The Invisible Heist: How AI Is Rewriting the Rules of Music Copyright—and Who Gets Left Behind
In the summer of 2023, a track called “Heart on My Sleeve” briefly took the internet by storm. It featured what sounded unmistakably like Drake and The Weeknd—the vocal inflections, the production style, the emotional cadence—yet neither artist had sung a single note. A creator using the pseudonym Ghostwriter had deployed AI tools trained on existing recordings to conjure a convincing simulacrum of two of the world’s most recognizable musicians. Universal Music Group moved swiftly to have it removed from streaming platforms, and the Recording Industry Association of America (RIAA) called it a clarifying moment. But what law, exactly, had been broken? The honest answer was: nobody was entirely sure.
That uncertainty—legal, ethical, and commercial—sits at the heart of one of the most consequential intellectual property battles of the digital age. As AI music generators like Suno, Udio, and Google’s MusicLM grow increasingly sophisticated, the music industry faces a fundamental question it has never had to answer before: when a machine learns to make music by consuming millions of copyrighted songs without permission, and then generates new songs that compete in the same marketplace, what does copyright law actually say? And perhaps more importantly, what should it say?
Copyright Law 101: What the Existing Framework Actually Protects
To understand where AI fractures music copyright, it helps to first understand what copyright law was built to do. In the United States, music copyright has long operated along two distinct but parallel tracks. The first protects the musical composition—the underlying melody, harmony, and lyrics—which is typically owned by songwriters and publishers. The second protects the sound recording itself, often called the “master,” which is usually owned by the performing artist or their record label. When you stream a song, royalties theoretically flow through both channels: one to the songwriters, one to the label.
This dual-track system is already complicated. When Pharrell Williams and Robin Thicke lost their 2015 lawsuit over “Blurred Lines”—a jury found they had copied the feel of Marvin Gaye’s “Got to Give It Up”—legal experts debated whether the verdict had effectively extended copyright protection to musical style and genre, which has historically never been protectable. Style, mood, and genre are considered building blocks of creative expression, free for any human artist to absorb, imitate, and reinvent. It is precisely this principle—that learning from prior art is not the same as copying it—that AI companies are now invoking to defend their training practices.
The critical distinction that copyright law draws is between ideas (not protectable) and expression (protectable). A blues chord progression is an idea. Robert Johnson’s 1936 recording of “Cross Road Blues” is a specific expression. The theory underlying most AI-generated music is that a model trained on millions of recordings absorbs patterns, structures, and styles—ideas—without directly copying protected expression. Whether that argument holds up in court remains one of the most contested legal questions of our time.
The Training Data Problem: Did Anyone Ask Permission?
The litigation that may ultimately define AI’s relationship with music copyright landed in June 2024, when the RIAA filed lawsuits against both Suno and Udio in federal court, alleging “massive copyright infringement.” The suits claimed that both companies had trained their AI models on vast libraries of copyrighted sound recordings without obtaining licenses, and that the resulting tools could generate outputs “indistinguishable in quality from those created by human artists.” The potential statutory damages cited in the filings were staggering—up to $150,000 per infringed work, with the RIAA alleging infringement of “thousands of copyrighted recordings.”
Suno and Udio both acknowledged using copyrighted material in training but advanced a fair use defense—the same legal doctrine that allows a critic to quote a book, a scholar to reproduce a painting, or a parody artist to riff on a pop song. Fair use in American law is assessed on four factors: the purpose and character of the use (commercial vs. educational), the nature of the copyrighted work, the amount of the original used, and the effect on the potential market for the original work. The companies argued their use was “transformative”—that ingesting songs to extract statistical patterns is fundamentally different from reproducing those songs.
This argument has some precedent. In Authors Guild v. Google (2015), a federal appeals court ruled that Google’s scanning of millions of books to build a searchable database was fair use because it was sufficiently transformative and did not substitute for the originals in the marketplace. But music industry attorneys point to a crucial difference: Google’s book-scanning did not produce a product that competed directly with the books it had scanned. AI music generators arguably do compete directly with the human musicians whose work trained them. When Suno generates a pop track with a melancholic female vocalist over synth-driven production, it is competing for the same listener attention—and the same Spotify placement—as the actual artists whose recordings it learned from.
In late 2024, both Suno and Udio settled with the major labels for undisclosed amounts, which means no court has yet definitively ruled on whether AI music training constitutes copyright infringement. The settlement preserved the legal ambiguity—and the fight—for another day.
The Output Question: Can AI-Generated Music Be Copyrighted?
The liability question runs in two directions. Labels worry that AI companies are stealing their copyrighted content. But a separate and equally thorny question concerns who, if anyone, owns what AI generates.
The U.S. Copyright Office has been remarkably consistent on this point: copyright requires human authorship. In a series of rulings beginning in 2023, the Office rejected copyright applications for AI-generated images and refused to extend protection to the AI-generated portions of a graphic novel (while protecting the human-written text). In its March 2023 guidance, the Office stated that it would not register works “produced by a machine or mere mechanical process that operates randomly or automatically without any creative input or intervention from a human author.”
For music, this creates a peculiar commercial paradox. If a human producer uses an AI tool to generate a beat, tweaks it, adds lyrics, and arranges a final track, the human contributions may be copyrightable while the AI-generated underlying music may not be. This is not merely a theoretical concern—it directly affects the business models of every company building AI music tools. If the outputs cannot be copyrighted, they cannot be exclusively licensed, which fundamentally undermines the commercial value proposition for artists and producers who might otherwise embrace these tools.
Some legal scholars argue the Copyright Office’s position is too rigid. Professor Ryan Vacca of the University of New Hampshire School of Law has argued that the law should recognize gradations of human-AI collaboration, since a musician who carefully prompts, selects, and edits AI outputs is exercising genuine creative judgment. Others warn that loosening the human authorship requirement could flood the copyright system with machine-generated works, making it harder for human creators to protect genuinely original work.
Meanwhile, AI music generators are themselves generating royalty-bearing content on platforms like Spotify and Apple Music. Suno-generated tracks are streamed. Money changes hands. But in the absence of clear copyright protection, the legal status of those earnings—and who is entitled to them—remains murky.
The Licensing Landscape: Early Deals and the Road Not Taken
Not everyone has chosen litigation. A small but growing number of companies have pursued licensing agreements with rights holders, attempting to build AI music tools on a foundation of explicit permission rather than contested fair use.
Stability AI signed a licensing deal with the stock music platform Pond5 in 2023. Boomy, a platform that lets users generate original songs with minimal input, has focused on royalty-free generation and has distributed over 14 million “created” tracks to streaming services. Most significantly, YouTube announced in 2023 that it had reached licensing agreements with Universal Music Group, Sony Music, and Warner Music Group for an AI music tool called “Dream Track,” which allowed a limited group of creators to generate short clips in the style of participating artists including John Legend, Charlie Puth, and Demi Lovato—artists who had explicitly consented to having their vocal style used.
The YouTube model is instructive because it threads several needles at once: it requires artist consent, it compensates rights holders, and it creates a contained, platform-controlled environment. It is, in other words, closer to a traditional licensing arrangement than to the “ask forgiveness later” approach taken by companies like Suno.
The music industry has licensed its catalog before in situations that initially seemed threatening. The ringtone market, digital downloads, and streaming services all required new licensing frameworks, and all were eventually negotiated—sometimes painfully, sometimes only after litigation. The difference, many argue, is scale and substitution. Streaming did not make it possible to generate a new song in the style of Taylor Swift with a text prompt. AI music generation does.
Some major labels have begun to explore AI revenue as an opportunity rather than purely a threat. Universal Music Group CEO Lucian Grainge announced in 2024 that the company was developing AI tools in partnership with technology companies—tools that would “protect and compensate” human artists. The implicit message was clear: the labels intend to be inside the AI music tent, not outside it.
The Artists’ Perspective: Between Erasure and Opportunity
For working musicians, the AI copyright debate is not abstract. It is a question about whether the market for their skills will exist in five years.
The concerns are particularly acute for session musicians, composers of library music, and producers who work in genre-adjacent commercial music—exactly the type of work AI tools currently do best. A 2023 study by the music industry research firm MIDiA Research found that 27% of professional musicians reported losing income to AI-generated music, a figure that was expected to rise sharply. Session vocalist Holly Herndon—herself an artist who has pioneered the use of AI in music—has argued that the problem is not AI per se but the absence of consent and compensation frameworks. In 2023, she launched “Holly+” a tool that allows others to use her voice in exchange for a share of any commercial earnings, a model she has called “AI for artists, by artists.”
On the other side of the argument, some artists see genuine creative potential. producers like Timbaland and will.i.am have been enthusiastic early adopters of AI composition tools, and a new generation of musicians who grew up with digital audio workstations see AI as another instrument in the toolkit. The Screen Actors Guild-American Federation of Television and Radio Artists (SAG-AFTRA) reached a groundbreaking deal with an AI voice company in 2023 that included informed consent provisions and payment structures for voice replicas—a template that music unions like the American Federation of Musicians (AFM) are watching closely.
The AFM has been less accommodating. In an open letter to major technology companies in 2023, signed by more than 200 prominent musicians including Nicki Minaj, Elvis Costello, and Billie Eilish, artists demanded that AI developers “cease the use of our work in AI training without consent, credit, or compensation.” The letter did not carry the force of law, but it carried enormous public weight—and it signaled that the creative community would not quietly accept a fait accompli.
What Comes Next: Courts, Congress, and the Coming Reckoning
The legal landscape is shifting fast, but it is shifting unevenly. American courts have not yet produced a definitive ruling on AI music training and fair use. The RIAA lawsuits against Suno and Udio settled before a judge could rule on the merits, leaving the central question—is training on copyrighted music without a license legal?—formally unanswered. Parallel cases in text and image AI will likely provide some guidance; a ruling against AI image generators like Stability AI could create precedent that applies directly to music.
In Congress, the NO FAKES Act (Nurture Originals, Foster Art, and Keep Entertainment Safe)—legislation that would create a federal right against unauthorized digital replicas of a person’s voice or likeness—has gained bipartisan support but has not yet passed as of this writing. The Music Modernization Act of 2018, which overhauled the licensing system for streaming, demonstrates that Congress can act when the industry presents a reasonably unified front. Whether labels, publishers, artists, and technology companies can agree on a framework remains uncertain.
Internationally, the picture is equally fragmented. The European Union’s AI Act, which entered into force in 2024, requires AI companies to publish “sufficiently detailed summaries” of copyrighted content used in training—a transparency mandate that stops short of requiring licenses but creates the conditions for rights holders to identify and potentially challenge unauthorized use. The UK has been debating an exception to copyright for AI training that would allow broad use of copyrighted material; the proposal has been fiercely contested by the creative industries.
What seems increasingly clear is that the music industry’s current copyright framework—designed for a world where copying required active human effort—is inadequate for a world where a machine can consume a billion songs, internalize their patterns, and produce competitive new work at scale. The question is not whether the rules will change, but whether those changes will be driven by litigation, legislation, negotiation, or some combination of all three.
The artists who have built careers on the assumption that their creative labor has lasting legal and commercial value are right to be alarmed. So are the technologists who believe that overly restrictive copyright regimes could stifle genuinely beneficial tools. Both sets of concerns are legitimate, and both deserve legal frameworks that take them seriously. What the music industry learned from Napster—eventually, painfully, and after enormous collateral damage—is that ignoring technological disruption is never a winning strategy. The question is whether this time, the reckoning can come before the damage is done.