retention, and deletion rather than assuming an upload is used only for streaming.
Document concerns about outputs: Save the output, prompt, service and date, and the comparison track. A measured report is more useful than an unsupported accusation.

Creators may also choose to license music for machine-learning uses, decline such licenses, or ask a qualified lawyer about available rights and remedies. Practical control can be limited where a creator has already assigned rights or where local law provides an exception.

What developers should do

For developers, dataset governance is a core part of product design. A defensible process starts before training: identify where each item came from, what rights or terms apply, and whether the intended use fits those terms. “Found online” is not a rights category, and a dataset assembled by a vendor still needs appropriate review.

Maintain a dataset inventory with sources, dates, license information, and any restrictions or removal requests.
Prefer music that is licensed for the intended machine-learning use, created for the project, or clearly in the public domain in the relevant territory. Confirm that a license covers both the recording and composition where needed.
Limit access and retention to what the project needs, and document how copies, backups, and derived data are handled.
Test for memorization and unwanted reproduction, including prompts that request a named artist or a particular song. Add safeguards and a process for reviewing reports.
Explain the system’s data practices accurately to users, and establish a channel for rights holders to raise concerns.

These steps do not guarantee legality or eliminate disputes. They make decisions more informed and auditable. Developers should get jurisdiction-specific legal advice for commercial projects, especially where datasets contain commercial recordings or the model is designed to imitate identifiable artists.

Common misconceptions and practical risks

“If a track is online, it is free to train on.” Availability is not the same as permission. Streaming access, a download, and a license for machine-learning use are different things.

“A model cannot infringe because it stores only numbers.” The technical representation does not decide the legal question. Training may involve copies, and outputs may matter if they reproduce protected expression. Whether a particular representation or use violates a right depends on facts and law.

“Giving credit makes the use lawful.” Attribution may be required by a license or expected as a matter of professional practice, but it does not substitute for permission when permission is required.

There are practical risks beyond a lawsuit: a dataset may have incomplete rights information, a model may unexpectedly echo training material, or a service’s terms may conflict with a user’s intended release. Developers and creators should avoid promising that a model is “copyright-safe” unless that claim is carefully defined and supported.

Frequently Asked Questions

Is training an AI model on copyrighted songs illegal?
Not necessarily in every case, but it is not automatically lawful either. The answer depends on the applicable law, the data source, permissions or exceptions, and how the model and its outputs are used.

Does fair use allow AI training on music?
Fair use is a U.S. legal doctrine assessed through multiple factors, not a blanket AI-training exemption. Whether it applies to a particular music dataset or use is unsettled and fact-specific.

Can I use music from a streaming service to train a model?
Streaming access generally does not itself grant permission to copy tracks for training. Check the service terms and obtain the rights needed for the intended use; do not treat access as a training license.

Can an AI-generated song infringe copyright?
It can raise infringement concerns if it reproduces protected expression, but similarity alone does not resolve the issue. The works, the nature of the resemblance, and the relevant law all matter.

How can musicians opt out of AI training?
Options depend on ownership, contracts, platform policies, and local law. A musician can review licenses and service terms, contact the relevant platform or rights holder, and seek legal advice about available options.

Bottom line

Copyrighted music may be used in AI training in some circumstances, but neither a universal ban nor a universal permission accurately describes the law. Creators should understand who controls each relevant right; developers should track provenance, secure suitable permissions, test outputs, and be transparent about uncertainty. For a specific catalog, model, or dispute, qualified legal advice is more reliable than a general rule.

Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.

More cards in this drawer

The whole Numbers and Powers drawer →