The AI Music Rights Stack

This is the pillar page for everything Clarity Unlocked publishes on the AI music rights beat. It anchors the cluster — every other piece on this topic links back here. If you're new to the subject, start with the original argument. If you came in through a settlement headline, jump to what the deals actually did. If you run a label, publisher, or distribution operation, jump to the pre-flight rights stack.

Why this matters now

The music business has spent two years arguing about whether AI music is "real." That argument is now over and it ended in a place nobody on either side predicted: the products are good enough that the labels are settling, the rulings have not arrived, and the infrastructure that would make any of this auditable does not exist.

This page exists because every operator I talk to — label execs, publishers, sync supervisors, lawyers, independent artists, the people actually running release calendars — is asking the same set of questions, and the answers are scattered across press releases, court filings, podcast interviews, and Discord servers. Pull them together, the picture is unambiguous: the platform that wins the next decade of music is not the one that ships the best generator. It is the one that knows who is allowed to make the song in the first place.

What the 2025–2026 settlements actually did

The numbers and dates that matter:

Warner Music + Suno · November 2025. Warner and Suno reached a licensing deal that requires Suno's training corpus to be licensed going forward and adds a paywall for users who want to download the tracks they generate. The product experience does not change dramatically — what changes is the legal substrate underneath it. Suno is now the platform that can credibly say its catalog is cleared, which is the entire moat in 2026.

Universal Music Group + Udio · October 2025. UMG's settlement with Udio is more architectural. It converts Udio from a free-text generator into what the company is calling a "fan engagement platform" — users remix, mash up, and prompt in the style of licensed artists, but the outputs cannot leave the platform. UMG is co-launching a licensed AI music platform in 2026 that runs on the same plumbing. The walled garden is now the product. It is also, not coincidentally, the model that the labels have always wanted: a downstream surface they control.

Sony Music · neither. Sony has not settled with Suno or Udio. The fair-use cases — brought by RIAA on behalf of the major labels — are heading toward a ruling expected in summer 2026. Whoever wins that ruling sets the precedent every AI music platform has to navigate for the next decade.

Independent musicians · October 2025 class actions. Independent musicians filed class actions against both Suno and Udio in October 2025, seeking to represent thousands of artists whose work was allegedly used in training. The class actions matter because they expose the gap the label settlements left: the labels can sign deals on behalf of their rosters, but they cannot license catalog they don't own, and a lot of what these models trained on belongs to writers and performers without major-label representation.

The opt-in licensing math problem

Every settled deal so far is opt-in. Artists and songwriters individually decide whether to license. This sounds reasonable. It is structurally fragile.

A modern pop song has anywhere from one to ten or more named songwriters. Hip-hop tracks frequently credit producers, sample sources, and feature artists into the writers' room. If a single dissenting writer on a 6-writer song declines a license, the entire song is disqualified from use under the deal. That is not a hypothetical — it is the math.

The result is that even the most successful licensing deals (Warner-Suno, UMG-Udio) cover a much thinner slice of the catalog than the press releases imply. A label can sign on behalf of its master rights. It cannot sign on behalf of every co-writer's publishing share. Every song has to be cleared at the song level, by every named contributor, before it can be used.

This is the part of the picture that the headlines miss. The settlements look like the labels rolling out the carpet. The catalog math underneath them looks much more like a thin strip of approved territory inside a much larger fenced zone of technically licensed, practically unusable material.

The pre-flight rights infrastructure that has to exist

Here is what every label, publisher, and operator I've spoken to wants — and what currently nobody is selling as a turnkey product.

A pre-flight rights system would do five things, in order, before a track ships:

1. Provenance verification. Document what trained the model that produced the audio. Was it a licensed corpus (Suno-W, Udio-UMG, future Sony deal) or an unlicensed scrape (most of the open-source models, Common Crawl-trained derivatives)? Provenance is the upstream question that every downstream question depends on.

2. Reference clearance. Document what prompted the model. Style references, audio uploads, lyric prompts, vocal cloning sources — every input that shaped the output. If the prompt said "in the style of Drake," that is a fact the rights stack needs to know.

3. Songwriter and contributor mapping. Document every human contributor whose share survives the release: producer, performing artist, lyricist, mix engineer, the operator who supervised the session. This is the cap table.

4. Release-readiness audit trail. Generate a signed, timestamped document — call it an RRA, release-readiness attestation — that names provenance, references, contributors, and licensing posture for every track. The RRA is the artifact that lives in the file the label, the DSP, and the litigation lawyer all want to see two years from now.

5. Distribution-side enforcement. DSPs and aggregators need to check the RRA before accepting a track. Today they check ISRC and metadata. They will eventually have to check rights provenance the way an ad network checks brand safety. That layer does not exist yet either.

Some of this infrastructure exists in pieces. Content ID is a downstream version of distribution-side enforcement. PROs have decades of contributor mapping data. Tools like Audoo and Pex do reference detection. What does not exist is a coherent pre-flight layer that bundles all five into a workflow a normal operator can run before hitting the upload button. Building that bundle is the trillion-dollar product opportunity sitting in plain sight.

The cap table for the song

If you take nothing else from this page, take this: every AI-assisted song needs a cap table, and the cap table needs to be approved before the song ships.

The cap table names everyone whose share survives the release:

• The songwriters whose underlying compositions show up in the prompt, the style reference, or the training data the model was tuned on
• The master rights holders whose recordings the model trained on (and the deal under which it trained — Warner-Suno, UMG-Udio, or unlicensed)
• The producer who finished the track
• The performing artist or operator who released it
• The AI platform whose model produced the audio
• The editor or A&R who signed off on distribution

Every line on that cap table needs two things: a documented permission and a documented royalty share. Today most AI-assisted tracks ship with zero of those documents. That is the entire liability surface. The work to fix it is not glamorous and not technically hard. It is overdue.

What the summer 2026 fair-use ruling does

The pending fair-use ruling is the first judicial test of whether training a generative music model on copyrighted recordings is fair use under U.S. copyright law. Two scenarios, both fully in play as of this writing:

Scenario A — creator-favorable. The court rules that training on unlicensed copyrighted recordings is not fair use. Every model trained on unlicensed catalog has to either pivot to an opt-in licensed corpus (the Suno-W and Udio-UMG path) or shut down generation on the disputed material. The settled deals become the floor, not the ceiling. Sony's holdout position becomes leverage to sign a deal on better terms than Warner or UMG got. Pre-flight rights infrastructure becomes a compliance requirement, not a nice-to-have. Independent musicians' class actions get fast-tracked toward statutory damages.

Scenario B — platform-favorable. The court rules that training is sufficiently transformative to be fair use. Suno's and Udio's settled deals look, in hindsight, like commercial accommodations they didn't have to make. The fight shifts to Congress — pressure to pass an AI-era copyright reform builds, but the legislative clock is multi-year. Sony's position weakens; the unlicensed open-source models proliferate; the labels' leverage erodes. Pre-flight rights infrastructure still gets built, but now as a commercial differentiator (the cleared-catalog platform earns the premium) rather than a compliance requirement.

Either way, the operators who already have an answer to "what did your model train on, what prompted this output, and who is on the cap table" are positioned for the next decade. The operators who don't are betting their release pipeline on whichever way the court rules.

What to do this quarter (operator checklist)

If you run a label, a publishing operation, a distribution business, or a single-operator music IP studio, here is the move set for Q2–Q3 2026:

1. Inventory your model provenance. For every AI tool in your workflow, document what it trained on, under what license. If a tool can't answer, treat that as a risk flag and source an alternative.

2. Build a cap-table template. One per release. Six fields: provenance, references, songwriter share, master share, AI platform share, editor signoff. Make it a hard prerequisite for distribution.

3. Get on the right side of the opt-in. If you control any of your own publishing or master rights, decide your opt-in posture now and document it. The first labels and publishers to publish clear AI-licensing policies are going to be the easiest to deal with — and the ones operators steer toward.

4. Pick a fair-use ruling stance. Pre-write what you'll do under Scenario A and Scenario B. If the ruling lands and you have to scramble to figure out which model is safe to use, you are already behind.

5. Track the cap table forward. The shape of the rights infrastructure you build in 2026 is going to determine how much of the upside you keep when this category matures. Building it as a one-line item on a release checklist today saves you from building it as a litigation response in 2027.