AI and Music: What the Law Actually Says in 2026

This article explains, in plain terms, where United States law currently stands on artificial intelligence and music. It is informational only. It is not legal advice, and it creates no attorney client relationship.

Three questions sit at the center of the fight, and it helps to keep them separate, because the law treats them very differently. First, can an AI generated song be copyrighted at all. Second, is it lawful to train a music model on copyrighted recordings in the first place. Third, who controls a real artist's voice once a machine can clone it. None of the three has a settled answer in 2026, but each one has a direction, and the directions are worth knowing before you make a decision that costs you.

Copyright and the human authorship requirement

Start with what you can own. The United States Copyright Office has held a steady line for years now. Copyright protects human creativity, so a work generated entirely by a machine, with no meaningful human authorship, cannot be registered. The Office laid this out in detail across its multi report study, Copyright and Artificial Intelligence, including its January 2025 report devoted to copyrightability.

Two decisions put flesh on that rule. In the matter of the graphic novel Zarya of the Dawn, the Office allowed the human author to register the text and the selection and arrangement of the images, but refused protection for the individual pictures a generative tool produced from her prompts. In Thaler v. Perlmutter, the courts upheld the Office's refusal to register an image the applicant himself described as authored by an AI system with no human creative input. Read together, the message is consistent. A machine is not an author, and a prompt is not the same as authorship.

For a working musician the practical rule is easy to state and worth tattooing somewhere you can see it. The more a person shapes the expression, the writing, the melody, the arrangement, the selection and editing of what the tool spits out, the more protectable the result becomes. A track summoned by a single sentence of prompting may carry thin protection, or none at all. Your fingerprints are what the law rewards.

Training data and the fair use fight

The louder battle is not about the songs that come out. It is about the songs that go in. In 2024 the major labels sued the two best known AI music generators. In UMG Recordings, Inc. v. Suno, Inc., No. 24-cv-11611 (D. Mass.), and in UMG Recordings, Inc. v. Uncharted Labs, Inc., the company behind Udio, No. 1:24-cv-04777 (S.D.N.Y.), the plaintiffs allege that the developers copied enormous libraries of copyrighted sound recordings, without a license, to train their systems. Both suits were still in active litigation in 2026, deep in the expensive machinery of discovery, and how they end will shape the entire category.

The developers lean on fair use, the doctrine that allows some unlicensed copying when the new use is transformative enough and does not gut the market for the original. That defense has had a rough start in the AI setting. In Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., 765 F. Supp. 3d 382 (D. Del. 2025), the court held that copying protected material to train an AI legal research tool was not fair use, and it leaned hard on the fourth fair use factor, the harm to the market for the originals. That case involved a tool that did not generate new expression, so its reach has limits, but it handed rights holders a blueprint they are now using everywhere.

The picture is genuinely mixed, which is why anyone who tells you the answer with confidence is selling something. In the Anthropic litigation over training large language models on books, a California federal court found that training on lawfully acquired copies could be transformative fair use, while separately holding the company liable for building a library out of pirated copies. Anthropic later agreed to a settlement reported at roughly 1.5 billion dollars. Related suits against image generators, including Andersen v. Stability AI and the Getty Images litigation, are grinding forward on similar questions. Every one of those cases involves text or images, not music, and the music cases will rise or fall on their own records. Read them as weather moving across the field, not as a final score for the recording industry.

Your voice, your face, your likeness

Now the third question, and the one most people get wrong. Copyright protects songs. On its own it does not protect the sound of your voice. That gap turned concrete in 2023, when an anonymous track built on AI imitations of two famous singers raced across every platform and pulled real numbers before it was taken down. The song infringed no copyright in the ordinary sense. It just wore two people's voices without asking.

The gap that fills that hole is the right of publicity, and it lives mostly at the state level, which means it changes when you cross a state line. Tennessee moved first and hardest with the ELVIS Act, effective July 1, 2024, which widened the state right of publicity to cover a person's voice specifically, and reached not just those who misuse a voice but those who make or distribute the tools whose primary purpose is unauthorized vocal cloning. At the federal level, the proposed NO FAKES Act would create a nationwide right against unauthorized AI replicas of a person's voice and likeness. As of 2026 it remained a bill, not a law, which means for now your protection depends heavily on which state you and the infringer sit in.

Where this leaves an independent artist in 2026

Put the three answers together and a careful posture appears. On ownership, keep your hands all over the work, because human authorship is still the line between owning your song and owning nothing. On training, understand that the legality of feeding copyrighted catalogs into a model is unresolved and moving under the courts month by month, so building a business on the assumption that it is clearly legal, or clearly illegal, is a gamble either way. On your voice, know that protection is arriving one state at a time, with a federal shield still waiting on a vote, so a cloned vocal that is actionable in Tennessee may be far harder to fight somewhere else.

One last time, because it matters. This is general information, not legal advice, and this corner of the law is rewriting itself constantly. Before you sign a contract, release a track, or send a threatening letter, talk to a licensed attorney about your specific facts.