All articles

How to Make an AI Song Copyrightable

I have registered my music with the US Copyright Office, and I have seen people realize too late that the AI song they used to launch a whole business belonged to nobody. Not to them, not to the platform, not to anybody. It resided in the public domain, ready to be appropriated by some unknown party and distributed, releasing income that had nothing to do with the creator.

And this is the problem that needs fixing. So here’s how to make an AI song copyrightable. It is possible. It is just impossible to do the usual stuff – type a prompt into Suno, get the AI song, and expect to get copyright protection because of this download. The following is a step-by-step guide to making an AI-generated song copyrightable, as it was outlined by the Copyright Office itself.

Short Summary

It is impossible to register a song that was generated from a text prompt by AI as a copyrightable material in the United States. However, a song where a human made a creative contribution can be copyrighted. And this is the way to go. You need to make sure that there is a real creative input from a human in the process, such as writing the lyrics, performing or replacing elements, making certain editorial decisions, and registering the song with an explicit declaration of the scope of your copyright.

The Copyright Office clarified this approach to AI-generated material in its report “Copyright and Artificial Intelligence, Part 2” that was issued on January 29, 2025. According to the law firm Rimon, in order for an AI-generated work to be copyrightable, it needs to reflect some human authorship, and the examination of human contribution takes place on a case-by-case basis. Everything mentioned below relates to making sure that you satisfy this requirement and prove that you did.

Why Just Inputting a Prompt Is Not Enough

People usually object to this part the most, so it needs to be discussed upfront. Even if you spent hours perfecting your prompt, re-generating the AI song dozens of times, and using style tags, none of this effort is recognized as human authorship by US law.

The Part 2 report mentioned the issue of prompting in detail. According to the analysis by Mintz, the Copyright Office concluded that prompting per se does not provide sufficient human authorship. A prompt is an idea and an instruction, and copyright protects the expression, not ideas. The AI model generates the melody, voicings, drum patterns, vocal phrasing, etc. You wanted a sad indie song about leaving Ohio, and the machine produced it.

Moreover, revising the prompt and re-generating it multiple times also does not affect this conclusion. Sidley Austin noted that regenerating the material does not mean having some kind of control over its expression; it just means having more lottery tickets. The Copyright Alliance says the same thing – current models fill in the gaps, ignore the instructions, and produce different output based on the same prompt, and that is why the human is not the author.

There is one exception though, that deserves special attention. If you fed the machine your original creative content (for example, recorded melody or written lyrics), and if this content is still perceptible in the output, then the copyright protection is available for this particular portion. And now we need to understand which portions can be covered.

According to the Part 2 report, there are three kinds of human authorship in the AI-assisted material that may be eligible for the copyright protection. Each of the following techniques correlates with one of these categories.

First, human-authored expression that is perceptible in the output. It means your lyrics, sung by an AI voice, or recorded melody, around which the AI arranged a song. Your words and tunes are copyrightable, even if the AI sings them.

Second, creative selection, coordination, or arrangement of AI material. Selecting which generated parts to use, determining the run of verse, chorus, bridge, half-chorus in a song, arranging stems, or trimming the unnecessary second verse can be considered authorship, just like making a collage from the found images.

Third, creative modification of AI output. Re-recording the vocal part, replacing generated bass line with your playing, rewriting generated lyrics, or restructuring the mix can contribute to your authorship of the song.

Notice that there is nothing on the list such as effort, time, money, or taste. The Copyright Office explicitly rejected the theory of sweat of the brow. You do not receive copyright by spending effort on prompting. You need to express something.

Write the Lyrics Yourself and Keep Your Drafts

This is one thing that needs to change about your workflow. Your lyrics that are copyrightable are a literary work protected by the copyright since the moment you fix them in a document (notes app will do the job). And when the AI sings them, they still remain your creative work in the context of the entire song. It is the clearest way to demonstrate the human authorship of the track, and the only one, in which there is absolutely no doubt about the authorship.

My workflow is the following. I write my lyrics in the document where version history is enabled (it can be Google Docs, for example). Then I date my drafts. Then only I paste the lyrics to the generator in the custom mode. If the generator misinterprets some line or generates additional phrases that I did not write, I either regenerate the song or note that this phrase is not mine, since it will be important for the registration of the song.

What does not work? It is the situation when you ask the AI to generate some lyrics, and then swap a couple of words with your own. If you want to receive copyright protection for the lyrics, they should be original. A good rule of thumb is the following – you should be able to show a judge the initial draft with cross-outs and additional notes. The AI does not produce this kind of content, only humans.

Put Some Real Performance into the Track

The second part of human authorship is adding or replacing audio with a human performance. Record your vocal on the instrumental track, perform a guitar solo, add some live bass, hand percussion, keyboard line that you have written. This approach is recommended in several producers’ guides, and MusicWave described it very well in the context of commercial AI music: treat the AI-generated material as the building blocks of your track, just like the loops of songs.

But there is also a technical bonus in this technique: US registration of the music involves two copyrights – the composition (melody, harmony, and lyrics) and the sound recording (a fixed performance). The fully AI-generated sound recording means that you have a weak claim about it. But if you added your vocal to the track, you created a real human performance in the sound recording, the proof of which an examiner would have no trouble finding.

If you cannot sing or play instruments, hire someone who can. The session vocalist on the work-for-hire agreement will cost you less than most people spend on plugins in a year. And it will turn your uncopyrightable demo into a real track with human performance in its center. I started thinking of the AI music platform as of the world’s fastest demo studio. The demo is not the product, and the record you build on its basis is.

Selection, Arrangement, and Editing That Can Count

Even without singing or playing, you can create authorship through editing, but the bar here is higher than people usually think, so be honest about it.

One song generation followed by trimming the intro is not the creative arrangement. Things that can be considered it: generation of stems or several sections of the song, assembling the song structure that AI would not generate itself. For example, taking the chorus from the 12th generation, the verse groove from the 7th generation, cutting them, writing a new transition, automating dynamics, and changing the key of one section to create a lift. As I said before, the Premier access to stems is very helpful for that. According to the guide to registering Suno tracks by RightsDocket, the access to stems helps with documentation, since it allows separating the modified portions from AI-generated material.

Keep the project file with the information about your edits, fades, and rearranged regions. DAW session (Logic, Ableton, Reaper, etc.) is the difference between saying that you made some choices and proving them. The screenshots of the project and export logs also help. The Copyright Office relies on the applicant honesty, but in case of court challenge of your registration, you will need a documentation.

Here is one thing that has nothing to do with the Copyright Office but kills your song anyway. Using a free plan when generating the song. Copyrightability and commercial rights are two different gates to pass, and the terms of service of the platform determine your ability to pass the second gate.

For Suno, the songs that were created on the Basic (free) tier are restricted to personal non-commercial use, and Suno owns them. Upgrading to other plans does not change this fact. According to the breakdown of the situation by Dynamoi, the upgrade does not retroactively license the songs that you made while being free user, so the creation date and your subscription status on that date matter. For Suno Pro, you will pay $10 a month (or $8 per month if billed annually) for 2,500 credits, and for the Premier level, it is $30 per month ($24 monthly if billed annually with 10,000 credits and studio features). If you want to release something, you need to make it on any subscription and save the subscription invoices.

One thing to know: after the settlement with Warner Music Group in November 2025, Suno has updated its ownership language, and the songs generated with the licensed models will replace the current ones, and the download limits will also come. According to Music In Africa, Suno has softened its initial statement that its songs belong to the paying subscribers. Translation: the ground is shifting under the platform, and your rights depend on the terms of service that were valid during the moment of the song generation, so you need to save the dated copy of the terms next to your song.

Udio is a completely different case right now. Following the settlement with Universal Music Group in October 2025, Udio disabled the audio download of its songs, which Chartlex describes as a walled garden where you cannot export anything, even if you subscribe to the Pro plan for $30 per month, which has commercial rights. Until the download is enabled again with the fully licensed version in 2026, it is impossible to release any Udio songs elsewhere. For everyone who wants to make a copyrightable song now, it eliminates Udio from the consideration, regardless of the sound you like or not.

How to Register the Song Without Getting Rejected

And now for the part that almost nobody explains properly: registration. You need to register via the Copyright Office’s website (eCO), and in case your song includes more than a trivial amount of the AI-generated material, you cannot use Single Application, and need to use Standard Application. RightsDocket’s guide is very straightforward about that – Single Application does not have an option to exclude some material from the claim, and exclusion is the key part of the registration of the AI-generated songs. The cost of Standard Application is $65, and it is non-refundable, so you need to fill the claim correctly.

You will need to choose between two forms. Form PA covers the musical composition (melody, harmony, and lyrics). Form SR covers the sound recording (the specific fixed performance). If you have both and want to register both with one filing, the SR form can be used for that, but think carefully which side of the song contains your human authorship. If you wrote the lyrics, but the AI generated the whole song, the strong claim will be made for the composition. If you sang and played over the generated backing, the sound recording claim will become much more legitimate.

The core of the application will be the Limitation of Claim section, where you will tell the truth in three steps. In the Author Created field, describe your human authorship in details – original lyrics, lead vocal performance, selection and arrangement of the instrumental material generated by AI, editing and structural revisions. In the Material Excluded field, disclose and disclaim the AI-generated portions of the song, for example, AI-generated instrumental accompaniment and backing vocals generated with a generative music tool. RightsDocket provides the examples of Limitation of Claim language for the AI-assisted songs, and you need to look at them prior to making your claim in order to avoid vague phrasing that results in correspondence from the examiner.

Do not try to be funny here. The requirement of disclosure is mandatory for any appreciable amount of the AI-generated material, and the attorneys of Finnegan warn that hiding AI involvement in your song will result in rejection of the registration. The Copyright Office relies on applicant honesty, and the registration based on a lie is a liability with a certificate attached. Also, do not list the AI tool as the author. It is a machine, authors are humans.

One useful piece of advice from the paralegal side of the industry: if you have a bunch of unpublished songs, you can use group registration for unpublished works (GRUW), and in this case you will be able to register up to 10 unpublished songs with one application, provided they have the same author and claimant. If you are an AI-assisted prolific songwriter and want to protect a batch of your unpublished demos, it is much cheaper than making individual filings. And if the examiner rejects your claim and you think that he or she made a mistake, there is a first level of appeal, but it costs $350, so it is yet another argument in favor of a clear and honest application.

The Documentation that Saves Your Back

Registration is the snapshot. Documentation is the weapon that will win you any dispute. Here is the set of files that I create for each AI-generated song, and I strongly recommend you do the same.

  • Dated drafts of lyrics with the visible revision history. Prompt logs and generation IDs not because prompting creates the rights (it does not), but because it establishes the timeline and the list of output that was used as the source material.
  • DAW project file with the edit history.
  • Stems, both original and modified versions, in order to demonstrate the difference.
  • Recordings of your performance, preferably with session dates.
  • Dated copy of the platform’s terms of service from the moment of the song generation.
  • Your subscription invoices, dated.
  • Work-for-hire or split agreement (if any).

It seems useless until the first time someone steals your song, or a sync licensor requests the warranty of your authorship, or a distributor flags your song and asks about the source. And at that point, this folder is the difference between five-minute email and lost opportunity. The Copyright Office decides these questions on a case-by-case basis, and the same will happen in courts. Cases are won by documentation.

Distribution Is the Other Battle

Copyright gives you the ownership. It does not mean that you can immediately distribute your song on Spotify or anywhere else, since there are separate rules for the streaming layer. And you need to consider these rules when you generate the song.

In 2025-2026, the distribution landscape changed. DistroKid accepts the AI music without restrictions, provided you have all the rights and completed the AI disclosure step, which now requires structured AI Credits data that Spotify uses. Jack Righteous made a detailed breakdown of current policies of distributors and streaming platforms regarding AI music, and the rule is quite consistent: disclosure is required, impersonation is forbidden, spam patterns are punished. TuneCore has tightened its policy and now does not accept the AI-generated songs unless the generator is in its partner list. CD Baby refuses to distribute fully AI-generated songs. Deezer flags detected AI tracks and excludes them from algorithmic recommendations.

Spotify’s bulk takedown should be understood properly since the big number in the headline is scary for the wrong reasons. They pulled down tens of millions of tracks, yet the majority of the removals concern non-disclosure, voice impersonation, duplication, and stream manipulation instead of AI generation as such. A single disclosed, original, and human-generated release is fine. Fifty clones a week is not, and if that’s your strategy, copyright concerns are the last of your troubles.

Notice the neat fit between the streaming guidelines and the copyright guidelines. Whatever makes an AI song copyrightable – lyrics, performance, arrangement, human creative decisions – prevents it from being spam and keeps it compliant with the distribution policy. One piece of work does the trick twice.

Some Uncomfortable Truths Before Building a Business On That

Here are a few uncomfortable facts before you start building a business on this.

Your registration will be more narrow than a fully human-made song. Elements created by the algorithm are not copyrighted, so somebody else could legally take the instrumental track and record their lyrics over it. What you copyright with your song is your lyrics, your performance, your arrangement, and your combination of those. That’s normally sufficient for an independent release, yet for a track that you plan to pitch to sync libraries, it is better to consult a lawyer since buyers will ask you for the warranties that an AI-generated composition complicates a lot.

Case-by-case analysis is truly fuzzy. Nobody, and not even the lawyers issuing client alerts about it, could give you an exact number of changes that would turn curation into creation. The Office stated that further progress in AI technologies could shift the line for the prompting analysis one day. So far, do your best in the areas where you could invest with certainty: the lyrics that you wrote yourself, the sounds that you performed.

And the platform layer is unstable. Suno moves toward licensing and download caps after the Warner deal. The catalog of Udio is locked on platform right now. Sony filed a lawsuit against Suno that was scheduled to reach the summary judgment phase in mid-2026. Nothing changes copyright law but affects your ability to practically distribute the file. A song that is copyrighted and could not be exported anywhere is an odd achievement.

What I’d Do This Week

Here is my recipe for one copyrightable AI song, shortened.

Write your complete lyrics in a dated document. Get a paid subscription before making anything – $10 on Suno Pro would cover any commercial use. Generate the track in custom mode with your lyrics, extract the stems, and rebuild the structure in your DAW. Record your vocals or hire a vocalist with a signed agreement. Export the song, and collect all the evidence in one folder. File a Standard Application via eCO ($65). Claim your lyrics, your performance, your arrangement, disclose and exclude the AI-generated material in the Limitation of Claim, keep the confirmation. When distributing, disclose the use of AI software.

That’s a weekend of work and a filing fee, and you get a song with your name on the certificate and authorship that is easy to prove and that won’t go away during any verification. The technology is amazing, yet the law requires a person somewhere inside your song. Be sure it’s you.

Sources

Common questions

Why does prompting not count as authorship for copyright?

Because a prompt is an idea and an instruction, and copyright protects expression, not ideas. The Copyright Office's January 2025 report concluded that prompting alone gives insufficient human authorship, and that regenerating dozens of times does not create control over the output, it only buys more lottery tickets. The Office also rejected any sweat of the brow theory, so effort, time and money spent on prompting earn nothing.

What parts of an AI song can be copyrighted?

Three kinds of human contribution qualify. Human authored expression perceptible in the output, such as your lyrics sung by an AI voice or a melody you recorded that the AI arranged around. Creative selection, coordination and arrangement of AI material, like assembling sections from different generations into a structure the AI would not produce. And creative modification, such as re recording vocals, replacing a generated bass line or rewriting lyrics.

Is writing my own lyrics enough to copyright an AI song?

It is the clearest and least disputable form of human authorship. Lyrics are a literary work protected from the moment you fix them in a document, and they stay yours when an AI sings them. Write them in a file with version history, date the drafts, then paste them into the generator in custom mode. Swapping a few words in AI written lyrics does not count, since you should be able to show a judge a draft with real cross outs and revisions.

Does adding my own vocals make an AI song copyrightable?

It creates a real human performance inside the sound recording, which strengthens the Form SR claim considerably. US music registration involves two copyrights, the composition and the sound recording, and a fully generated recording gives you a weak claim on the latter. If you cannot sing or play, a session vocalist on a work for hire agreement turns an uncopyrightable demo into a track with a human performance at its centre.

Does editing an AI song in a DAW count as creative arrangement?

Only if the editing is substantial. Trimming an intro after one generation is not enough. Taking the chorus from one generation and the verse groove from another, writing a new transition, automating dynamics and changing the key of a section can qualify. Keep the project file with edit history, stems before and after, and screenshots, because the Copyright Office relies on applicant honesty but a court challenge will require proof.

Do I need a paid Suno plan to copyright my song?

Copyrightability and commercial rights are separate gates. Songs made on Suno's free tier are owned by Suno and limited to personal use, and upgrading later does not change that, so a copyrightable but unreleasable song is pointless. Suno Pro costs $10 a month or $8 annually and Premier $30 or $24 annually. Save dated invoices and a copy of the terms in force when you generated, since Suno's ownership language changed after the Warner settlement.

Can I copyright and release a song made with Udio?

Not practically right now. After settling with Universal Music Group in October 2025, Udio disabled audio downloads, so even Pro subscribers with commercial rights cannot export their songs. Until a fully licensed version restores downloads, the article rules Udio out for anyone who wants to release a copyrightable song elsewhere.

How do I register an AI assisted song with the Copyright Office?

File a Standard Application through eCO for $65, since the cheaper Single Application has no way to exclude material. Choose Form PA for the composition or Form SR for the sound recording, or SR for both. In the Limitation of Claim section, describe your human contribution in the Author Created field and disclose the AI generated parts in Material Excluded. Never list the AI tool as an author, and never hide AI involvement, which leads to rejection.

Can I register several AI assisted songs at once?

Yes, if they are unpublished. The group registration for unpublished works lets you register up to 10 songs with the same author and claimant in one application, which is far cheaper than individual filings. If an examiner refuses your claim and you believe it was a mistake, a first level appeal costs $350, which is one more reason to file a clear and honest application the first time.

What documentation should I keep for an AI assisted song?

Dated lyric drafts with revision history, prompt logs and generation IDs to establish the timeline, the DAW project file with edit history, original and modified stems, recordings of your performances with session dates, a dated copy of the platform terms, subscription invoices, and any work for hire or split agreements. This folder is what settles a theft dispute, a sync licensor's warranty request or a distributor flag in minutes rather than costing you the opportunity.