If you can type “sad indie pop track about breakups in the rain” and get a polished song back in seconds, it is natural to wonder: is this “your” song?

Generative music tools are moving faster than copyright law. You can already make full tracks with AI systems like Suno, Udio, Stable Audio, or custom models plugged into ChatGPT, Claude, or Gemini. Record labels are suing, lawmakers are holding hearings, and musicians are asking whether they are being quietly trained out of a job.

Underneath all of that drama is one deceptively simple question: who owns an AI-generated song? You, the AI company, nobody, or the original artists used in training? The uncomfortable answer right now is: “it depends,” and in some cases, “maybe no one at all.”

In this post, you will get a clear, practical overview of how music copyright normally works, how AI tools complicate it, and what you can realistically do today to protect yourself if you are experimenting with AI music.

Before we add AI to the mix, it helps to remember how music copyright works in the analog world.

In most countries, including the US, copyright protects “original works of authorship” that are fixed in a tangible medium—like recording a track to your laptop or writing a lead sheet on paper.U.S. Copyright Office, “What Is Copyright?” For a typical song, there are two separate copyrights:

  • The musical work (composition): melody, harmony, lyrics.
  • The sound recording: that specific recording of the performance.

If you write and record a song from scratch, you are usually the copyright owner of both. You can license them, assign them to a label or publisher, or keep them.

Two key ideas matter here:

  • The work must be original (not just a copy of someone else).
  • The work must be created by a human author.

That second one is the land mine for AI-generated songs.

AI and the “human authorship” problem

In 2023, a US federal court in Thaler v. Perlmutter confirmed what the U.S. Copyright Office has been signaling for years: purely machine-generated works without human authorship are not eligible for copyright.Loeb & Loeb summary of Thaler v. Perlmutter The court upheld the Office’s refusal to register a visual artwork that the applicant claimed was created entirely by an AI system.

The U.S. Copyright Office later issued formal guidance on AI works: if an AI system “determines the expressive elements” of the output, that part of the work is not protected by copyright; only the human-authored parts can be.U.S. Copyright Office, “Copyright Registration Guidance: Works Containing Material Generated by AI”

Translated into music:

  • If you simply type a short prompt into an AI tool and accept the result, large chunks of that track may not be protected by copyright at all.
  • If you do significant human editing—rewriting melodies, changing lyrics, rearranging structure, recording your own vocals or instruments—you may own the copyright in the parts you contributed, but not the unedited AI material.

Practically, that means:

  • There might be no enforceable copyright in a “one-click” AI instrumental you generate for a YouTube video.
  • You still might own the copyright in custom vocals you record on top, or a heavily reworked composition derived from the AI sketch.

The US is not alone. Many jurisdictions still implicitly anchor copyright in human authorship, even if they have not tested AI music in court yet. The bottom line: AI on its own does not qualify as an “author” under current law.

So who owns an AI-generated song right now?

Because of this human-authorship rule, the usual question “who owns it?” splits in two:

  1. Who (if anyone) owns copyright in the track?
  2. Who has contractual rights to use or control it?

For a typical text-to-music output where you do minimal editing:

  • In the US, there may be no copyright at all in the AI-generated parts.
  • Any human contributions (lyrics you write, melodies you explicitly compose, performances you record) can be protected, but the AI-generated backing track itself may be in a gray area or unprotected.

Some AI music platforms acknowledge this directly in their terms of service: they say you “own” the outputs, but they often mean contractually, not that the law automatically grants you traditional copyright. Others give you a license as broad as possible, but still within the unsettled legal environment.

2. Contractual rights (the fine print you click through)

Even if the law does not clearly grant copyright, contracts can.

When you use an AI tool—whether that is ChatGPT with a music plugin, a standalone app like Suno or Udio, or a DAW plug-in powered by a generative model—you agree to its terms. Those terms often decide:

  • Whether you can use the outputs commercially.
  • Whether the platform can reuse your prompts or outputs to further train its models.
  • Whether you can claim to be the owner (for practical purposes) of the generated audio.

This is why two tools can feel similar but be very different legally. One might give you broad commercial rights and promise not to reuse your song; another might retain sweeping rights and restrict how you can monetize it.

If you plan to release AI-assisted tracks on streaming services, sync them to video, or pitch them for ads, read the section about “ownership” or “license to outputs” in the tool’s terms before you do anything else.

What about the artists in the training data?

A separate, fiery debate is about how these models were trained, not just who owns the outputs.

Training powerful music generators typically requires ingesting vast amounts of existing recordings. Labels and artists argue that using their catalogs without permission infringes reproduction and other rights; AI companies often counter that training is a kind of analysis and should be treated like text and data mining or fair use.

In June 2024, major labels including Sony Music sued AI music startups Suno and Udio in US federal court, alleging that their systems were trained on massive catalogs of copyrighted sound recordings without authorization.”Artificial intelligence and copyright” overview of Suno and Udio lawsuits Suno and Udio have pushed back, arguing that their systems learn patterns and styles rather than storing or reproducing specific tracks, and that this kind of data mining is lawful.LegalClarity, “Music Copyright AI Lawsuits: Where Every Case Stands Now”

Meanwhile, Stability AI—behind the Stable Audio system—faces lawsuits from a musician claiming his works were used in training despite contractual limits and opt-out requests.Digital Music News, report on Anders Manga v. Stability AI

In the EU, lawmakers have tried to get ahead of this with explicit text and data mining (TDM) rules. Under the EU’s Copyright in the Digital Single Market Directive and the new AI Act, rightsholders can reserve their rights against TDM for commercial AI training; providers of “general-purpose” AI models are then expected to respect those opt-outs and publish summaries of training data.AI Act Service Desk, Recital 105 on text and data miningEuropean IP Helpdesk, “Artificial intelligence and copyright”

For you as a creator using these tools, the key takeaway is:

  • These training-data lawsuits target the AI providers, not end users.
  • However, if courts eventually decide that certain models are unlawful, that could create uncertainty around songs heavily built on those tools.

Human + AI collaboration: the sweet spot (for now)

The most legally comfortable place to be is AI-assisted, not AI-only.

If you use AI to generate:

  • rough chord progressions,
  • melodic ideas,
  • drum grooves,
  • or sound design elements,

and then you substantially rewrite, arrange, and perform the piece yourself, you strengthen the argument that your final track is a human-authored work with copyright protection. That is true whether the initial ideas come from a local model, a plugin tied to something like Gemini, or a commercial service.

Think of AI as:

  • A session musician or co-writer who never asks for credit, or
  • A smart instrument that reacts to your prompts instead of MIDI notes.

You still need to do enough creative lifting that a court (or a copyright office) would see the end result as primarily your expression, not the model’s.

In the US, if you try to register an AI-assisted song, you are expected to:

That is a hassle, but it is also a roadmap: it tells you where the law thinks your protectable authorship lives.

Practical tips if you are using AI to make music

Here are some concrete ways to stay on firmer ground while experimenting:

  • Stay in the driver’s seat creatively. Use AI for drafts and textures, but make clear human choices about structure, melody, lyrics, and performance.
  • Keep project files. Save stems, DAW sessions, and prompt histories. If anyone ever questions whether you did meaningful authorship, your process is evidence.
  • Record something yourself. Even if the backing track is AI-generated, adding your own vocals, guitars, or synth parts creates more clearly human-authored material.
  • Watch for “sound-alike” risks. If a model spits out something that obviously mimics a famous artist—especially in vocals or a distinctive riff—treat it as radioactive. Change it heavily or toss it.
  • Read the tool’s terms for output rights. Make sure they at least grant you a broad license to use outputs commercially before you release them.

And remember that tools like ChatGPT, Claude, and Gemini can also help you on the “boring” rights side—drafting split sheets, license requests, or even simple language to add to your website clarifying how you used AI—once you have the legal basics in mind.

Where this is heading

Courts and regulators are still in the early innings. The US Copyright Office is running a multi-part study on AI and copyright, including one report focused on generative AI training and how copyright should apply.U.S. Copyright Office, “Copyright and Artificial Intelligence, Part 3: Generative AI Training” The EU’s AI Act will start to bite as its obligations phase in, forcing big model providers to be more transparent about training data and opt-outs.

In the meantime, expect:

  • More lawsuits over training data from labels and collecting societies.
  • Gradual clarification about whether large-scale training is fair use, licensed use, or infringement in different jurisdictions.
  • Platforms tightening or revising their terms of service as rulings come down.

But none of that means you have to sit on your hands creatively.

Actionable next steps for you

If you are making or planning to make AI-assisted music, here is what you can do this week:

  1. Pick one track and “humanize” your process. Take an AI-generated idea and rewrite at least the melody or lyrics yourself. Record at least one live part. Save all project files and prompts so you have a clear authorship trail.
  2. Audit the tools you rely on. For each major AI music tool or model you use, skim its terms of service for sections labeled “Ownership,” “License,” or “Use of Outputs.” If the commercial rights are vague or restrictive, consider switching tools before you invest heavily.
  3. Start a simple rights log. For every release that involves AI, write down which tools you used, what they generated, and what you changed. This does not have to be complicated—a single text file per project is enough—but future you (and your lawyer, if you ever need one) will be grateful.

AI will not replace the messy, human part of music that makes it worth listening to—but it is changing how songs are made and how copyright works. If you understand the basics of who can claim what, you can play with the new tools without stepping into legal quicksand.