All articles

At What Point Does Human Contribution Make an AI-Assisted Track Copyrightable?

It isn’t about hitting 30 percent, 29 of which won’t help you either. There is no 30 percent magic number or even some bounce at 29%, and I used to believe there is while starting to register AI-assisted tracks. I have learned one thing for sure since then: it isn’t about how much you have done, it’s about how you did it.

This is a brief overview, and here is the detailed analysis of copyrightability and protection of human work in AI-assisted music. Prompts aren’t enough no matter how detailed or how many times you regenerate them. Writing your own lyrics, performing or playing any of your parts, creatively selecting, arranging, and reworking the generated material can be copyrighted, but only those parts of your work that you have authored yourself. AI-generated audio in its raw form cannot be copyrighted even if you layer your contributions on top. Here is how you can currently protect your music in the United States, and the rest of the post is about doing it effectively.

American copyright law protects works of original human authorship, and the latest decisions confirm it every time. Stephen Thaler tried to register an image called A Recent Entrance to Paradise created by his AI system as the sole author. The Copyright Office refused to accept the application, and the DC Circuit concurred in March 2025. According to Skadden’s overview of the court’s decision, the problem is that ownership, duration of copyright, and inheritance are assumed to be human.

On March 2, 2026, the Supreme Court refused to take the case on appeal, and according to Mayer Brown’s overview, the human authorship rule is confirmed. However, no matter what constitutes sufficient human authorship was not decided, which matters. The floor is set: machines cannot be authors, but the line between assistance and generation is determined on a case-by-case basis by examiners on a tool-by-tool basis.

Why Making Your Prompt More Detailed Gets You Nowhere

The Copyright Office addressed the question of prompts in Part 2 report on copyrightability of works made by artificial intelligence published on January 29, 2025. According to the Office, prompts alone aren’t enough human control to make the creator of the prompt an author of the work. Yeah, it hurts if you spent 40 minutes creating the perfect style description. The reason is that a prompt is a set of instructions sent to a system that generates the material and makes its own choices, which is closer to a commission rather than to a composition. One nuance is that this conclusion is based on today’s behavior of existing systems, and the Office will have to reassess once the technology changes. No tool currently allows that level of user control, and you have to plan accordingly.

The best example is Jason Allen’s Théâtre D’opéra Spatial, the Midjourney image that won a 2022 Colorado State Fair art contest. According to the artist, he spent at least 624 iterations trying to achieve his goal, but the Review Board refused to grant the copyright in September 2023 as the prompts weren’t sufficient to determine the expressive elements generated by the system. IPWatchdog reports that the Office confirmed it in its 2025 analysis.

The takeaway for music is clear: no matter how much time you spend on typing “melancholy southern gothic ballad, sparse piano, weary female vocal, 70 BPM” into a generator, it still isn’t enough. You just re-roll the dice and pick the one you like best, and that doesn’t make you a dice creator.

The Three Doors That Remain Open

The same report describes three scenarios under which you can protect your work, and these lines match perfectly with how the music is being created. First, human-created expression which survives in the final work can be protected – your lyrics, your performed vocal, and any melody you have composed stay under your copyright in AI-assisted tracks. Second, a sufficiently creative selection, coordination, and arrangement of AI-generated material can be protected as a whole, despite inability to protect the parts. Third, modifying AI output in a sufficiently creative way is also sufficient for protection of the modified material.

Second scenario has a precedents although it involves images. On January 30, 2025, the Office registered A Single Piece of American Cheese as a composite created with the help of Invoke’s inpainting. The registration covers the “selection, coordination, and arrangement of material generated by artificial intelligence,” and The IPKat notes that the application was previously rejected but accepted after the CEO of Invoke provided a detailed whitepaper demonstrating all of the creative choices he made. The lesson: evidence matters and should be collected accordingly.

Music has its own precedent in the form of Zarya of the Dawn, the 2023 comic case. Sterne Kessler sums up the decisions of the Office: Kris Kashtanova preserved protection for the text and arrangement of the images, while the individual Midjourney images were excluded from copyright protection. Replace images with AI-generated audio and you will see the shape of your song’s protection. The way you arrange and coordinate the machine’s output can belong to you.

What That Means Inside of a Session

I want to be concrete as the legal language tends to hide the ordinariness of the work which can be protected. Contributions that can count towards copyright include lyrics you have written, a melody you have hummed or played before generating the material, vocals you performed yourself, instrument parts you performed, song arrangement you constructed with the help of cutting and reordering generated material, and substantive production work such as rewriting the harmony under a chorus or comping a lead.

Contributions that aren’t copyrightable include the genre and mood description you have typed, the number of regenerations, selection of your favorite one among the eight outputs, the title change, trimming silence off, and loudness normalization. And this is another trap: using AI as a tool for creating music isn’t going to hurt you. The assistance of AI or AI-generated material in human-created work doesn’t deprive it of the copyright. The assistive tools are allowed. The problem is in the contributions in which the machine made all of the expressive decisions.

One example is clear. On my recent track, the instrumental material is generated by the tool, the lyrics are written by me in advance, the lead vocal is my own performance, and the final song arrangement is created with the help of cutting and reordering the generated parts. When I apply for copyright, I claim the lyrics, the vocal melody, the performance, and the selection and arrangement of the generated material and disclaim the generated audio itself. That is exactly what allowed to get the American Cheese registration, and that is the closest thing AI musicians currently have to protect the whole song rather than the parts.

Lyrics Are the Most Important Part You Have

Here is the real-world example of copyright in music. Xania Monet is an AI R&B persona created by Mississippi poet Telisha Jones. She writes all of the lyrics herself, sends them to Suno for production of the vocals and instrumentals and got into Billboard charts and was signed by Hallwood Media in a deal worth up to $3 million in a bidding war. According to Forbes, the current revenues from the first five songs are around 17 million US streams in two months and $52 thousand in revenue.

Here is what that means for copyright in her catalog. The lyric sheets are plain human works which can be protected traditionally, without any caveats due to AI. The recordings are unclear because Suno’s model creates the performance and production of the song, and Billboard notes that while Jones’s camp claims she owns masters and compositions due to the human authorship, legal experts consider those claims untested. From my perspective, as I analyze this situation weekly, the words are copyrightable, and the master is a question mark. That mismatch has to be taken into account while building your catalog. Whatever you want to own, create it yourself.

The trap that gets people caught in more than anything else in this post. The ownership is granted by the Suno’s paid plans as they say you own your songs, but it isn’t copyright. It is a contract between you and Suno, and a contract binds the two parties, not the random person who rips your track and uploads it.

The tiers matter a lot. With the free Basic plan of Suno, Suno retains ownership of the tracks and limits your use to non-commercial while the paid Pro plan at $10/month and Premier plan at $30/month provide you with the ownership of the tracks created during subscription period and commercial usage rights even after cancellation of the subscription. These rights aren’t retroactive either. Dynamoi’s guide on Suno’s commercial use rules explains that the track created on the free tier will remain non-commercial even if you subscribe to the paid tier later and these terms changed after the licensing agreement with Warner Music Group and now involve download cap and voice option for artists.

But the platform does acknowledge the problem. Suno’s article on copyright protection of songs states that the material “may not be eligible for copyright protection.” A platform can grant you all of the rights it has, but it cannot create the copyright which isn’t created by US law. So, treat platform ownership as the license for commercial use and treat copyright as something you have to earn separately via human authorship. You need both, and they come from different sources.

By the way, be honest while disclosing your work to the service. As LabelGrid’s overview of the stream services’ policies shows, they differ in approach to AI detection and labeling but share one reflex: they punish the tracks which concealed their origin. Deezer already detects and labels AI-generated tracks on the service, and any covert uploading will be revealed pretty soon.

How I Register an AI-Assisted Song

The process is clearly stated in the Copyright Office’s guidance on registration of works containing AI-generated material: you have to disclose. If your track includes more than de minimis amount of AI content, you have to disclose it in your application and exclude the material from your claim. It isn’t just a formality you can neglect because the Office was reviewing the applications which included such content when issuing the guidance, and Part 2 report notes that by early 2025 the Office already registered hundreds of works with AI content.

Practically, the disclosure pushes you to file the Standard Application at $65 rather than cheaper Single Application at $45 because the latter doesn’t include necessary Limitation of Claim fields. RightsDocket’s overview of the fields adds two important warnings I would emphasize: vague authorship description and non-disclosure of the AI use is the quickest way to get in trouble, even cancellation of the registration. I specify the Material Excluded with phrases such as “sound recording and instrumental accompaniment generated by artificial intelligence” and Author Created field specifies: lyrics, vocal melody, lead vocal performance, arrangement and editing of AI-generated material. A short Note to the Copyright Office with a description of your process in plain language costs nothing and avoids the examiner’s back and forth later.

One music-specific issue needs special attention, and it is discussed by SoundGuard AI in detail. A song has two separate copyrights: the composition (lyrics and melody) and the sound recording (actual performance). While you contributed to the lyrics and melody but the model performs and produces the song, you can register the composition confidently and think carefully whether you should register the sound recording. Overclaiming is the misrepresentation which can spoil your application. While registering the human-created composition is easy, registering the sound recording created with AI’s help may be risky.

Also, if you create a lot of music, batch the applications. Group Registration of Unpublished Works allows registering up to 10 unpublished works of the same type in one application for $85 as explained by Trademarkia. This way, a stack of completed demos becomes an affordable registration. I do it every few months, disclose the AI-generated material in my batch and sleep better.

Two timing points. First, registration is what gives your rights teeth because timely registration can allow for statutory damages up to $150,000 and attorney fees for the willful infringement, per Accelerate IP. Second, act quickly. A Federal Register notice issued on March 20, 2026, proposes raising the price of the Standard Application from $65 to $85 and cancelling the Single Application completely. Public comments are already closed in May, and registering the human layer of your catalog will never become cheaper than right now.

The Workflow I Use for the Music I Create

It is simple: design the song in a way when the copyrightable layer would survive if all of the AI-generated content would be stripped. I write the lyrics first, in a dated document, before any generation because this is the cleanest proof of human authorship you will have. Then I generate the instrumental material and immediately download stems (separated vocal and instrumental tracks) for the further work in a DAW.

From there the track gets reconstructed in the DAW, and this is the moment when actual authorship takes place. I strip the AI arrangement, rearrange the track, remove repeating sections, record my own vocal over it or sing the topline my own way, replace at least one core element by my performance and mix with my own production choices. It isn’t an excuse I can provide to the examiner – it just makes the song better. AI arrangements tend to be lazy and overly verbose.

Here is what I don’t do: I don’t claim that my prompts were creative, I don’t claim AI-generated vocals which I just selected from a batch and I don’t submit the application before the human layer of the song is ready because it will be harder and more frustrating to correct it later.

And, of course, I document everything: my prompts, all of the rejected generations, project files, session screenshots, bounce dates. The second attempt of American Cheese registration succeeded because the applicant demonstrated the Office each creative decision in a step-by-step manner. The process which is existing only in your head – is a story, but it doesn’t mean anything without the files proving that it happened.

Reality Check on What Protection Gives You

Be honest about the limitations because I haven’t been at first. All of the work you have left AI-generated and unprotected, and according to Skadden’s overview of the Part 2 report, such content can be copied by anyone. Also, the report considers other countries and discovers that the jurisdictions ruling on this issue so far require human authorship, so any “loophole” offshore is just a myth.

The fighting of the prompting fight is far from over too. Jason Allen tried to sue the Office and the response brief shows that the dispute isn’t over yet as the parties still disagree on whether 624 prompts can be considered as authorship. Follow this case if you like courtroom drama but don’t build a business plan based on it.

Also, understand the environment your track will enter. According to Deezer’s July 2026 statement, more than 50 percent of all daily uploads are AI-generated and on some days this number reaches 90,000 tracks. Deezer’s statistics show that they get only 1–3 percent of streams, and much of them are fraudulent. Billboard has reported that Spotify removed millions of spammy tracks due to its new policy of low-effort AI flooding, and LabelGrid has noted a study proving that 97 percent of listeners cannot distinguish AI music from human music.

Sit with it and understand what it means. Copyright is just one of the survival axes now. Disclosure rules, platform tagging, fraud filters, and listener trust are the others, and the human contribution helps in all of them. It is more than a legal work for most of them.

Where That Leaves You

Forget about how much you have done and ask yourself how much will remain if you strip the AI from the work. If the answer is silence, you don’t have anything. But if the answer is your lyrics, your voice, your song structure, and your edits, then the copyright issue answers itself: those parts are yours, so register them properly, disclose the rest, and keep the proof that you have created them.

The law isn’t settled yet but its movement is quite predictable since 2023 guidance through Thaler’s cert denial, and I wouldn’t bet my catalog on this changing. Only creators who make a human layer in their workflow have something to protect if the next ruling appears. Everyone else is renting.

Sources

Common questions

How much human contribution makes AI music copyrightable?

There is no fixed percentage of human contribution. What is crucial is the ability to control expression: writing lyrics, performing or playing parts, creatively selecting and arranging generated material, or making modifications to it, not how detailed the prompt was or how many times it was regenerated.

Can a detailed prompt make you the author of an AI song?

No. The US Copyright Office treats a prompt as an unprotected idea while a machine generates an expression, which is why no amount of detail in prompt and its regeneration makes prompter an author of output.

What elements of an AI-assisted track can be registered?

Perceivable human-created expression, such as your lyrics or performance, your creative selection and arrangement of generated material and modifications of output, each evaluated on a case by case basis.

Should you disclose the use of AI when registering a song?

Yes. You are required to disclose the presence of more than a de minimis amount of AI-generated material in your work and exclude it from your copyright claim; failure to do so would be reason for canceling a registration.