AI music licensing explained: who owns the songs AI makes?

Who owns AI music in 2026? In most cases, nobody — not the way a songwriter owns a song. What an AI music service sells is a contractual license: permission to use the output under conditions that vary by tier. Under current US Copyright Office doctrine, a track generated entirely by AI likely has no copyright for anyone to own at all. And the training side of the question spent two years in federal court: the major labels sued the two leading consumer AI song generators in June 2024, UMG and Warner struck settlements in late 2025, Sony Music is still litigating, and as of July 2026 no court has ruled on whether AI music training is fair use. This guide separates what the marketing says from what the terms grant — tool by tool, every claim dated.

What did the record-label lawsuits actually change?

In June 2024, Universal Music Group, Sony Music, and Warner Music Group — coordinated by the RIAA — sued the two leading consumer AI song generators, alleging their models were trained on copyrighted recordings without a license. For roughly sixteen months the cases moved toward the central legal question: is training a music model on copyrighted recordings fair use?

Then the settlements started. On October 29, 2025, UMG settled with one of the two services, pairing the deal with plans for a licensed, "walled-garden" platform — and that service disabled downloads the following day. Its users' creations became stream-only inside the platform: as of July 2026 they cannot be exported or released to outside services, and the joint licensed platform promised for the first half of 2026 has not fully shipped. In November 2025, Warner settled with the other, larger service on similar structural terms: licensed models rolling out through 2026, free-tier downloads ending in favor of stream-and-share, and download caps coming to paid tiers.

What has not happened matters just as much. Sony Music remains in litigation against both services, UMG's case against the larger one continues, and a summary-judgment hearing on the fair-use defense was scheduled for July 2026 in the District of Massachusetts. Every headline so far has been a settlement, not a precedent: as of July 2026, no US court has ruled on the merits of fair use for AI music training.

The practical consequence for creators is the shape of the deals. The settlements traded licensed training data for walled gardens, download suspensions and caps, and deprecated models — which means whether a tool lets you take your files and leave has become a real differentiation axis in AI music, arguably the axis. (For the streamer-specific view of all this, see AI music for streamers.)

Do you actually own the music an AI generates for you?

Two different questions hide inside that one: what does the vendor promise you, and what does copyright law recognize?

Vendor "ownership" is a contract. When a music tool's terms say "you own the output," they are promising not to assert rights against you and, usually, to pass along whatever rights they can. That promise is real and worth having — it binds the vendor. What it cannot do is manufacture a copyright.

Copyright is a different test. The US Copyright Office requires human authorship for registration, and purely AI-generated material does not qualify; the federal courts agreed in the Thaler litigation, and the Supreme Court declined to take the case in January 2026 (copyright.gov/ai). The doctrine is about authorship, not medium — it applies to music the same way it applies to images. Human contributions can change the analysis: lyrics you wrote, arrangements you shaped, performances you recorded into the track may each be protectable. But the fully generated audio itself likely is not, which means "you own it" is best read as "no one will stop you — and you probably cannot stop anyone else."

AIVA is the exception that proves the rule, because it sells copyright transfer as a product tier. On its free plan, AIVA keeps the copyright and limits you to attribution-required non-commercial use; on Standard (€11/month), AIVA still keeps the copyright while licensing monetization on a short list of platforms; on Pro (€33/month), the copyright transfers to you, as of July 2026 (aiva.ai). A transfer conveys whatever exists — under US doctrine that may be very little for fully generated material — but the contract's practical effect is real: the broadest grant AIVA offers, with a paper trail a client can file. Other jurisdictions are still working out their own authorship rules, which is exactly why that paper trail sells.

How do the major AI music tools license output in 2026?

The market has settled into six recognizable shapes.

1. The copyright-transfer ladder (AIVA). Rights escalate with price and culminate in transfer, as described above — the transfer itself is the paid feature.
2. The perpetual license with composition retention (Soundraw). Tracks you make while subscribed stay licensed to you permanently, and you keep 100% of recording royalties — but Soundraw retains the underlying composition copyright, and its ~$11/month Creator tier covers background use only. Releasing to streaming services requires Artist tiers at roughly $19–33/month, which also add WAV and stems, as of July 2026 (soundraw.io).
3. The placement exclusion (Mubert). Royalty-free tiers run from free to $199/month, but no plan — none — licenses YouTube Content ID registration, streaming-service release, or stock-site resale, as of July 2026 (mubert.com). Distribution and commercial use are different permissions, and Mubert is the clearest proof.
4. The use-case carve-out (ElevenLabs). Its music product generates full songs with vocals and is positioned as trained on licensed data only. Commercial use starts on the $6/month Starter plan, but self-serve plans exclude film, TV, and studio games — those need an Enterprise agreement, as of July 2026 (elevenlabs.io/pricing).
5. Open weights on licensed data (Stable Audio). Stable Audio 3.0 shipped on May 20, 2026, with tracks past six minutes, and its Small and Medium models are open weights under commercially permissive terms, trained on fully licensed data — 806,000 AudioSparx files plus Creative Commons Freesound audio (stableaudio.com). Instrumental only, no vocals; the free tier is 10 tracks/month non-commercial, and commercial use starts at $11.99/month.
6. The walled garden (the leading consumer song platforms). Post-settlement: licensed models, suspended or capped downloads, stream-and-share tiers. On the largest platform, free-tier songs remain the platform's property, and subscribing later does not retroactively grant commercial rights to tracks made on the free tier, as of July 2026.

Tool	What the free tier grants	Commercial use starts	Copyright position	Notable restriction (July 2026)
AIVA	3 downloads/mo, non-commercial, attribution	€11/mo (listed platforms only)	Transfers to you on Pro (€33/mo)	Full monetization only at Pro
Soundraw	Preview only — no downloads	~$11/mo (background use only)	License; Soundraw keeps the composition	Streaming-service release needs Artist tiers
Mubert	25 tracks/mo, personal use, attribution	$14/mo (social) / $39/mo (full digital)	License only	No Content ID or DSP release on any plan
ElevenLabs	~11 min of music, no commercial use	$6/mo	License only	Film/TV/studio games need Enterprise
Stable Audio	10 tracks/mo, non-commercial	$11.99/mo	License; open-weight models governed separately	Instrumental only — no vocals

One reading note on that table: outside AIVA's Pro transfer, "ownership" language across this market is contractual license, not copyright — see the previous section for why that distinction is structural, not pedantic.

What should you check before paying for AI music?

Six questions, in the order they bite:

1. Which tier grants commercial use — and commercial for what? Mubert's $14 tier covers social use, not everything; ElevenLabs' self-serve commercial rights exclude film and TV. "Commercial" is a scoped word.
2. Can you download files, or are you renting access to a stream? Free tiers of Soundraw and Beatoven are preview-only, and the largest consumer platforms are converting downloads into a capped, paid feature under their label deals.
3. Is releasing to streaming services allowed? Distribution is its own permission. Mubert bans it outright; Soundraw sells it as an upgrade.
4. What survives cancellation? Soundraw's license is perpetual for tracks made while subscribed; AIVA's Pro transfer endures by nature. Many tools are silent — assume nothing that is not written down.
5. Who keeps the composition? You can hold a broad recording license while the vendor retains the underlying composition, which matters the day a sync request or cover version shows up.
6. Do failed generations bill? Across the category, credits are use-it-or-lose-it and failed generations consume them anyway — the most common complaint pattern in AI music reviews.

For a full criteria walkthrough with tool recommendations per question, see how to choose an AI music generator.

How does PrismPoster handle music rights?

PrismPoster's Music Studio generates original songs — vocals and lyrics in 35 languages, with stems — at 15 credits per song, inside the same studio as video generation and a timeline editor. You approve every track before it plays or publishes anywhere, and on paid plans you can release tracks publicly to Discover and PrismRadio, PrismPoster's own surfaces. Streamers get the approval-first live workflow described in AI music for streamers, and the broadcast terms define what that does and does not promise.

On ownership, PrismPoster's terms take the same position this article does: outputs pass through whatever rights the producing systems grant, and there is no copyright guarantee — because under current US doctrine, nobody can honestly offer one for purely AI-generated music. What is guaranteed instead is provenance: every export carries embedded machine-readable AI-provenance metadata, on every plan.

Where PrismPoster is the wrong choice for licensing reasons: if a client contract requires a copyright-transfer document, AIVA Pro is the tool built around exactly that. If your workflow is MIDI-first composition in a DAW, AIVA again — PrismPoster exports stems, not MIDI. And if you want a distribution pipeline to outside streaming services, note that PrismPoster's public release surfaces are its own. Where the studio earns its keep is everything around a finished track — how to make a music video with AI shows that path end to end.

Frequently Asked Questions

Who owns a song made with AI?

Usually no one holds a registered copyright in it. The vendor grants you a contractual license — sometimes broad, sometimes narrow — and under US Copyright Office doctrine, purely AI-generated music is not eligible for copyright, though human-written lyrics or recorded performances added to it may be.

Can I put AI-generated music on Spotify?

Only if your tool's license allows distribution, and many do not: Mubert prohibits streaming-service release on every plan, Soundraw gates it behind Artist tiers, and the leading consumer song platforms currently restrict downloads under their label settlements, as of July 2026.

Is training AI on copyrighted music fair use?

Undecided. As of July 2026 no US court has ruled on the merits; UMG and Warner settled their claims in late 2025, Sony Music is still litigating, and a summary-judgment hearing on fair use was scheduled for July 2026 in the District of Massachusetts.

What is a "walled garden" in AI music?

A platform where your creations live only inside the service — streamable and shareable, but not downloadable or releasable elsewhere. The label settlements of late 2025 pushed the biggest consumer AI music platforms in this direction, with downloads suspended on one and caps rolling out on the other.

Does subscribing later fix tracks I made on a free tier?

On the largest consumer platform, no — free-tier songs stay under the platform's ownership terms, and upgrading does not retroactively grant commercial rights, as of July 2026. Assume the tier you were on when you generated a track is the tier that governs it.