AI developers need large amounts of text, music, images and video, but access to those materials does not automatically establish a right to train on them. For rights holders, copyright and licensing determine which uses require authorization, which may fall within a legal exception and what an agreement must cover.

The analysis is especially important for music catalogs, where a recording and its underlying composition have different owners. It also requires separating training inputs from model outputs: permission to ingest a work does not necessarily authorize every use of content a model generates.

This guide focuses on U.S. law, with a separate discussion of European rules. It provides a framework for rights and business affairs teams, not a legal opinion on a particular model.

Separate the activities in the AI pipeline

“AI training” can describe several legally distinct activities. A developer might download files, retain a source library, normalize audio, create training copies, fine-tune a model or retrieve material when answering a user’s request. Those activities should not be treated as one undifferentiated use.

Under 17 U.S.C. § 106, copyright owners control specified acts, including reproduction and preparation of derivative works, subject to statutory limitations. Downloading and preprocessing protected material can involve reproduction even if the final product does not display the source file.

By contrast, copyright does not protect facts, ideas or musical style as such. Nor does a model’s exposure to a work automatically establish that its parameters are an infringing copy.

A useful copyright and licensing assessment therefore identifies each act, the protected material involved and the legal basis asserted for that act. Evaluating only the final model can miss problems in how the training corpus was acquired.

Music training requires separate master and publishing analysis

A commercial recording usually embodies at least two copyrighted works: the sound recording and the musical composition. Lyrics generally form part of the musical work. Ownership and licensing authority may differ across those layers.

A label’s authority to license a master does not necessarily extend to the song. A publisher’s authority over the composition does not necessarily authorize copying a particular recording. Co-ownership, administration agreements and territorial limits can further complicate clearance. The distinction between musical works and sound recordings is the starting point for identifying the relevant owners.

Existing licenses must be read rather than assumed to cover training. A mechanical license for making and distributing phonorecords is not a general authorization to train a generative model. A synchronization license likewise does not automatically permit unrelated model development.

For music catalogs, copyright and licensing diligence should connect each proposed training file to the relevant master and composition permissions, unless an applicable exception supplies the legal basis. A dataset vendor’s statement that its audio is “licensed” is insufficient without identifying what was licensed and by whom.

Fair use is a factual defense, not a blanket training exemption

U.S. fair use can permit otherwise restricted uses without the owner’s authorization. Section 107 requires consideration of four factors: the purpose and character of the use, the nature of the work, the amount used and the effect on the work’s potential market.

For model training, relevant questions include:

• Purpose: Is the use analytical, expressive, commercial or directed toward generating substitutes for the source material?

• Nature: Does the corpus contain factual material, highly creative works or unpublished content?

• Amount: Were complete works copied, and how does that amount relate to the asserted purpose?

• Market effect: What evidence shows substitution, harm to relevant markets or other legally cognizable market effects?

Commercial use does not automatically defeat fair use, just as describing training as “transformative” does not automatically establish it. Copying a complete work can sometimes be justified, but the justification depends on the use.

The U.S. Copyright Office’s May 2025 analysis of generative AI training, available among its copyright and artificial intelligence reports, examines these distinctions rather than treating all training alike. Its analysis is influential guidance, not a binding judicial ruling.

For copyright and licensing decisions, fair use should be documented as a use-specific legal position. It should not be reduced to a checkbox stating that the material was publicly accessible.

What the 2025 court decisions actually established

Two June 2025 district-court decisions illustrate why acquisition, training and market evidence require separate analysis.

In Bartz v. Anthropic, the court treated training on the books at issue as fair use but distinguished that activity from acquiring and retaining pirated books in a central library. The decision did not make unlawful sourcing irrelevant simply because some material would later support training.

In Kadrey v. Meta Platforms, the court granted summary judgment to Meta on the authors’ training-related copyright claims before it. The decision emphasized deficiencies in the plaintiffs’ evidence of market harm and did not establish that every model trained on copyrighted books qualifies for fair use.

These were district-court rulings addressing particular records, not universal permission for developers to ingest any catalog. Books, music and images may present different evidence about copying, substitution and commercial markets.

Rights teams should use such decisions to sharpen copyright and licensing questions, not replace factual investigation. For a live dispute or transaction, counsel should verify subsequent rulings, appeals and the controlling law in the relevant jurisdiction.

Verify dataset permissions, not just dataset access

A dataset’s availability tells you little about its chain of rights. Material may be accessible through a subscription, platform API, archive or download link without being authorized for model training.

Start by distinguishing three things: the right to access the source, the right to copy its contents and the right to use those copies for the proposed training activity. A dataset provider may control its selection or arrangement without controlling every underlying work.

Open licenses also require attention to their actual terms. Attribution, noncommercial restrictions or conditions on adaptations may matter. Whether a condition reaches a particular model or output can require legal analysis, so a label such as “Creative Commons” is not a complete clearance result.

For each source, retain the operative terms, acquisition date, permission documents and relevant ownership records. Record whether authorization comes from the owner, an authorized agent, a sublicense or a claimed statutory exception. Contract warranties are useful, but they do not create rights the seller lacks.

Good provenance supports copyright and licensing decisions by making the permission claim testable. The same discipline used in rights metadata and legal data governance helps connect files to owners, contracts and evidence rather than relying on a vendor’s description of the dataset.

European rules require a separate analysis

The European Union has express text and data mining provisions in Directive (EU) 2019/790. Article 3 addresses qualifying scientific research by research organizations and cultural heritage institutions with lawful access. Article 4 provides a broader exception subject to conditions, including lawful access and the absence of an appropriate express reservation of rights.

For content made publicly available online, Article 4 specifically addresses machine-readable rights reservations. A reservation can therefore affect the availability of that exception, but it should not be treated as a universal prohibition enforceable identically in every jurisdiction.

The EU AI Act adds obligations for general-purpose AI model providers. Article 53 establishes copyright-policy and training-content-summary requirements, subject to applicable provisions and implementation timelines. Those requirements are not a substitute for a license or an exception.

A cross-border review should map where relevant copying occurs, which entities perform it and which national laws apply. U.S. fair use and EU text and data mining rules are different legal frameworks, not interchangeable labels.

Negotiate copyright and licensing terms that match the model

An AI training agreement should describe the permitted activity more precisely than “use for artificial intelligence.” The commercial value and risk differ substantially between internal experimentation, fine-tuning an existing model and releasing a general-purpose model to outside users.

A rights holder should be able to identify what is being licensed, who can use it and what happens after training. Broad assurances about responsible AI are not a substitute for operative contract terms.

Define the corpus, training scope and downstream users

Specify the covered works and rights, including catalog identifiers, ownership shares, territories and exclusions. Where the licensor controls only one layer of a music recording, the agreement should say so rather than implying complete clearance.

Separate pretraining from fine-tuning, evaluation, retrieval and other uses. Retrieval can involve storing or presenting source material at inference time, which is different from using a work only during training. Decide whether the license includes affiliates, contractors, model hosts or downstream customers.

Also address whether the developer may distribute model weights, provide an API or sublicense access. Granting access to a hosted service and releasing downloadable weights create different practical limits on monitoring and control.

Precise copyright and licensing language prevents a narrow catalog deal from becoming an unintended grant for unrelated products. Restrictions should address observable conduct where possible, such as prohibited distribution, unauthorized dataset transfers or specified output uses.

A training license also should not be confused with a complete clearance of generated material. The distinction between input permissions and copyright risks in AI outputs remains relevant after the agreement is signed.

Set reporting, payment and exit terms

Pricing depends on the agreement, not a universal statutory rate for AI training. Parties can negotiate fixed fees, usage-based payments, revenue participation or a combination, but each structure needs a measurable payment trigger.

The following questions help turn commercial expectations into contract terms:

Deletion deserves particular attention. Removing a source file is not the same as reversing its influence on an already-trained model. A contract should distinguish deletion of retained copies, exclusion from future runs, restrictions on deployed models and any agreed retraining or unlearning obligations.

For copyright and licensing negotiations, avoid accepting an undefined promise to “remove the content.” Establish what the developer must do, how completion will be verified and what remedy applies if the promised action is not technically achievable.

Allocate responsibility for rights defects and infringement claims as well. Representations, indemnities and liability limits should reflect which party controls sourcing, model development and downstream distribution.

Audit an existing model or investigate suspected training use

A rights holder investigating suspected use needs evidence, not merely a resemblance between its catalog and generated material. A model’s ability to produce work in a familiar genre does not, by itself, prove that a particular recording was copied into its training corpus.

Useful evidence can include dataset manifests, acquisition records, developer statements, archived source pages and reproducible output tests. Each item establishes something different. A manifest may support inclusion in a dataset without proving that a particular model version used it; an output may support a reproduction claim without establishing the full training history.

Preserve original files and document how evidence was collected. Record model versions, prompts, settings and test dates so that output comparisons can be evaluated. Do not bypass access controls or contractual restrictions merely to obtain stronger evidence.

For an internal review, prioritize provenance gaps, works subject to clear contractual exclusions and missing permissions for separately owned rights. For an external claim, distinguish confirmed facts from inferences before making allegations.

A disciplined copyright and licensing investigation should identify the act at issue, the owner entitled to assert the right and the evidence connecting the defendant to that act. Counsel can then assess defenses, registration requirements, available remedies and whether a negotiated license is commercially preferable to litigation.

Frequently asked questions

Does publicly available content count as free training data? No. Public availability does not itself grant permission to copy or train on copyrighted material. A license, fair use or another applicable exception may supply a legal basis, but accessibility alone does not.

Does training always require a license? No. Some uses may qualify for fair use or another statutory exception. The answer depends on the activity, material, jurisdiction and facts. A developer should document the basis it relies on rather than assume that all training is exempt.

Can a record label authorize training on an entire song? Not necessarily. A label may control the sound recording without controlling the composition. The proposed use requires separate analysis of those rights and the label’s contractual authority.

Does a rights reservation prevent AI training everywhere? No. Reservations have specific significance under EU text and data mining rules, but their effect depends on applicable law and implementation. They do not automatically create an identical worldwide prohibition.

Does a training license clear every model output? No. Copyright and licensing permissions for training inputs do not necessarily authorize reproduction of protected expression in outputs. A training agreement can address output controls and responsibility, but its scope must be read alongside the relevant law.

Put the rights position in writing before approving a deal

Before approving a dataset purchase or catalog license, assemble a short decision record: the covered works, relevant owners, proposed activities, jurisdictional analysis and permission or exception relied upon. Attach the supporting contracts and provenance evidence.

Then resolve the terms most likely to become disputed later: downstream access, reporting, payment and the treatment of existing models after termination. A documented, use-specific position gives legal and commercial teams a firmer basis for approving training, negotiating compensation or challenging unauthorized copying.