# Whose Song Is This? AI Music Broke the Question Platforms Cannot Stop Asking

> TikTok and Universal keep re-signing the same promise to remove unauthorized AI music, which reveals a system that can no longer answer its core question. The opportunity is the layer that decides what a sound is, who controls it, and what a platform may do with it.

**Published:** 2026-07-15
**Canonical URL:** https://yihuisong.com/article/ai-music-copyright

On May 22, 2026, Universal Music Group and TikTok [announced a new multi-year global licensing agreement](https://newsroom.tiktok.com/universal-music-group-and-tiktok-announce-new-global-licensing-agreement?lang=en). The deal matters, but the AI language matters more. The companies committed to "remove unauthorized AI-generated music from the platform, while further improving artist and songwriter attribution."

This is the second major agreement between these two companies in just over two years. The previous one, in [May 2024](https://newsroom.tiktok.com/universal-music-group-and-tiktok-announce-new-licensing-agreement?lang=en), ended a three-month standoff during which UMG had pulled its entire catalog from TikTok. That deal carried the same promise, nearly word for word. TikTok committed to working with UMG to remove unauthorized AI-generated music from the platform, along with tools to improve artist and songwriter attribution.

So the 2026 sentence is a repetition with one word added. "Further improving" concedes that the improvement is unfinished. Obligations get renegotiated when they are hard to meet, and for a growing share of the audio flowing through the platform, the system can no longer treat "whose song is this?" as a settled question.

TikTok answers that question billions of times a day. Every video with music in it depends on the answer, and the platform has to produce one before the video goes live. Before tracing how the fight got here, it is worth slowing down on the machine that produces those answers.

## How Music Rights Become Product Decisions
### The Bar Pays and Stops Asking
The machine runs on rules written long before platforms existed. The easiest place to see them is a bar on a Friday night, where a band is covering a song everyone in the room knows.

Someone is getting paid for this. Who, and for what?

Every song is two separate pieces of property. The composition is the song as written, the melody and the lyrics, and that copyright belongs to the songwriters and usually to a publisher who administers it. The master recording is the specific recorded performance, and that copyright typically belongs to a label, or to the artist if they kept it.

The band is making its own sound in the room, so it touches the composition and nothing else. When a DJ plays the original track, both copyrights come into use.

This is also why the industry says "rights holder" instead of "copyright owner." The owner, the administrator, and the party that collects are frequently three different entities. The safer mental image is an institution, not a singer.

The bar doesn't handle any of this. It pays a flat annual fee to the performing rights organizations and is covered for everything in their catalogs. The bartender never determines which song is playing. That blanket license is a deliberate abstraction, and what it abstracts away is the question itself.

![Figure 1. Who gets paid when music plays in a bar?](/images/ai-music-copyright/bar-royalties.jpg)

The bar pays a flat fee and stops asking. That option is not available to a platform.

### Every Sound Is a Decision
A blanket license covers a room, but it cannot route a payment. The platform has to precisely pay per use, attribute per rights holder, and enforce per territory, so it has to know what every sound is before it can do anything with it. Here is how its music platform machine works:

* First, it has to identify what the audio actually is. A licensed track from a label's catalog is one thing. User-uploaded original audio is another. A cover, a remix, a sped-up version, a stem-isolated edit, or something a model generated are all different again, and they arrive looking roughly the same: a waveform attached to a video.
* Next, that sound has to be mapped onto rights: which composition it contains, which master, who holds each one, in what territory, for how long, and under which negotiated terms.
* Then the mapping becomes product behavior:
  * The sound goes into an authorized library or it doesn't.
  * It gets recommended or suppressed, monetized or not, and if it is monetized the money routes to specific parties in specific proportions.
  * It plays in one country and stays muted in another, because the underlying deal differs by market.
  * When a label, a publisher, an artist, or a creator objects to any of it, the platform needs a process, a record, and an answer.

![Figure 2. Who gets paid when music plays on TikTok?](/images/ai-music-copyright/tiktok-royalties.jpg)

TikTok operates this as product infrastructure. The system identifies each sound, attributes it to rights holders, authorizes what can be done with it, and enforces the result, continuously, against everything that gets uploaded.

This is the trade that makes short-form video work. A creator opens the app, taps a track, and the song is just there, cleared, attributed, and paid out to the right parties in the background. That experience is the output of a resolution layer doing enormous unseen work. It only holds as long as the layer can answer its central question reliably:

*What is this sound, who owns or controls the rights to it, and what is the platform allowed to do with it?*

## When the Music Industry Meets a New Platform
### The Catalog Pull
UMG's licensing agreement with TikTok [expired on January 31, 2024](https://www.prnewswire.com/news-releases/universal-music-group-agreement-with-tiktok-to-expire-on-january-31-2024-302048628.html), and UMG did not renew. It pulled its catalog instead. Music from artists like Taylor Swift and Billie Eilish came off the platform, and videos built on those tracks went silent, which cost TikTok a meaningful piece of its library for roughly three months. It also cost UMG the discovery engine its own artists depend on.

This is the only real move a rights holder has, and it cuts both ways, which is what makes the threat credible.

What UMG bought with it is the interesting part. The day before the contract expired, UMG [published an open letter](https://www.universalmusic.com/an-open-letter-to-the-artist-and-songwriter-community-why-we-must-call-time-out-on-tiktok/) naming three grievances. It said it had been pressing TikTok on appropriate compensation for artists and songwriters, on protecting human artists from the harmful effects of AI, and on online safety for TikTok's users. Here is how I read them:

* The first was the oldest complaint in the business. Rate disputes settle on a number, and that part of the letter could have been written in 2006 about YouTube.
* The second was new, and it cannot settle. There is no number that finishes AI protection, so it returns at every renewal, which is what the 2026 deal shows.
* The third was not a term at all. It widened the audience. A rate dispute is a conversation between two companies. A safety complaint is addressed to artists, press, and regulators, which is why the letter ran the day before the deadline rather than the day after.

The standoff [ended on May 2, 2024](https://newsroom.tiktok.com/universal-music-group-and-tiktok-announce-new-licensing-agreement?lang=en), with a deal that restored the catalog and added what TikTok called "industry-leading protections with respect to generative AI." That closed the rate dispute. I read the AI language as the more durable win, because it established AI protection as part of what a platform pays for.

Two years later TikTok and UMG signed again, and AI was still the headline. That was the return on the pull. UMG played the sequence well, and the sequence itself is older than this fight.

### Why the Industry Defends First
UMG's 2024 pull was the industry's standard move in a new technology cycle.

Napster made other people's music collections instantly copyable. The labels sued, Napster shut down in 2001, and the old album-sales business never fully recovered.

Streaming offered all the music for a monthly fee instead of a purchase. The labels resisted, Spotify did not launch in the US until 2011, and once streaming became unavoidable the fight moved to the rate per stream.

Short-form video turned songs into raw material for other people's content. The labels licensed faster this time, but they kept catalog access and takedown pressure in hand. The 2024 catalog pull was that leverage at full volume.

The cycle repeats: technology changes how music circulates, the industry defends, and the settlement trades distribution control for rights leverage.

The industry does not defend by accident. The response comes from how the business is built.

* The three major label groups control [roughly two thirds](https://www.mordorintelligence.com/industry-reports/music-market-landscape) of global recorded music revenue. That concentration is why one company's catalog pull could silence a platform. Few industries let a single participant do that.
* What those companies sell is permission: the right to say yes or no to someone else's use of a catalog. A company that sells units wants volume. A company that sells permission wants control over the conditions, which means it defaults to no until the terms are right.
* Underneath that is a data problem the industry has never fully solved. Music rights metadata is messy. Ownership splits across parties and territories. Collection and enforcement run through labels, publishers, PROs, and distributors whose records do not always agree. The industry is less digitized than the platforms that now depend on it.

Taken together, this makes the industry's defensiveness look rational. The rights system is built around permission, and its infrastructure was never built for platform-scale ambiguity. When something increases ambiguity, the trained response is to withhold permission until the ambiguity is priced.

In every prior platform shift, the industry still knew what the underlying work was. The fight was over distribution, payment, or usage. AI-generated music changes the fight because the identity of the work itself becomes unstable.

### Why AI Breaks the Playbook
Those fights were all about existing recordings. Whatever else was contested, the object itself was legible. There was a composition, a master, and parties who owned each. The argument was about what happened next.

AI-generated audio does not reliably produce that object. A track can sound like a specific artist without copying any specific song. A synthetic voice can carry an unmistakable resemblance to a singer without touching that singer's master recording. A model can be trained on a catalog and generate output with no traceable fragment of any individual work in it. A creator can make something genuinely new that still raises questions about likeness, attribution, training data, and derivative work, all at once, with no clean answer to any of them.

It helps to separate two failures that get discussed as one.

The first is *unknown origin*. The platform cannot determine what copyrighted works, if any, sit behind a generated output. The provenance is simply opaque. This is a detection problem that gets harder as generation gets cheaper.

The second is *known imitation*. The platform can tell perfectly well that a track is trying to sound like a specific artist. The intent is legible to any listener. At the launch of this year's IFPI Global Music Report, Sony Music's Dennis Kooker said the company had [requested removal of more than 135,000 AI deepfake tracks](https://www.billboard.com/pro/ifpi-global-report-2026-highlights-vinyl-china-ai-deepfakes/) impersonating its artists, among them Beyoncé and Harry Styles. Everyone understands what the track is trying to do. What's missing is a clean copyright category to hang the claim on, because sounding like someone has never been the thing copyright was built to adjudicate. This is a new category problem, which does not get solved by better classifiers.

Both failures land in the same place: a video is being uploaded, the platform has to make a decision, and the rights system may not have an answer ready.

![Figure 3. What happens when AI music enters the pipeline?](/images/ai-music-copyright/ai-pipeline.jpg)

The blanket license already lets the bar stop asking whose song this is. It pays a flat fee and routes around the question, so if an AI-generated track turns up in its playlist, the bar's exposure does not really change.

TikTok cannot stop asking, because answering per song is the whole model. AI-generated music breaks the exact question the platform is built to answer. The system that resolves music identity at scale is now being fed audio whose identity may not resolve.

## What the Commitment Requires
So look again at the 2026 deal. TikTok and UMG committed to removing unauthorized AI-generated music from the platform and improving artist and songwriter attribution. Both halves of that sentence run through the same machine.

### Running the Promise Through the Machine
To remove unauthorized AI-generated music, the system has to determine that a track is AI-generated, and then that it is unauthorized. The first is the identification stage, and it is where *unknown origin* bites. The second is the rights mapping stage, because "unauthorized" only means something if the system can identify whose authorization was needed. That is where *known imitation* bites.

Improving attribution depends on both. A system that cannot identify the sound or map the rights cannot reliably say who should be credited. That is a real commitment made against a capability that is still being built.

### The Tools Only Go So Far
Regulators have moved first, and disclosure is the piece with a date on it.

The EU AI Act's [Article 50](https://artificialintelligenceact.eu/article/50/) requires providers of AI systems that generate synthetic audio, image, video, or text to mark their outputs in a machine-readable format detectable as artificially generated or manipulated. Those transparency obligations apply from August 2, 2026, while generative systems already on the market before that date get [until December 2, 2026](https://artificialintelligenceact.eu/transparency-rules-article-50/) to meet this requirement under the May 2026 provisional agreement.

A mandate can mark the audio. It does not tell the platform what to do when an unmarked track arrives, how to weigh a marked track against a rights holder's objection, or who to pay when the marking says generated and a label says that is our artist's voice.

Provenance standards fill part of the gap. The [C2PA specification](https://spec.c2pa.org/specifications/specifications/2.4/explainer/Explainer.html) defines Content Credentials, signed manifests that travel with a file and record the asset's origin and edit history, covering audio alongside image and video. The European Commission's [draft Code of Practice on Transparency](https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content) points to them as an example mechanism for satisfying the marking obligation. The direction is a real shift, away from detecting what is fake after distribution and toward proving what is real at creation.

Provenance is narrower than it first appears, and the specification is honest about this. Content Credentials do not assert that provenance claims are true. They assert that the claims are well-formed, untampered, and signed by someone on a trust list.

A valid manifest can travel with a track whose training data is contested, whose voice resembles a living artist, and whose rights holder is disputed. Provenance tells you where a file says it came from, it does not tell you who owns what is inside it.

### Rules Ahead of Law
Marking helps the platform identify a file's declared history. It does not settle rights-mapping challenges.

Sony Music asked streaming platforms to take down more than [135,000 songs](https://www.musicbusinessworldwide.com/sony-music-has-targeted-135000-deepfakes-of-its-artists-music-for-removal-from-streaming-platforms/) it said were created with generative AI to impersonate artists on its roster, including Beyoncé and Harry Styles. These were enforcement demands against specific tracks rather than vague objections to AI. Existing copyright categories do not handle those claims cleanly, but platforms are still the only parties positioned to act on them at scale.

Whatever gets enforced in that gap becomes the working rule. It is not law or a settled doctrine yet. It is the platform, the rights holder, and the policy team making the least-bad decision before the legal system catches up.

That is why it would be a mistake to read the 2026 TikTok-UMG agreement as a reconciliation. UMG needs platforms to enforce boundaries at a scale it cannot reach alone. TikTok needs workable rules that do not suppress the creator behavior its product depends on. Both are still negotiating from different incentives.

What changed is that the old tools no longer reach the whole problem. A label can withhold permission when the dispute is over a known catalog. A platform can detect infringement when an upload matches a known reference file. AI-generated music is harder because the dispute may not trace back to one song, one master, or one rights holder whose permission would settle it.

That mutual shortfall is what puts both parties in the same room. The 2026 deal reads less like a resolution than a forced attempt to keep music usable on the platform while the rules for AI-generated audio are still being written.

## Where the Question Becomes Product
Regulators can write disclosure rules. Labels can renegotiate contracts. But the platform still has to decide what happens when the next sound is uploaded.

So the challenge moves into the product layer, where unsettled law shows up as product decisions. A few questions show the shape of the problem.

First, what similarity threshold makes an AI-generated track unauthorized? Melodic overlap is measurable. Style, timbre, and vocal resemblance are not. Somewhere between those categories, platforms and rights holders will draw a line that decides how much AI music gets blocked, licensed, or allowed to circulate.

Second, what counts as a rights object? In the old model, the platform mostly had to resolve the composition and the master. AI adds pressure to treat voice, likeness, provenance, and training history as fields the system may need to track. If that happens, the music layer needs more than better detection. It needs new fields.

Third, who owns platform-native AI music? If a creator generates a track inside a platform tool using a model trained on licensed catalog, the output has multiple plausible claimants and no obvious default. The creator is no longer just carrying music from labels into videos. The creator may become part of the rights chain.

The next fight may not be over royalty rates. It may be over who gets to define, verify, and control what a piece of music is on a social platform.

Whoever owns that layer owns the answer to the platform's most important question:

***Whose song is this?***