Part 3. How is copyright law being applied to AI?

Four Considerations for Copyright, Music, and AI

Part 3 of this primer draws primarily from a legal sidebar entitled “Generative Artificial Intelligence and Copyright Law” that was prepared by the Congressional Research Service, a nonpartisan research institute under the Library of Congress. This sidebar provides a comprehensive overview of the current state of AI and copyright in the U.S. It was first published in February 2023 and last updated in July 2025. I will be summarizing the document below along with additional context.

Consideration 1: Authorship

Although the Constitution and Copyright Act do not define the term “author,” current standards in U.S. courts and the Copyright Office do not recognize nonhuman authors, a norm that has been upheld in court in recent years.

Consideration 3: Does the Training of AI Systems Infringe on the Copyrights of Other Works?

AI companies have been the subject of lawsuits by rightsholders because they routinely make unauthorized copies of copyrighted works during the training process.

Many AI companies claim that the doctrine of fair use allows them to make unauthorized copies of copyrighted works. Fair use is dictated by four factors:

1. Purpose and character of use (e.g., commercial, nonprofit, educational).

   1. “Transformative use” is a critical part of fair use; this considers how the new work differs from the original, e.g., a parody or critique/commentary/analysis of the original work.

2. Nature of copyrighted work (e.g., creative work vs. factual/public information).

3. How much of the work is used and how substantial that part of the work is.

4. Effect of use on market or value of original work.

Bartz v. Anthropic PBC and Kadrey v. Meta Platforms, Inc. are two lawsuits related to books and AI models in the U.S. that illustrate the legal complexity of these issues. Bartz v. Anthropic PBC resulted in a split summary judgment. The court determined that Anthropic’s use of the books for training was transformative and fair use but downloading and retaining pirated books in a central library was not. Anthropic ultimately settled this issue for $1.5 billion.

The court also found in Kadrey v. Meta Platforms, Inc. that Meta’s use of the books was transformative and fair use based on the record presented. However, the court stated that “it seems likely that market dilution will often cause plaintiffs to decisively win the fourth [fair use] factor” and that companies “will generally need to pay copyright holders for the right to use their materials.” The plaintiffs’ failure to provide evidence of market dilution because of AI in this case ultimately resulted in the ruling in Meta’s favor.

Consideration 4: Does the Output of AI Systems Infringe on the Copyrights of Other Works?

It can be difficult for rightsholders to litigate if they determine that the output from an AI system resembles their work. Plaintiffs must prove that:

• The AI system had access to their work.

  • AI companies do not usually disclose their training data, which makes this task difficult. Recent journalism has shed some light on music training data, including data used by Suno and large music datasets used by developers.

• The AI system created “substantially similar” output.

  • This is not a well-defined test, which makes it difficult to prove in court.

A separate issue relates to generating works in the style of another artist. For example, Suno and Udio do not allow the use of artist names as part of a written prompt. The issue at hand here is not so much a violation of copyright law but more so the potential of violating Right of Publicity laws if a system generates a voice that is similar to an artist’s voice. Right of Publicity, or likeness laws, govern the use of someone’s name, image, likeness, and voice. These are separate from music copyright laws and are governed at the state level. At the federal level, work on a bill to address federal Right of Publicity laws related to AI has been ongoing since 2023 with the NO FAKES Act.

Detecting and Tracing AI-Generated Music

Given that rightsholders may want to track potentially infringing output, there is industry interest in AI-detection tools. The three major labels have either partnered with, invested in, or acquired companies that manage the use of intellectual property in AI (detecting infringing use, managing licensing, etc.).

One potential method of detecting AI-generated content is for the platform itself to build watermarking into the generation pipeline. Google has built-in watermarks for its generative AI products, called SynthID, including audio. However, audio watermarks have limitations that enable removal.

Another approach is to use new services that have been built to detect AI-generated music and/or to trace AI-generated output back to the original copyrighted training data. For example, Deezer, a French streaming service, claims to detect “100% AI-generated music with 99.8% accuracy”. However, its researchers outline ongoing challenges with AI-generated music detection in a 2025 paper:

• Commercial detection services are closed-source and lack published research, which makes it difficult to scrutinize their accuracy.

• Accuracy numbers can be misleading because model performance degrades when:

  • AI-generated audio is manipulated (e.g., pitch-shifted).

  • The model is exposed to output from models that were not part of its training data.

Part 3: Wrapping Up

Under current U.S. copyright law, human authorship is necessary for copyright protections. Whether or not the training and output phases of AI infringe on copyright is an open, evolving question. Although there is interest in systems to detect AI-generated music and track infringement, the reliability and scalability of such systems have yet to be proven.