- The Supreme Court denied certiorari in Thaler v. Perlmutter on 2 March 2026, which settles that a machine cannot be an author under US copyright law and settles nothing else.
- The open question is the one every working musician has: how much human input is enough. Neither the courts nor the Copyright Office has drawn that line in a number.
- The Copyright Office concluded in January 2025 that generative outputs are protectable only where a human author determined sufficient expressive elements, and that the mere provision of prompts is not that.
- Registration requires you to disclose AI generated material and disclaim it, so the practical question is not what you own but what you can describe accurately on a form.
- Your real exposure is evidentiary. If you cannot show what you did, you cannot claim it, and session files are the cheapest insurance in this whole subject.
- Copyright is only one of three risks people merge into it. Training data claims and voice rights are separate, and a distributor's terms can pull a track whatever the law says.
Can you copyright a track you made with AI? The honest answer has two halves, and almost every article you will read only gives you the first one.
The first half is settled. On 2 March 2026 the Supreme Court denied certiorari in Thaler v. Perlmutter, docket 25-449, on the question of whether works output by an AI system without a direct traditional authorial contribution by a natural person can be copyrighted. The denial leaves the D.C. Circuit's ruling standing: US copyright requires a human author. A machine cannot hold the authorship. That is now as fixed as anything in this area gets.
The second half is where you actually live, and it is untouched. Thaler was an unusual case because Dr. Thaler said outright that the work was created autonomously by his system with no traditional human authorship. Almost nobody making music is in that position. You wrote the lyrics and generated the vocal. You composed the part and let a model arrange it. You recorded a real performance and used a separation tool on the stems. None of that was decided, and the amount of human contribution that suffices remains undefined.
Everything below is United States law. Copyright is national, other jurisdictions answer these questions differently, and none of this is legal advice.
What did the courts actually settle?
One narrow thing: a non human cannot be named as the author. That is a statement about who can appear in the author field, not about how much of a work a machine may have touched.
The distinction matters because it is routinely reported as if the courts had banned AI from music. They have not. The ruling that stands says the Copyright Act requires human authorship, which was already the Copyright Office's position when it refused Thaler's application in the first place. The case travelled a long way to confirm a rule rather than to change one.
What did not travel with it is the threshold. There is no percentage, no minimum number of edits, no test you can apply to a session and get a yes. Anybody quoting you a figure invented it.
How much human input is enough?
The most authoritative answer available is a standard rather than a threshold. In Part 2 of its report on copyright and artificial intelligence, released on 29 January 2025, the Copyright Office concluded that outputs of generative AI can be protected only where a human author has determined sufficient expressive elements, and stated that protection does not extend to the mere provision of prompts. It also concluded that existing principles are flexible enough to handle the technology, which is a polite way of saying no new law is coming to rescue you.
Read that standard closely, because it tells you where to put your effort. The question is not how long you spent or how many tools you used. It is whether the expressive choices in the finished work trace back to decisions you made. A prompt describes a target. An arrangement decision is the thing itself.
Iterating on a prompt two hundred times does not convert prompting into authorship. The Office's position is about which elements of the output you determined, not about how much labour you expended getting there. This is the single most common misreading in the guidance people share with each other.
Five production scenarios, and what each one gets you
Here is the same standard applied to the ways people actually make records. The right column is the one to act on, because registration asks you to describe your own contribution and to disclaim what a machine generated.
| How the track was made | Likely registrable? | What you would claim | What you must disclaim |
|---|---|---|---|
| Fully generated from a text prompt, released as it came out | No | Nothing. There is no human authored expression to point at. | The whole work |
| Generated stems, then arranged, edited and mixed by you | Partly, and this is the common case | Your selection, arrangement and edits as a compilation of preexisting material | The generated stems themselves |
| You composed and performed it, AI used for mastering or stem separation | Yes | The composition and the performance, in the ordinary way | Nothing, if the tool processed rather than generated |
| Your lyrics and melody, an AI generated vocal performance | Partly | The lyrics and the musical composition | The generated vocal performance |
| Human performance, AI assisted mixing decisions | Yes | Composition and performance as usual | Nothing, if the mixing was your decision executed by a tool |
Two things about that table are worth naming. The second row is where most people sit and it is the row with the least certainty, because everything depends on whether your arrangement rises to expression a court would recognise. And the third and fifth rows are reassuring in a way that gets lost in the panic: a tool that processes your work is not the same as a system that generates work, and the vast majority of AI in a modern studio is the former.
What do you keep on file?
This is the part that decides real cases, and it is nothing to do with law. If you cannot demonstrate which elements were yours, you cannot claim them, and the moment to assemble that evidence is while you are making the record rather than two years later when somebody disputes it.
Keep the project file with its dates intact, not a bounced master. Keep revision history, because a chain of versions shows decisions being made rather than a single arrival. Keep raw stems and alternate takes, which show what you rejected as well as what you kept. Write arrangement decisions down somewhere dated, even in a text file, because the reasoning behind a choice is the clearest evidence that a person made it. Log which tools and model versions touched the work and at which stage. And archive prompts alongside the outputs they produced, which sounds like evidence against you and is the opposite: it draws a clean boundary between the generated material you will disclaim and the work you did to it.
The last one deserves emphasis. Disclaiming generated material is not a confession, it is a requirement, and an application that quietly omits it is worth less than one that describes the work accurately. An unclear claim is fragile in exactly the situation you bought it for.
Three risks that are not copyright
People fold all of these into one worry, and they have different owners, different timelines and different fixes.
Training data claims. Whether a model was lawfully trained is a dispute between rightsholders and model providers. It is live litigation, it does not depend on what you did, and it can still reach you indirectly if a provider loses and changes its terms or withdraws a model you built a catalogue on. Your protection here is not legal, it is operational: know which tool made what, so a change in one provider's status does not require you to re examine everything you have ever released.
Voice and likeness. Generating a vocal that sounds like a specific identifiable singer raises rights that have nothing to do with copyright in your composition, and they vary sharply between jurisdictions. This is the one where a track can be perfectly registrable and still get you sued.
Platform terms. Independent of law, a distributor or a streaming service can refuse, remove or demonetise a track under its own rules. Those rules are moving quickly and they do not wait for a court. The same pattern played out on social platforms, which we covered when AI posts started losing reach under new platform rules: the terms of service changed faster than any regulation, and the practical consequence arrived first.
Why distribution is the real constraint
The volume numbers explain the platform behaviour better than any policy document. Deezer reported on 21 July 2026 that fully AI generated tracks passed 50% of total new music uploads at peak in June 2026, around 90,000 tracks a day. The same release says that AI generated music accounts for between 1% and 3% of total streams, and that up to 85% of the streams those tracks generated in 2025 were fraudulent.
Sit with the gap between those figures. Half of new uploads, a few percent of listening. That is not a market forming, it is an upload channel being flooded, and platforms respond to floods with filters. Deezer excludes detected AI tracks from algorithmic and editorial recommendation, which means the practical penalty for a fully generated track is not a legal one. It is that nobody hears it.
For a small label, that reframes the whole question. Registrability tells you what you could defend. Discoverability tells you whether there is anything to defend. A track that cannot be recommended has no revenue to protect, and the two problems have the same root: the platform needs to be able to tell what a human contributed. Detection systems are getting better at that quickly, a trend we traced in the piece on AI content detection becoming harder to fool.
Disclosure is arriving from the other direction too
While US copyright asks what you may claim, European transparency rules ask what you must tell people. Under the EU AI Act, Article 50 requires deployers of systems that generate or manipulate audio content to disclose that the material has been artificially created or altered, with those obligations applying from 2 August 2026. There is an exception where the content is part of an evidently artistic, creative, satirical or fictional work, in which case the disclosure has to exist without spoiling the enjoyment of the work.
That exception is doing a lot of work for musicians and it should not be over read. It shapes how you disclose, not whether. If you release in Europe, the safe posture is a factual line in the metadata or the release notes rather than an argument that music is art and therefore exempt.
Who asks you this question first?
Not a court. In practice the first person who makes you answer is a buyer. A sync agent placing a track in an advert, a label taking on a catalogue, a game studio licensing a cue, a distributor processing a release. Every one of them now has a clause about AI generated material, and the clause asks the same thing: warrant what you own and identify what a machine made.
That is why the scenario table above is operational rather than academic. A warranty you cannot support is worse than a narrower one you can. If a track sits in the second row, the correct thing to tell a buyer is that you hold the arrangement and edits and that the underlying stems were generated, which prices the deal accurately instead of collapsing it eighteen months later. Buyers are not looking for purity, they are looking for a clean description they can pass to their own legal team.
The same is true in the other direction with collaborators. If a co writer brought a generated element into a session, that needs to be recorded in the split sheet alongside the percentages, because the split sheet is the document that survives the relationship. A verbal understanding about who used what tool does not.
The catalogue question nobody wants to answer
If you have already released material and have no idea which tracks used what, you have a records problem rather than a legal one, and it is finite. Work backwards from the releases that earn, not from the top of the catalogue. For each one, establish two facts: whether any part of it was generated rather than processed, and whether the session files still exist. That gives you three piles. Tracks you can document. Tracks that used no generative tools at all, which is most older material. And tracks where you genuinely cannot tell.
The third pile is the one to handle deliberately. Do not warrant those to a buyer, do not register them on a guess, and if one starts earning, reconstruct what you can while people who worked on it still remember. Nothing bad happens to a track you never made a claim about. Bad things happen to a claim you cannot support.
Going forward the fix costs nothing. One line in the session notes at the moment of bouncing a master, naming what was generated and what was played. Two years of that habit is worth more than any amount of reading about case law.
What to do before the next release
Three decisions, in this order, and none of them requires a lawyer to start.
Decide per track which of the five rows above it sits in, and write that down at the point of release rather than reconstructing it later. Assemble the evidence file while the session is open. And if you register, describe the work honestly: claim the human authored elements, disclaim the generated ones, and accept that a narrower registration you can defend beats a broad one that collapses under a question.
Then keep a written policy for your own catalogue about which tools may touch which stage of a release. It sounds bureaucratic for a two person operation and it is the thing that saves you when a provider changes its terms or a distributor asks what you used. The same logic we set out for businesses adopting AI in our own AI policy applies to a label: name the data and the stage, not the brand of tool, because the tools change and the stages do not. The same discipline holds anywhere you generate assets you intend to sell, which is why we wrote the equivalent guidance for what a generated product image can honestly do.
One more thing about timing. The Copyright Office has said a third part of its report is still to come, covering the legal implications of training models on copyrighted works and how liability gets allocated. That part will not change the human authorship rule, which is now settled at the highest level available. It may well change which tools you are comfortable using and what your provider is willing to promise you, so treat your tool log as a live document rather than an archive.
The rule underneath all of it is short. You own what you decided. Write down what you decided, and you can prove it.