Who Owns an AI Song With Multiple Collaborators

Last spring, three of us finished a song in a weekend. I wrote the lyrics, a friend ran the generations on her Suno Pro account, and a drummer we both know replaced the programmed kit with a real one in my studio. When a small sync library, the kind that licenses music for ads and film, asked who owned the track, we gave three different answers, and only one of them was close to right.

Here’s what I learned, and what the law and the platforms actually say as of October 2026. The AI-generated layer of the song belongs to nobody under US copyright law. The human layers belong to whoever made them, and if two or more of you added protectable work intending to merge it into one song, you’re joint authors of that human layer and, without a written agreement, you share it equally. The commercial permission from the platform goes to one person only: the account holder who generated and downloaded the output. So who owns an AI song with multiple collaborators is never one answer. It’s three overlapping answers, and the only way to make them line up is a signed agreement before the first upload.

Copyright Ownership and Platform Permission Are Not the Same Thing

The confusion starts with the word ownership, because the two systems that actually decide the question each mean something different by it.

Copyright is the first system. In the United States, copyright only attaches to work created by a human being. On March 2, 2026 the Supreme Court declined to hear Thaler v. Perlmutter, which left standing the D.C. Circuit’s ruling that a work generated autonomously by an AI system has no author and no copyright. As Finnegan’s note on the denial points out, that isn’t a ruling on the merits, but the appellate decision is now the controlling authority. The Copyright Office had already said the same in its January 2025 report on copyrightability: material whose expressive elements were determined by a machine is not protected, and prompts alone don’t give the person typing them enough control to be the author of what comes out.

The second system is contract. When you pay Suno for Pro or Premier, its terms assign you whatever right, title and interest Suno has in the output you generate, and they grant you permission to use it commercially under conditions. In the same paragraph, Suno says plainly that because of how machine learning works it makes no promise that any copyright exists in that output at all. So the platform can hand you a permission slip. It can’t hand you a copyright that the law never created.

Why does this matter for a group? Because the permission slip is written to one person. Your cowriter who typed half the prompts from your login is not a party to that contract. Your singer who never opened Suno has no relationship with Suno at all. Everyone’s copyright position depends on what they personally created, and everyone’s contractual position depends on whose account did the generating. Those two maps rarely overlap by accident.

What the Prompter Actually Owns

I’ll say this gently, because I know how much craft can go into a good prompt. The person who wrote the prompt owns the prompt. They do not own the song the model produced from it.

The Copyright Office’s reasoning in the Part 2 report is that a prompt is an instruction that conveys ideas, and ideas aren’t protected. The model fills in gaps, ignores parts of what you asked for, and produces different results from identical inputs. The Office concluded that this isn’t enough human control to make the output yours, however many times you regenerated. The report leaves open that far more precise tools could change that analysis; today’s generators aren’t there.

There are three ways a human does get copyright in a track that started as AI output, and all three show up constantly in collaborations. The first is contributing your own expressive material that remains perceptible in the result, such as lyrics you wrote and fed in, or a melody you sang into the upload box. The second is modifying the output in a way that rises to creative expression, such as rewriting the bridge in a DAW, the software you record and edit in, re-tracking the vocal, or re-harmonizing a section. The third is the thinnest: selecting and arranging AI-generated pieces in a sufficiently creative way, the same theory the Office used when it registered the comic book Zarya of the Dawn for its arrangement but not for its images.

That third path is where prompters usually try to land, and I want to be clear about how thin it is. A registration for selection and arrangement protects your particular sequence of generated sections. It does not stop anyone else from using the sections themselves. If your contribution to the group project was choosing the best of forty generations and ordering them, you may have a slim claim to that ordering and nothing more.

When a Collaborator Becomes a Joint Author

US law defines a joint work in section 101 of the Copyright Act as a work prepared by two or more authors who intend their contributions to be merged into inseparable or interdependent parts of a single whole. Courts have layered two tests on top of that definition, and the Ninth Circuit’s model jury instruction on joint authors lays them out clearly. Each author must have contributed material that could have been independently copyrighted, and each must have intended, at the time, to be a co-author rather than a helper.

The first test knocks out ideas, suggestions, and prompts: in Childress v. Taylor, an actress who fed research and story ideas to a playwright was not a joint author, because ideas aren’t expression. The second knocks out real contributors who lacked shared intent: in Aalmuhammed v. Lee, a consultant who wrote actual scenes for the film Malcolm X still lost, because Spike Lee controlled the work and nobody treated him as an author. Control and credit are the evidence courts look at.

Map that onto a typical AI collaboration and the roles sort themselves quickly. The lyricist who wrote the words is contributing copyrightable expression, so they can be a joint author of the composition. The singer who replaces the AI vocal with a real performance is an author of the new sound recording and, if they rewrote the melody, of the composition too. The drummer who plays a real kit over the track is an author of the recording. The mix engineer paid a flat fee to balance the stems, meaning the separated instrument tracks, is usually not an author of anything. The prompter, as we’ve covered, is contributing instructions.

Now the part that surprises people. Once two or more of you qualify as joint authors, the default under US law is that you own equal, undivided shares of the whole work regardless of how much each person did. As the Lexology primer on song ownership puts it, if one writer did ninety percent and the other did ten, without a written agreement they each own half. Any joint owner can license the song nonexclusively without asking the others, but as the collaborator guide from the Law Offices of Edward R. Hearn explains, they must account to the others for the money. Nobody can grant an exclusive license without every co-owner signing. That default is why split sheets exist, and it applies to the human layer of an AI song exactly as to one written at a piano.

Who Suno Actually Gives Rights To

Suno is the platform most collaborative projects run through in 2026, because it’s the only major generator that currently lets you take a file off the service. Its terms changed on September 3, 2026, and the changes matter more for groups than for solo users.

The assignment of Suno’s rights goes to the paid subscriber, for output generated from that subscriber’s own submissions. It’s perpetual, it survives cancellation, and it’s unaffected by later changes to pricing or download allowances. Your submissions, meaning the lyrics, audio, and voice you upload, stay yours. Suno takes a broad, perpetual license to them, but no ownership.

Then comes the condition that didn’t exist before September: commercial use is now tied to a permitted download. You may only exploit output commercially if you obtained an authorized download of that track within your tier’s allowance, and Suno says it may fingerprint or watermark files to show which tier they came from and whether the download was permitted. Pro includes twenty song downloads a month and Premier sixty, and both plans let you buy more. Undetectr’s pricing breakdown reports extra downloads at $2.99 each. The free plan has no monthly downloads at all, and free output is restricted to personal, non-commercial use.

The prices below are Suno’s US list prices from its pricing page. Annual billing lowers Pro and Premier to $8 and $24 a month, and prices vary by region.

Two more clauses in those terms are written specifically for collaborators, and both are traps. If you turn on the feature that lets other users remix your output, every remix is deemed a joint work owned equally by you and the remixer, and every remix is locked to personal, non-commercial use regardless of plan or download status. So the easy in-app way to collaborate on Suno produces a track nobody can release. The other trap is the voice model. Suno’s terms say you may only create a voice model of your own voice, so your singer can’t upload their vocal into your account as a persona. They need their own paid account if they want the assignment to cover their generations.

Udio is the other name people ask about, and my advice for group projects is simple: don’t build on it right now. Udio switched off downloads on October 29, 2025 as part of its settlement with Universal Music Group, reopened them for 48 hours in early November, and has kept them off since. Digital Music News reported that songs pulled during that window kept Udio’s older terms; everything since sits under transition terms limited to personal, non-commercial use. The AI Tools Directory review that checked Udio’s pricing and help center on October 4, 2026 found the licensed platform still hadn’t launched. Standard costs $10 a month and Pro $30, but neither plan currently gives you a file, and a song you can’t export is a song you can’t co-own in any useful sense.

Four People, One Song, Layer by Layer

Let me run the weekend song through all of this.

Maya generated the track on her Suno Pro account from my lyrics, picked the best of thirty takes, and downloaded it within her monthly allowance. Under copyright, Maya owns a thin claim at most, in her selection and arrangement. Under contract, Maya holds Suno’s assignment and the commercial permission for that download. Nobody else does.

I wrote the lyrics in a notebook before Maya opened Suno. The lyrics are my copyright, full stop, and they remain my submission under Suno’s terms. Because I intended them to merge into one song and so did Maya, I’m a joint author of whatever human-authored composition exists. If Maya’s arrangement claim holds up, we co-own that composition in equal halves by default.

Dev re-tracked the drums in my studio, and we kept the AI bass, keys, and lead vocal. Dev is an author of the new sound recording that includes his performance. Suno’s output inside that recording is still unprotected AI material, so the recording copyright covers Dev’s playing and our production choices, not the generated stems underneath.

Priya mixed it for $300 and no points, meaning no percentage of the master. She signed nothing. A mix can involve creative choices, and I’ve watched engineers argue authorship after a track blew up. We fixed that afterward with a single page work-for-hire letter, which is what I should have done before she opened the session.

So who owns the song? Three of us co-own a human layer, one of us holds the platform permission, and nobody owns the parts most listeners would call the song. If Maya cancels her subscription she keeps the permission for that download, but if she falls out with the rest of us, she’s the only one who can lawfully distribute the Suno audio. That’s the answer the sync library needed, and it’s why we wrote everything down the following week.

The Agreement to Sign Before Anyone Uploads

A split sheet is the first document, and for an AI song it needs more than the usual columns. The standard version records each writer’s percentage of the composition, their performing rights organization, which is the body that collects performance royalties, their IPI number, the identifier those organizations use to track writers, and their publisher if they have one. Interspace Music’s guide makes the point I’d underline most: composition splits and master splits are different agreements, and mixing them on one page causes rejections at PROs and administrators. I keep two sheets, one for the song and one for the recording.

For an AI collaboration I add four things. First, a column recording what each person actually contributed, in words, because the human contribution is what any later registration, PRO filing, or dispute will turn on. Second, a line identifying the account holder and the subscription tier, the generation IDs, and the download date, as proof the download was permitted. Third, a clause in which the account holder grants the other collaborators, and any label or distributor, a nonexclusive license to exploit the downloaded output under the rights Suno assigned, so the permission isn’t trapped in one person’s login. Fourth, a work-for-hire acknowledgment for anyone paid a fee with no share, signed before they do the work.

On percentages, I’d push back on the instinct to hand the prompter a big songwriter share. If the prompter contributed no lyrics, melody, or edits, a songwriting share is a gift rather than a legal entitlement, and gifts are fine as long as everyone knows that’s what they are. A common split is 50 percent to the lyricist and 50 percent to whoever rewrote the melody or arrangement on the composition side, with the account holder compensated through a share of the master or a flat fee. Whatever you choose, make the numbers total exactly 100 percent on each sheet. Sync Summit’s splits guide notes that an agreed email chain can document ownership, but I’d still get signatures.

Keep the receipts too: the DAW project with your human parts on separate tracks, the dated lyric file, the Suno generation history, and the download confirmation prove your song has a human layer worth protecting.

Registering, Filing With Your PRO, and Getting It Distributed

Registration with the Copyright Office is optional, but you can’t sue for infringement in federal court without it, and for a multi-author work it costs $65 on the Standard Application, according to the Copyright Office’s current fee page. The $45 Single Application is only for one author who is also the sole claimant, so a collaboration can’t use it. The Office’s new schedule, sent to Congress in July 2026, raises those fees to $85 and $55, and Interspace Music’s analysis puts the effective date around the middle of November 2026, so file before then if the song is ready.

On the application you must disclose AI-generated material and exclude it from your claim, describing the human contribution you’re registering, such as lyrics, a vocal melody, and the arrangement. List every human co-author. Don’t list the AI, and don’t list the prompter as an author of music they didn’t write.

For performance royalties, ASCAP, BMI, and SOCAN aligned their policies in October 2025 to accept partially AI-generated works that combine AI content with human authorship. BMI’s own AI page is blunt that prompt-only works with no meaningful human creative contribution are not accepted, that only humans can be listed as writers, and that qualifying hybrid works are paid the same as fully human songs. Register the human writers at the percentages on your split sheet and answer the AI disclosure question truthfully.

Distribution is where the one-account problem bites. DistroKid accepts AI music, but its help center requires the uploader to own 100 percent of the rights to everything in the track, and since April 2026 its upload form asks which parts were made with AI and sends that as AI credits that Spotify, Apple Music, and YouTube can display. That’s why your agreement needs the license clause: the uploader has to be able to say truthfully that the group’s rights and the account holder’s Suno permission are both behind them. Wavmark’s July 2026 distributor comparison found that TuneCore only takes generative AI music from tools trained on fully licensed datasets, while CD Baby, under the Downtown Music content policy, won’t accept AI-generated content at all, even partial. Undetectr reports TuneCore began accepting tracks from Suno’s partner models on September 8, 2026; confirm that on TuneCore’s site first. Spotify itself doesn’t ban AI music. Its September 25, 2025 policy update targets impersonation and spam and backs the DDEX standard, the industry’s metadata format, for AI disclosure in credits, and the AI Credits beta went live on April 16, 2026 through DistroKid first.

If You’re in the UK or the EU

The US answer above is not the UK answer, at least not yet. Section 9, subsection 3 of the Copyright, Designs and Patents Act 1988 gives a computer-generated work with no human author a copyright lasting 50 years, and hands it to the person who made the arrangements necessary for its creation. In a group that’s usually the account holder, though no UK court has yet decided how the provision applies to generative AI. The government’s March 18, 2026 report proposes repealing it, with no timetable, and confirms that AI-assisted works with a human creative contribution stay protected, as HSF Kramer’s summary of the report explains.

In the EU there’s no equivalent provision. Copyright requires the author’s own intellectual creation, so a purely generated layer is unprotected in much the same way as in the US, and joint authorship turns on each person’s creative choices. Mishcon de Reya’s horizon scan notes that the EU AI Act’s transparency rules for AI-generated content took effect on August 2, 2026, which affects labeling rather than ownership. Suno’s terms are governed by Massachusetts law wherever you live, so the contract layer works the same in London or Berlin as it does in Boston.

An Honest Reality Check

Let me say the quiet part. In a typical AI collaboration, the thing that sounds like the song, the groove, the chord changes, the timbre of that generated voice, is the part no one owns. Your paperwork protects the human additions and allocates the platform’s permission. It can’t stop a stranger from generating something similar, and Suno says as much in its terms. Suno also reworded its own ownership FAQ in December 2025, as Digital Music News documented, from telling paid users they own their songs to telling them they generally aren’t considered the owner even with commercial rights.

That’s why the right answer for a track with real commercial ambitions is often to replace the generated layer. When we rebuilt the weekend song with a real bass part, real keys, and a vocalist singing my lyrics to a melody we rewrote, every element of the finished recording had a human author, the registration covered the whole thing, and CD Baby stopped being off limits. Doing that yourselves in a DAW works if you have the players. Hiring session musicians, or a studio that re-records AI tracks with real performers, is the other path, and either way you end up with something ownership law can actually hold onto.

This area also moves monthly, with Suno rewriting its terms twice in 2026, so treat this article as accurate on October 14, 2026 and reread the platform terms before every new group project.

What to Do This Week

If you have a shared AI song sitting in someone’s library, do four things before Friday. Write down who did what, in sentences, and get everyone to sign it. Confirm which account generated and downloaded the file, and that the download was made on a paid tier within the allowance. Draft the composition and master split sheets with a license clause from the account holder to the group. Then decide, with clear eyes, whether the generated layer is good enough to live with its ownership gap, or whether you’d rather re-record it and own all of it.

Who owns an AI song with multiple collaborators comes down to what each person made, whose account made the rest, and what you all signed. Two of those three are already fixed by the time the song exists. The third is still yours to write.

Sources

• U.S. Copyright Office, Copyright and Artificial Intelligence Part 2: Copyrightability: https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-2-Copyrightability-Report.pdf
• Finnegan, Supreme Court Declines to Hear Thaler v. Perlmutter, Leaving Human Authorship Requirement Intact: https://www.finnegan.com/en/insights/ip-updates/supreme-court-declines-to-hear-thaler-v-perlmutter-leaving-human-authorship-requirement-intact.html
• United States Court of Appeals for the Ninth Circuit, 17.9 Copyright Interests, Joint Authors: https://www.ce9.uscourts.gov/jury-instructions/civil/chapter-17/17-9-copyright-interests-joint-authors-17-u-s-c-101-201-a/
• Lexology, Music Law 101: Who Owns the Copyright in a Song?: https://www.lexology.com/library/detail.aspx?g=152cdcf7-36b9-4f58-a473-95db73d65af5
• Law Offices of Edward R. Hearn, Collaborator/Songwriter Agreements: https://www.internetmedialaw.com/articles/collaborator-songwriter-agreements/
• Suno, Terms of Service, effective September 3, 2026: https://suno.com/terms-september-2026
• Suno, Pricing: https://suno.com/pricing
• Undetectr, Suno Pricing 2026: Free vs Pro vs Premier, Which to Buy: https://undetectr.com/blog/suno-free-vs-pro-vs-premier
• Digital Music News, Suno Previews 2026 Changes Under Warner Music Deal: https://www.digitalmusicnews.com/2025/12/22/suno-warner-music-deal-changes/
• Digital Music News, Udio Opens Downloads for 48 Hours After UMG Deal Backlash: https://www.digitalmusicnews.com/2025/10/31/udio-umg-deal-downloads-resumed-48-hours/
• AI Tools Directory, Udio Review 2026: Pricing, Licensing And No Downloads: https://www.aitools-directory.com/blog/udio-ai-review-2026/
• Interspace Music, What is a Splits Sheet? The Document That Decides Who Gets Paid: https://interspacemusic.com/blog/glossary/what-is-splits-sheet/
• Sync Summit, Splits: https://syncsummit.com/splits/
• U.S. Copyright Office, Fees: https://www.copyright.gov/about/fees.html?lang=en
• Interspace Music, The US Copyright Office Fee Rise Is Not 55% for Music: https://interspacemusic.com/blog/?p=19702
• BMI, BMI and AI: https://www.bmi.com/specials/bmi-and-ai
• Music Week, ASCAP, BMI and SOCAN Align on Musical Compositions Using AI Tools Alongside Human Authorship: https://www.musicweek.com/publishing/read/ascap-bmi-and-socan-align-on-musical-compositions-using-ai-tools-alongside-human-authorship/092955
• Wavmark, Every Major Music Distributor’s AI Policy, Compared: https://wavmark.io/blog/music-distributor-ai-policies-2026
• Undetectr, DistroKid AI Policy 2026: Does DistroKid Allow AI Music?: https://undetectr.com/blog/distrokid-ai-policy-2026
• Spotify Newsroom, Spotify Strengthens AI Protections for Artists, Songwriters, and Producers: https://newsroom.spotify.com/2025-09-25/spotify-strengthens-ai-protections/
• HSF Kramer, UK Government Report on Copyright and AI Concludes More Evidence Is Needed Although Section 9 Subsection 3 CDPA Could Go: https://www.hsfkramer.com/notes/ip/2026-03/uk-government-report-on-copyright-and-ai-concludes-more-evidence-is-needed-although-s9-3-cdpa-could-go
• Mishcon de Reya, GC Horizon Scan July 2026: https://www.mishcon.com/reports/mishcon-gc-horizon-scan-july-2026/page-11