Why AI-Generated Music Deserves the Same Copyright Rules as Everyone Else

In 2019, a jury decided that Katy Perry owed 2.8 million dollars for a sequence of eight descending notes on a minor scale. The notes formed an ostinato, a short repeating line of notes, in her hit, “Dark Horse,” and the rapper Flame claimed that she took them from his song, “Joyful Noise.” Three years later, the Ninth Circuit threw the verdict out in Gray v. Hudson, stating that the pattern was made entirely of commonplace musical elements that no one can own.[1] The court warned that giving this protection to something as simple as that ostinato could essentially create a precedent of handing one artist a monopoly over a basic building block of music itself. Today, those same eight notes of the ostinato can create a piece of music from a simple text prompt. New AI tools like Suno and Udio generate full songs in minutes, trained on more music than any person could hear in a lifetime. These rapid developments have already pulled such companies into federal court; sooner or later, a court will once again have to refine how musical copyright infringement exists in the context of AI. 

When that moment comes, the refinement must include a simple principle: copyright law should not be applied more restrictively to AI-generated music than to human-made music. Courts handling AI music cases should apply the same “thin protection” standard for musical building blocks that they already apply to human songs, because shared musical elements have never been sufficient to prove copying; because AI models learn a genre by absorbing its common patterns; and because any stricter standard invented for AI would inevitably get cited against human artists.[2]

First of all, two songs sounding alike is where a copyright case starts, and it is not at all where the case ends. According to Skidmore v. Led Zeppelin, to prove unlawful copying, a plaintiff must prove two things: first, that the defendant had access to the original song, and second, that the two works are substantially similar in their protected expression.[3] In most human-made copyright cases, access is easy to establish, and rarely where cases are won or lost. Substantial similarity is where the real fight begins, and it has never worked cleanly in music because two songs can share a tempo, a chord progression, and a drum pattern and still sound nothing alike, and a song can also share none of these things yet sound very alike. This is why Gray v. Hudson overturned the “Dark Horse” verdict, finding that the ostinato consisted of commonplace musical elements and that the similarities did not arise from any original combination of them.[4]

The best way to see why Gray got it right is to simply listen. The same descending pattern in the ostinato appears in Bach’s Violin Sonata in F minor, in the Christmas carol “Jolly Old Saint Nicholas,” and in the theme to the original 1954 Godzilla movie, as the musician Adam Neely demonstrates in his breakdown of the case.[5] Nobody thinks a Christmas carol copied Bach, and nobody thinks Godzilla copied Bach or the Christmas carol. You could, of course, make the argument that there were intermediate songs that copied each other that slowly spread this ostinato around, but even then, that shows both how difficult it is to prove that “Dark Horse” copyrighted “Joyful Noise” specifically and how the versatility of this singular ostinato makes it such a fundamental musical building block that it should not be copyrighted.

The same principle governs the larger patterns that define a genre. In William v. Gaye, the “Blurred Lines” case, the court let a jury verdict stand on a collection of common funk elements, and Judge Nguyen’s dissent warned that the majority was effectively letting a musical style, a “groove,” be copyrighted.[6] Two years later, an en banc court in Skidmore v. Led Zeppelin pushed back hard, holding that commonplace musical building blocks should receive only thin protection, meaning a plaintiff has to show near-identical copying rather than a loose family resemblance.[7] Those elements are the vocabulary of a genre, not anything one artist invented, and the past decade of court cases shows courts deciding how much of that vocabulary any one artist can own. 

This is where AI music comes in, as genre resemblance is what generative models produce genre resemblance because that is what they are built to learn. As a quick, high-level look at how these models are trained, a musical model is generally shown an enormous number of songs and trained to predict what comes next. It cannot memorize them all, so it is instead pushed to learn the common characteristics most songs in a genre share, such as which chord progressions appear again and again, how a drum pattern fits inside a beat, and why sad songs often live in minor keys. When it generates, it assembles new music out of these characteristics, so the output, consequently, sounds like the genre, as well as many other songs that also live in this genre.[8] As such, an AI-generated song that sounds like a genre is not evidence of copying any specific work within it. This leads to the conclusion that the law does not, and should not, protect the building blocks that many artists build from, even from artificial intelligence.

However, this is not a complete defense of a model that copies songs exactly. It is still entirely possible that a model memorizes and produces an output that is too similar to a specific recording, even when accounting for the building blocks of a genre. On these occasions, such music should be judged as current copyright law has intended and as courts have done for decades. In Grand Upright Music v. Warner Bros. Records, lifting a piece of a recording without a license was treated as plain and clear infringement.[9] A memorized output would not hold up well in a similar case. The nuance is that AI models internalize a genre through these characteristics and then produce songs based on what they have learned, much like a producer listens to various songs in a genre and is subsequently influenced to write their own songs using that genre's vocabulary. The legal treatment of this process should not change simply because such training happens over the course of a machine learning algorithm rather than a childhood of listening. It is not illegal for an AI music model to listen to a big catalog of music, much like it is not illegal for humans to do so as well.

The strongest argument for treating AI music the same as human music is not about AI at all. Suppose courts ignored this parallel and loosened the similarity standard for AI, treating shared genre elements as evidence of infringement whenever the defendant is artificial intelligence. The problem is that the law does not change depending on who the defendant is or whether the defendant is human when applying its rules. Establishing the precedent that a generic chord loop infringes on copyright when created by artificial intelligence becomes a tool the copyright owner can use against any producer, human or not, and is dangerously close to a single entity monopolizing an entire characteristic of music. Loosening this standard to catch machines makes it more difficult for future producers to make music, as they must consider which musical characteristics have already been copyrighted and cannot be used in their music. 

And even then, this is assuming that such a rule against AI can be contained, which it likely cannot. A realistic defendant is likely not a fully autonomous AI song, but a scenario in which a producer sketched an idea using a generative musical tool and reworked the song from there. Behind the closed doors of a production session, who’s to say what is AI-generated and what is human-created? A court facing that question would have to determine which fractions of the music were made by whom and judge each piece separately, an exercise that is impossible to do practically and accurately. And treating the whole song as AI-generated by using a generative tool just runs into the problem initially described when considering loosening the similarity standard.

The primary administrable rule, even in today’s world with AI musical models, should be the “thin protection” rule. In the time that thin protection has been standard, it has never been an excuse to copy music. It exists so that the law does not hinder artists from making music, without giving too much leeway that would override preexisting precedent and allow artists to completely copy a piece of music. And the moment to settle on this rule is now. It took the better part of a decade to undo the one mistake in William v. Gaye, which chilled the music industry and forced it to exercise greater care when producing new music. In today’s world, as AI becomes increasingly prevalent, it’s important that we allow AI to define its role in music, where it has real potential to provide substantive value. While it’s clear that some rules will be needed to limit the revolutionary power it has, definitive ownership over the building blocks of music should not be one of them. Courts will have one chance to choose the right standard the first time, and that choice will define the future relationship among humans, AI, and the music they make.

Edited by Leena Mehta

Sources

[1] Gray v. Hudson, 28 F.4th 87 (9th Cir. 2022).

[2] Skidmore v. Led Zeppelin, 952 F.3d 1051 (9th Cir. 2020) (en banc).

[3] Skidmore, 952 F.3d 1051.

[4] Gray, 28 F.4th 87.

[5] Adam Neely, “Why the Katy Perry/Flame Lawsuit Makes No Sense,” YouTube video, August 2, 2019, https://youtu.be/0ytoUuO-qvg 

[6] Williams v. Gaye, 895 F.3d 1106 (9th Cir. 2018) (Nguyen, J., dissenting).

[7] Skidmore, 952 F.3d 1051.

[8] U.S. Copyright Office, Copyright and Artificial Intelligence, Part 3: Generative AI Training (Washington, DC: U.S. Copyright Office, 2025), https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-3-Generative-AI-Training-Report-Pre-Publication-Version.pdf 

[9] Grand Upright Music, Ltd. v. Warner Bros. Records Inc., 780 F. Supp. 182 (S.D.N.Y. 1991)