Analyzing the definitive conclusions of Part 3 of the AI Report on whether using copyrighted music for AI model training is protected by the Fair Use doctrine.
Table of Contents
In May 2025, the U.S. Copyright Office (USCO) released the third and final part of its comprehensive report on Artificial Intelligence, focusing squarely on the central legal conflict: **Fair Use**. This report provides the definitive analysis for judges, lawyers, and creators on whether AI companies can use copyrighted works—including music, literature, and art—for model training without permission. [1]
The USCO concluded that AI training is **not categorically Fair Use**. Instead, the legality of the practice must be determined on a fact-specific, case-by-case basis using the traditional four-factor test. Importantly, the USCO emphasized that two factors weigh heavily against AI developers, supporting the arguments of creators.
The report rejected the argument put forward by many AI companies that training models is **"inherently transformative"** simply because it results in a statistical model (the **weights**) rather than a direct copy. [2]
The USCO determined that while the use of copyrighted works to train AI models is unprecedented in scale, the existing legal framework of Fair Use is flexible enough to address it. However, the report makes it clear that the analysis depends on the context, the source of the training data, and the nature of the output. [2]
This factor examines whether the new use is **transformative**—meaning it has a different purpose or character from the original work—and whether the use is commercial.
The USCO concluded that the potential market effect is the **most important element** of the fair use analysis in the AI context, and it strongly favors the arguments of the copyright owners. [1]
The report asserts that AI training **"threatens significant potential harm to the market for or value of copyrighted works."** This threat must be viewed broadly, including effects on lost licensing opportunities and market dilution.
The USCO identified three distinct ways in which unlicensed AI training harms the market for music and other creative works: [1]
| Type of Harm | Description |
|---|---|
| **Lost Licensing Opportunities** | AI developers bypass the need to negotiate licenses, depriving copyright owners of compensation for the use of their property as data. |
| **Lost Sales/Substitution** | AI systems capable of generating works that substitute for the works they were trained on (e.g., an AI song replaces a human song) directly cannibalize the market. |
| **Market Dilution** | The **speed and scale** at which AI generates content (the "AI Slop" discussed in a prior article) poses a serious risk of diluting markets, making it harder for human-authored works to be found and valued. |
The USCO explored potential legislative solutions, such as compulsory licensing (government setting the price) and Extended Collective Licensing (ECL). However, the Office ultimately recommended **allowing the nascent licensing market to evolve organically** without immediate government intervention. [3]
The US Copyright Office's final report on AI training is a major victory for creators and rights holders. By explicitly stating that AI training is not *per se* fair use and underscoring the severe threat of market harm, the Office has provided judges with the analytical tools needed to rule against unlicensed AI training in the currently pending copyright lawsuits. This report confirms that the future of AI music relies on a compensated, licensed ecosystem, not one built on free access to creative works.
