AI Music and Copyright: What the Law Actually Recognizes

In February 2023, the U.S. Copyright Office made a decision that seemed narrow at the time. It refused to register a comic book whose images had been generated by Midjourney, ruling that works produced entirely by artificial intelligence — without human creative control — are not eligible for copyright protection. The decision was about a comic book. But the principle it established reaches far beyond comics.

Music is next. The question of whether AI-generated music can be copyrighted — and if so, by whom — is now working its way through courts, regulatory proceedings, and legislative hearings in the United States, the European Union, and the United Kingdom simultaneously. The decisions being made in the next two to three years will determine who owns the music of the next decade.

The current legal framework was not built for this. Copyright law in the United States requires human authorship. The Copyright Act of 1976 doesn't mention AI because AI, in its current form, didn't exist. Courts have interpreted the human authorship requirement strictly: a photograph taken by a monkey is not copyrightable. A painting generated entirely by a computer program is not copyrightable. The question is where AI-assisted music falls on the spectrum between 'human made this' and 'a machine made this.'

That spectrum is genuinely complicated. When a producer uses Suno or Udio to generate a track, how much creative control did they exercise? If they wrote a detailed prompt — specifying tempo, key, instrumentation, mood, structure, and lyrical themes — does that constitute authorship? The Copyright Office has suggested it might, in some cases. But the threshold is unclear, and the burden of proof falls on the person claiming copyright.

The current legal uncertainty is not neutral. It benefits the platforms and the AI companies, who can operate in the gray zone while the rules are being written.

Studio Talk, Issue 001

The training data question is separate but equally consequential. Several major record labels — Universal Music Group, Sony Music, and Warner Music Group — have filed lawsuits against Suno and Udio alleging that the companies trained their models on copyrighted recordings without permission. The labels are seeking damages and, more importantly, injunctions that could force the AI companies to either license the training data or stop operating.

The AI companies argue that training on copyrighted material constitutes fair use — a doctrine that allows limited use of copyrighted material without permission for purposes such as commentary, criticism, and transformation. The argument is that training a model on music is transformative: the model doesn't reproduce the music, it learns from it. Courts have not yet ruled definitively on this question in the music context, though a series of decisions in the visual art space have gone both ways.

What makes the music cases particularly complex is the nature of what's being copied. A painting is a single work. A musical recording is a layered artifact: there's the composition (the melody and lyrics, protected by one copyright), the sound recording (the specific performance and production, protected by a separate copyright), and the arrangement (which may or may not be separately protectable). AI models trained on recordings are potentially infringing multiple layers of copyright simultaneously.

Independent artists are watching these cases with a particular kind of dread. The major labels have the resources to litigate. They also have the leverage to negotiate licensing deals if the courts rule against the AI companies — deals that would compensate the labels for the use of their catalogs and potentially give them a stake in the AI music ecosystem. Independent artists have neither. If the labels settle, the settlement will likely not include provisions for the thousands of independent musicians whose work was also used to train the models.

The European Union is taking a different approach. The EU AI Act, which came into force in 2024, includes provisions requiring AI companies to disclose what copyrighted material was used to train their models. This transparency requirement doesn't resolve the ownership question, but it creates a paper trail that could be used in future litigation or licensing negotiations. Several EU member states are also considering legislation that would create a right to remuneration for artists whose work is used in AI training — a model similar to the royalties that radio stations pay to musicians.

The United Kingdom has proposed a different solution: an opt-out system that would allow copyright holders to exclude their work from AI training datasets. Critics argue that an opt-out system places the burden on creators rather than on the companies profiting from their work. Supporters argue that it's more practical than requiring affirmative consent for every work in a training dataset that may contain millions of recordings.

None of these frameworks fully addresses the output side of the equation — the question of who owns the music that AI systems produce. The Copyright Office's position is that AI-generated music with minimal human input is not copyrightable. But 'minimal human input' is doing a lot of work in that sentence. A producer who spends hours crafting prompts, selecting from dozens of generated options, editing the results, and arranging the final track has arguably exercised significant creative control. Whether that rises to the level of authorship is a question courts will have to answer.

What's clear is that the current legal uncertainty is not neutral. It benefits the platforms and the AI companies, who can operate in the gray zone while the rules are being written. It disadvantages creators, who can't enforce rights they're not sure they have. And it creates a chilling effect on the kind of creative experimentation that might actually produce interesting results — because artists who invest time and effort in AI-assisted work don't know whether they'll be able to protect what they make.

The music industry has been here before. The introduction of sampling in the 1980s created a similar period of legal uncertainty, during which artists and labels operated without clear rules until a series of court decisions and licensing frameworks gradually established norms. That process took roughly a decade. The AI music question is moving faster — the technology is advancing more quickly, the economic stakes are higher, and the number of affected parties is larger. The law will catch up. The question is what gets destroyed in the meantime.