Udio, Suno and Stability AI now state licensed music, while court filings still name tens of thousands of recordings and works allegedly used without it.
Figures here are the stating party's own claims or court records, dated in the sentence; the method is set out in how we source, date and check every figure.
This page treats music as a data type: its layers, the models that train on it, the filings, the routes and the licensing decision. Label deals and artist consent have their own page. The wider frame is AI training data licensing.
What Counts as Music Training Data: Recordings, Stems, Compositions, Lyrics and Metadata
Music training data comes in separable layers: the recording, its stems, the composition, the lyrics and the metadata, and each layer can carry a different rights holder.
The table compares the six forms of music data by definition, rights layer and where each appears on record (as of 2026-11-16, values checked 2026-11-09).
| Form | What it is | Rights layer and usual holder | Where it appears on record | Source |
|---|---|---|---|---|
| Finished recordings (masters) | the mixed and mastered audio of a track | sound recording; label or artist | Suno and Udio complaints; PLAI and STIM masters | court dockets; gema.de; stim.se |
| Stems and multitracks | the separate instrument and vocal tracks behind a mix | same as the recording | no acquirer on record states a stems requirement | n/a |
| Compositions (musical works) | the melody, harmony and structure as authored | composition; publisher, songwriter, society | GEMA v. Suno, six musical works; PLAI composition rights | justiz.bayern.de; gema.de |
| Lyrics | the words of a song, used as text | composition (lyrics); publisher, society | Concord I (lyrics in training and outputs); GEMA v. OpenAI, nine lyrics | SEC S-1/A; justiz.bayern.de |
| Production music | music made for licensing into media | composition and recording, often one holder | PLAI lists production music | gema.de |
| Metadata | titles, genres and descriptive tags per file | catalog owner | PLAI lists metadata and 60+ genres | gema.de |
A license names which layers it covers, so a recording licensed without its composition leaves half the rights open; the other data types are compared under data types AI companies buy.
Which AI Models Train on Music: Generators, Lyric-Trained Text Models and Multimodal Systems
Three kinds of AI models train on music: music generators that learn from recordings, text models whose training included lyrics, and multimodal models that declare audio among their training data.
- Music generators: Suno and Udio build models on recordings; Stability AI states that Stable Audio 3.0 is "trained on fully licensed data" (Stability's claim, 2026-05-20), while its agreements with UMG (2025-10-30) and WMG (2025-11-19) name tools, not training.
- Lyric-trained text models: the Concord I publishers allege lyrics in Claude's training and outputs (N.D. Cal. 5:24-cv-03811-EKL), and the Munich court held memorized lyrics in GPT-4 and GPT-4o to be reproduction (judgment of 2025-11-11, not final).
- Multimodal systems: SpaceXAI's Grok 4.7 summary (2026-09-21) puts its audio training data at "More than 1 million" hours and marks commercial licenses "Yes" for text and image only.
The audio band does not separate music from speech, and the Commission's summary template asks only whether licensed data was used, per modality, so licensed music would show as "Yes" for audio.
Regulation (EU) 2024/1689 Article 53(1)(d) requires that summary from general-purpose model providers, and its obligations apply from 2025-08-02, as set out in the AI Act training data summary.
How Much Music Court Filings Say Went Into Training
Five filings put a number on the music at issue, from 60,202 recordings in the labels' second action against Suno to six works in GEMA's Munich case.
The counts are works or recordings that plaintiffs assert or a court decided on, not the size of any training set; no filing on record states how large Suno's or Udio's training data is.
The table lists the five filings by count, with the acquisition or retention as alleged or held and the status as of 2026-11-09.
| Case, court and docket | Music data at issue | Count | Acquisition or retention, as alleged or held | Status (as of 2026-11-09) | Source |
|---|---|---|---|---|---|
| UMG Recordings et al. v. Suno, D. Mass. 1:24-cv-11611-FDS; second action 1:26-cv-14275 (filed 2026-09-18) | sound recordings | 60,202 recordings (second action) | copying recordings to train; YouTube stream ripping under DMCA 1201, alleged | pending, discovery; Warner dismissed (Doc. 174) | complaint; Doc. 174 |
| Sony Music Entertainment et al. v. Uncharted Labs (Udio), S.D.N.Y. 1:24-cv-04777-AKH; second action 1:26-cv-06120-AKH | recordings and works | 30,117 works (second action, per Udio's answer, Doc. 25, 2026-08-27) | training on recordings; YouTube rolling cipher under DMCA 1201, alleged | pending for Sony; UMG and Warner settled | complaint; Doc. 25 |
| Concord Music Group et al. v. Anthropic PBC (Concord II), N.D. Cal. 5:26-cv-00880-EKL (filed 2026-01-28) | musical works | 20,517 works | allegedly used to train Claude, including acquisition by downloading | pending, motion-to-dismiss stage | complaint; Dkt. 141 |
| GEMA v. OpenAI, LG München I, 42 O 14139/24 (judgment 2025-11-11) | lyrics | 9 lyrics | memorization in GPT-4 and GPT-4o held to be reproduction, § 44b not covering it | not final | court press release 11/2025 |
| GEMA v. Suno, LG München I, 42 O 763/25 (judgment 2026-07-31) | musical works, lyrics not at issue | 6 works | memorization in v3.5 and v4 on servers in Germany held to infringe | not final | court press release 16/2026 |
The acquisition alleged is YouTube stream ripping or a rolling cipher under DMCA section 1201 (Suno and Udio) and downloading (Concord II); these are allegations, and none of the three actions is decided.
Who Licenses, Acquires or Handles Music Data
Music reaches AI training through AI music developers that take licenses, societies that license on their members' behalf, and apps and platforms that collect or list catalogs.
The table groups ten parties by role, alphabetically within a role, with what each licenses, acquires or handles (as of 2026-11-16, values checked 2026-11-09).
| Company | Role for music | What it licenses, acquires or handles | Source |
|---|---|---|---|
| ElevenLabs (Eleven Labs Inc.) | AI music developer with a label agreement | licensed AI music platform with UMG, 2026-09-10 | universalmusic.com |
| Klay Vision Inc. | AI music developer with a label agreement | agreement with WMG, 2025-11, variable economics | WMG 8-K exhibit |
| Stability AI Ltd | AI music developer with label agreements | Stable Audio 3.0 "trained on fully licensed data" (Stability's claim); tools agreements with UMG and WMG | stability.ai |
| Suno, Inc. | AI music developer, licensee | new models on licensed music under the WMG agreement, 2025-11-25 | suno.com |
| Udio (Uncharted Labs, Inc.) | AI music developer, licensee | recorded-music and publishing licenses from UMG; license from WMG | universalmusic.com; Dkt. 139 |
| GEMA | society, licensor on members' behalf (not a buyer) | PLAI curated music dataset | gema.de |
| STIM | society, licensor on members' behalf (not a buyer) | opted-in works and their master recordings | stim.se |
| Kled AI (Nitrility, Inc.) | contributor app, licensee of uploads | music among photos, video and transcriptions | kled.ai terms |
| Credtent | licensing platform, lists catalogs on rights holders' behalf | catalogs including music | credtent.org |
| Troveo | video intermediary | excludes unlicensed music from footage it licenses | troveo.ai |
Kled takes uploaded music under its own license, which is exclusive to the company for developing and training third-party AI models, so an uploader cannot license the same file elsewhere for that use.
Troveo's seller guide excludes unlicensed music, so soundtracks are cleared separately from the video (checked 2026-11-09).
Routes From a Music Catalog to a Training Set: Label Agreements, Society Licenses, Curated Datasets and Apps
A catalog reaches a training set by one of five routes, and they differ most on who signs, what is delivered and whether the license can be narrowed later. The label-by-developer deals and the society terms are covered in full on the page for musicians, labels and publishers licensing music.
The table compares the five routes from the most to the least centralized signer, each value being the stating party's own statement (as of 2026-11-16, checked 2026-11-09).
| Route | Example on record | What is delivered or licensed | Scope and exclusivity | Pay basis stated | Source |
|---|---|---|---|---|---|
| Label agreement | UMG and Udio (2025-10-29); WMG with Suno, Udio, Stability AI, Klay | recorded-music and publishing licenses; terms not disclosed | not disclosed | WMG: "all of our deals are based on variable economics" | universalmusic.com; WMG 8-K |
| Society opt-in license | STIM collective AI license (from September 2025) | opted-in musical works and corresponding masters; attribution technology required | only works whose creators opted in; work-by-work approval or exclusion from 2026-06-01 | fixed license fee plus a percentage of the AI service's income; no rates | stim.se |
| Curated society dataset | PLAI by GEMA (GEMA states it has offered an AI licensing model since 2024-09) | ca. 57,000 works, ca. 178,000 audio files, 60+ genres, composition and master rights, metadata | curated and adaptable per customer; delivered exclusively to each customer | contributors paid "anteilig und angemessen" (proportionately and appropriately); no figures | gema.de |
| Catalog listing platform | Credtent content licensing | catalogs of rights holders, music included | standard terms; other terms not stated | annual training fee plus revenue share; 85 % to rights holders under standard terms, paid quarterly (Credtent's claim) | credtent.org |
| Contributor app | Kled data contributor app | individual uploads, music included | transferable, fully sublicensable, worldwide, perpetual, irrevocable; exclusive for training third-party AI | rewards in fiat at Kled's discretion or third-party tokens, no guaranteed value | kled.ai terms (revised 2026-07-21) |
Exclusivity runs both ways: GEMA delivers PLAI exclusively to each customer, while Kled takes the uploader's grant exclusively for itself.
Only STIM lets a creator narrow the grant work by work, since its members approve works one at a time or exclude the AI rights category from 2026-06-01. The same comparison across other program types is kept in the program terms compared, dated.
Pricing a Music Catalog as Training Data: Shares, Fees and Undisclosed Terms
No per-track, per-hour or per-work price for music training data appears in any filing, release or society page on record; what parties state is a pricing basis. WMG states that "all of our deals are based on variable economics" (8-K exhibit, 2026-03-03), and UMG's release on Udio gives no terms.
STIM states a fixed license fee plus a percentage of the AI service's income, with no rates. GEMA states that contributors are paid "anteilig und angemessen" and publishes no figures, and Credtent states 85 % to rights holders under standard terms (Credtent's claim, checked 2026-11-09).
Kled pays rewards in fiat at its discretion or in third-party tokens with no guaranteed value. Unit prices stated for other data types are collected under AI training data prices per work, hour, image.
Preparing a Music Catalog Before It Is Licensed
A catalog is made licensable by establishing who holds each right and describing the content in countable terms.
Splitting Master, Composition and Lyric Rights per File
Every file in a music catalog carries at least two rights, the recording and the composition, and a license covers only the layers the licensor controls. PLAI offers both composition and master rights, and Troveo's guide shows the reverse: a soundtrack the footage owner does not hold is left out.
Under UK law the employer is first owner of a musical work made by an employee in the course of employment, subject to any agreement to the contrary (Copyright, Designs and Patents Act 1988 s.11(2)). Co-written works and samples are checked holder by holder, and the steps are set out in the rights audit.
Describing the Catalog: Works, Audio Files, Stems, Genres and Metadata
A catalog is described to a licensee by counts and rights, the way GEMA describes PLAI: number of works, number of audio files, genres, rights per layer and metadata.
- Works: the number of distinct compositions
- Audio files: the number of recordings, including versions
- Genres: the labels the catalog carries
- Rights held per layer: recording, composition and lyrics
- Metadata: titles, tags and identifiers, with a flag for whether separate stems or multitracks are available, yes or no
GEMA gives about 57,000 works and about 178,000 audio files for PLAI. A data manifest describing a dataset records them without handing the files over.
Attribution and Usage Reporting After Delivery
Licenses that pay by use need a way to trace an output back to the works that influenced it, and STIM makes independent attribution technology a condition of its license.
STIM states that creators receive royalties during model training, during use of the AI service and when the music it generates is used (STIM's claim, checked 2026-11-09).
Credtent pairs its annual training fee with a revenue share. The models are compared under attribution and per-use payment models.
Licensing Music Data: Benefits, Risks, Downsides and Alternatives, Including Not Licensing
Licensing a catalog as training data can turn unused recordings into recurring income, but the money is rarely disclosed, some grants are perpetual and a trained model cannot give the music back.
Benefits
- STIM states royalties at three points: during training, during use of the service and on generated music.
- WMG states that its AI deals rest on variable economics, so income can follow usage.
- PLAI licenses composition and master rights together and pays contributors proportionately, according to GEMA.
Risks
- Models retain works: the Munich court held memorization in the model to infringe in GEMA v. OpenAI and GEMA v. Suno (both not final).
- Kled's grant is perpetual and irrevocable.
- Licensing does not stop others copying: Concord II alleges acquisition by downloading.
Downsides
- No route publishes rates; PLAI states no figures and STIM states no rates.
- Credtent's 85 % is its own claim, and its other terms are not stated.
- Work-level control exists only at STIM among the five routes.
Alternatives
- An opt-in society license, as STIM offers
- A curated dataset, as PLAI offers
- A machine-readable reservation of rights under Directive (EU) 2019/790 Article 4(3), which Germany applies in § 44b(3) UrhG
- Credtent's free opt-out notices, with follow-ups at 30, 60 and 90 days
- Not licensing at all, weighed under whether to sell or license your data
The right option depends on which rights the holder controls.
Music Next to Speech, Voice, Footage and the Seller Steps
This page covers music as data; the label deals, society terms and artist consent behind a license, speech corpora, a singer's own voice and footage soundtracks each have their own page. This page is information, not legal advice.
Neighboring pages:
- Collective licensing for AI training: societies and their terms
- Speech and audio data: spoken-word corpora
- Selling your voice to AI: voice cloning licenses
- Selling video for AI training: footage and soundtracks
- Text corpora for AI training: lyrics licensed with text
- Voice and likeness rights in AI training data: publicity law
- Copyright and AI training across jurisdictions: fair use and TDM
- EU text and data mining exception: the opt-out
Trackers: AI copyright lawsuits over training data and AI data licensing deals, with Universal Music Group AI licensing deals and Warner Music Group AI licensing deals for the label deals.
Seller steps for music:
- Read how to sell or license company data for the whole process.
- Check the clauses in AI data licensing agreements.
- Learn when AI data buyers pay.
- Keep the chain of rights as set out in training data provenance.
Frequently Asked Questions
What music do AI companies train on?
Recordings, compositions and lyrics. The filings on record name 60,202 recordings against Suno and nine lyrics against OpenAI, while Udio, Suno and Stability AI now state licensed music (as of 2026-11-09).
Is there a licensed music dataset for AI training?
Yes: GEMA offers PLAI, about 57,000 works and 178,000 audio files with master and composition rights, delivered exclusively per customer. STIM licenses opted-in works. Open and scraped datasets are listed under major AI training datasets.
Can musicians upload songs to apps that pay for AI training data?
Kled accepts music, but its terms take an exclusive, perpetual and irrevocable license and pay rewards at its discretion. Recordings signed to a label are usually not the uploader's to license. Other options for individuals are listed under creators selling content for AI training.
Advisers and agents can use the introducer form, and new music deals arrive as deal tracker updates by email.