1. Copyright Claims Based on AI Training Data
A training data claim identifies protected works and specific copying. Federal copyright law protects original expression in articles and books, but not facts, ideas, or writing styles. Section 106 grants exclusive rights, including reproduction, subject to limitations such as fair use.
Ownership and License Scope
Publishers must establish their right to enforce the copyright. Author agreements, assignments, and exclusive licenses determine which rights they hold. Publication alone does not establish ownership.
Permission to display excerpts or distribute ebooks may not extend to model training. A copyright licensing review examines the agreement’s permitted uses and restrictions. Whether exceeding those terms supports infringement, breach of contract, or both depends on the provision involved.
Different Uses Require Different Analysis
| Challenged Use | Central Question |
|---|---|
| Downloading and storage | How did the developer obtain and retain copies? |
| Model training | Was training-related copying authorized or fair use? |
| Generated outputs | Does the response reproduce protected expression? |
Downloading and storage
- Central QuestionHow did the developer obtain and retain copies?
Model training
- Central QuestionWas training-related copying authorized or fair use?
Generated outputs
- Central QuestionDoes the response reproduce protected expression?
A training-copying claim does not necessarily require an infringing output. An output resembling an article also does not, by itself, establish the source of the training data.
Claims involving removed copyright management information under §1202 require separate proof of the prohibited conduct and statutory knowledge requirements. Missing attribution alone is insufficient.
2. Fair Use Defenses and Market Harm
Section 107 requires courts to weigh four factors together. Commercial use, copying an entire work, or asserting a new technological purpose does not decide the defense alone.
The Four Factors
Courts consider the purpose and character of the use, the nature of the copyrighted work, the amount used, and the effect on its potential market or value.
Developers may argue that training uses text for a different function. Publishers may argue that AI products replace their content or impair licensing opportunities. The analysis must connect those arguments to the particular works and uses.
Lost subscriptions, substitution for original content, and licensing evidence may matter. A traffic decline alone does not establish that the challenged copying caused legally relevant market harm.
Limits of Earlier Rulings
The June 2025 fair use ruling in Bartz v. Anthropic distinguished training use from acquiring and retaining pirated books. Its treatment of training did not excuse those separate acquisitions.
The June 2025 ruling in Kadrey v. Meta favored Meta on the training claim based on the evidence presented, including weaknesses in the plaintiffs’ market-harm showing.
Neither district court ruling establishes a nationwide exemption for AI training. Their reasoning highlights the importance of the challenged conduct and the evidentiary record.
3. Evidence Publishers Need
A useful claim file connects identifiable works to the defendant’s alleged use. General assertions about internet scraping leave ownership, copying, and permission unresolved.
Records That Address the Dispute
Publishers should preserve:
- Original works, publication dates, and registration records.
- Ownership agreements and third-party licenses.
- Available dataset references and crawl records.
- Complete prompts, outputs, dates, and model versions.
- Customer and licensing records relevant to claimed harm.
Testing context matters. A response produced after a user supplies the article differs from one produced without that text. The complete interaction helps show what the test actually demonstrates.
A dataset reference may support further investigation without proving that the defendant trained on a particular work. Discovery can seek source records, training documentation, and testimony to address that gap.
4. Federal Litigation and Available Relief
Copyright Act infringement claims belong in federal court. Related contract claims require separate analysis of the agreement and applicable law.
Registration and Case Development
For U.S. .orks, §411 generally requires registration or refusal before filing an infringement action, subject to exceptions. An application alone ordinarily does not satisfy that prerequisite. Copyright Office filing records also help establish registration timing and coverage.
A complaint identifies the works and challenged acts. A motion to dismiss tests legal sufficiency; surviving it does not prove infringement. Discovery develops the record, and summary judgment may resolve issues without trial. A settlement does not establish a binding fair use rule.
Remedies and Their Conditions
Potential relief includes actual damages and attributable profits, or statutory damages when available. Section 412 can restrict statutory damages and attorney’s fees based on registration timing.
Injunctions and fee awards require separate analysis. The developer’s total revenue is not automatically recoverable as infringement-related profits.
5. Frequently Asked Questions
Public access and later changes to crawling practices raise additional questions.
Free online access does not place an article in the public domain or automatically authorize training. Protected expression, permissions, copying, and fair use still require analysis.
A crawler restriction does not establish what happened before the restriction or resolve earlier copying claims. Historical access and dataset records remain relevant.
6. Reviewing a Potential Training Data Dispute
When specific works and credible copying evidence emerge, an attorney can assess ownership, permission, preservation needs, and potential defenses.
A focused AI copyright litigation review connects those materials to the alleged acquisition, training, or output use. It should identify both supporting evidence and questions that require further investigation.
13 Oct, 2026
