AI Training Data & Copyright
AI training data refers to the datasets used to train machine learning models — in entertainment contexts, these datasets typically include copyrighted text, images, audio recordings, and video. The assembly and use of training datasets raises copyright questions at two stages: (1) whether reproducing copyrighted works into training datasets constitutes infringement, and (2) whether the resulting model outputs infringe the training data. The training data copyright question is the central legal dispute defining the AI industry's relationship with creative rights holders.
Training data is the foundational input that determines an AI system's capabilities — and its legal exposure. Entertainment companies whose content was scraped for training data without permission are the primary plaintiffs in the most significant pending AI copyright litigation. Understanding how training datasets are assembled, what rights are implicated, and what licensing frameworks exist is essential for attorneys advising either side of these disputes.
The exclusive right to reproduce copyrighted works — implicated when training datasets are assembled by copying copyrighted content.
Fair use — the primary defense for training data assembly without license.
Prohibits removal or alteration of copyright management information (CMI) — relevant when AI scraping tools strip metadata, watermarks, or attribution from training data.
Confirms that training involves reproduction of copyrighted works and requires case-by-case fair use analysis — no blanket exemption.
Whether Suno's AI music generator trained on major label recordings without license infringed those recordings' copyrights.
Settled in 2025 for undisclosed terms — establishes real infringement exposure for AI systems trained on commercial recordings without licenses. The settlement value implied significant liability.
Whether Stability AI's training on Getty's licensed image database without permission infringed Getty's copyrights and trademarks.
Claims survived dismissal — case proceeding through discovery. Getty's allegations include both copyright infringement for training data use and trademark infringement for outputs bearing Getty watermarks.
Class action by visual artists alleging AI image generators infringed copyrights by training on artworks scraped from the internet.
Direct infringement claims survived early dismissal — established that artists can state a viable copyright claim against AI image generators for training data use.
The training data dispute has created a new licensing market that did not exist three years ago. Major music labels, book publishers, and news organizations have signed training data licenses with AI companies — establishing a market that strengthens rights holders' fair use arguments by demonstrating that a licensing market exists. Entertainment companies that have not yet evaluated whether their catalogs are being used for AI training are missing both a risk management opportunity and a potential revenue stream.
Advise catalog-owning clients to audit whether their content has been used in AI training datasets — several public tools now allow content searches in known training datasets.
For clients considering AI training license negotiations, the existence of licensing deals by comparable rights holders strengthens market value arguments — research comparable deals before entering negotiations.
Include CMI protection provisions in any content licensing or distribution agreement — stripping copyright management information from training data is independently actionable under § 1202.
For AI company clients assembling training datasets, recommend proactive licensing for commercially sensitive content categories (major label recordings, stock image libraries, major publisher catalogs) where litigation risk is highest.
Monitor whether 'robots.txt' compliance or web scraping terms of service violations create independent legal exposure beyond copyright — several pending cases include computer fraud and breach of contract claims alongside copyright.
Assembling AI training datasets by copying copyrighted works implicates the exclusive reproduction right — fair use is the primary but unresolved defense.
Stripping copyright management information (watermarks, metadata, attribution) from training data creates independent liability under 17 U.S.C. § 1202.
A licensing market for AI training data now exists — which strengthens rights holders' market harm arguments in fair use analysis.
The RIAA v. Suno settlement establishes real commercial consequences for training on commercial recordings without licenses.
Entertainment companies with large content catalogs should both audit AI training data use and evaluate proactive licensing as a revenue opportunity.
Several public tools allow content searches in known datasets (LAION, Common Crawl, etc.). For music, the RIAA and publishers have conducted audits. Legal discovery in pending cases has also surfaced training dataset contents. The most practical approach is a targeted audit of major AI training datasets for client content.
Potentially — beyond copyright, scraping in violation of terms of service may implicate the Computer Fraud and Abuse Act and breach of contract claims. Several pending AI cases include these claims alongside copyright infringement.
Market is still developing. Deals reported range from low six figures for niche datasets to hundreds of millions for major label or publisher catalogs. Valuation factors include dataset size, quality, uniqueness, and the AI company's projected commercial use.