Elena VossSeven hundred was not a plaintiff count. In May 2024, Sony Music warned 700 AI companies, developers, and streaming platforms about unauthorized training use. It also invited licensing discussions.
Music labels now use AI copyright to govern generative media: which models can access valuable catalogs, which outputs can enter paid distribution, and who gets paid when software helps make music.
In 2020, OpenAI’s Jukebox drew attention because a neural network could produce raw audio with rudimentary singing in different genres and artist styles. By August 2026, Sony Music and Warner Chappell were suing Anthropic while Sony Music, Universal Music Group, and Warner Music Group were financing Stability AI. The industry had moved from asking whether AI could make music to asking which companies could do so with permission.
The lawsuits create a bid-ask spread
By pairing prohibition with an invitation, Sony Music turned a flat no into a negotiation over price. The major labels then supplied the downside: RIAA-led suits against Suno and Udio alleged mass infringement in model training, with reported statutory exposure of up to $150,000 per infringed work. Sony Music’s second lawsuit against Udio alleged unauthorized copying of 30,117 sound recordings.
Sony Music Publishing and Warner Chappell later pursued Anthropic over Claude’s training, alleging use of tens of thousands of copyrighted songs. Rights holders turned a general training claim into a work-by-work demand, naming songs, seeking documentation, and attaching potential damages to the absence of permission.
Suno and Udio argue that training on proprietary music qualifies as fair use. Courts may accept that defense, reject it, or draw narrower boundaries. The labels need not win in court to force price discovery. They need only create enough credible exposure that both sides must value an agreement.
A settlement can become a product specification
Universal, Warner, and Sony Music entered reported discussions with Suno and Udio about catalog licenses alongside settlements. The talks did not produce one industry-wide template. Suno’s negotiations with Universal and Sony Music later stalled because the labels argued that tools dependent on human-made music should pay.
Warner and Suno reached a different endpoint. Their strategic partnership settled prior litigation, called for licensed AI models, and included Suno’s acquisition of Warner’s Songkick service. UMG and Udio announced a subscription service planned around creation with licensed songs, partly resolving their copyright dispute.
A damages payment closes a historical claim. A creation license governs future conduct. It can specify eligible catalogs, participating artists, permitted transformations, subscription access, distribution rights, and a share of downstream revenue. Some music-AI arrangements reportedly borrow the revenue splits that labels already use for YouTube user-generated content.
Labels and model builders agree that access has commercial value. They have not agreed on the clearing price or the rights it includes.
Royalty eligibility requires a machine-readable chain
Once labels license selectively, platforms must know what entered the model, what the model produced, whether the user had permission, and whether the output qualifies for distribution or payment.
Sony Music said it had developed technology to identify copyrighted music inside AI-generated songs so rights holders could receive compensation.
Fingerprinting answers one question: what does this output resemble? A provenance credential answers another: where did this media come from? C2PA uses cryptography to encode origin information for audio, images, and video. Neither system alone decides who gets paid. Contracts must connect identity and origin to permission and royalty rules.
Together, these systems feed a payout ledger. Rights registries establish ownership; model records establish authorized access; output detection identifies possible use; platform policy decides commercial eligibility.
Tidal has already demonstrated the exclusionary version. The service tags wholly AI-generated music, blocks it from royalties, and removes AI tracks that impersonate artists. That policy makes provenance a condition of payment.
Platforms still face an unresolved middle. They can identify an AI-generated track without knowing which licensed works influenced it. They can know the training catalog without attributing a specific output. They can authenticate an output without deciding whether an imitation is commercially acceptable. To compensate creators, platforms must connect those measurements to contracts.
Catalog owners are financing the compliant lane
Stability AI raised a $76 million Series B from UMG, Warner, Electronic Arts, Sony Music, and other investors. Stability already had agreements with UMG, Warner, and EA to build models using their intellectual property. Stability presented its Stable Audio 3.0 family as trained on licensed data; the largest model could generate six-minute songs.
The same rights holders can sue one model builder and invest in another because the two positions hedge different exposures. Litigation limits uncompensated use. Equity captures part of the value created by permissioned use. The catalog sits on both sides of the transaction.
More than 100 million people had used Suno since 2023, and more than 2 million were paying subscribers even as UMG and Sony Music sued the company. The service had reached mass scale before the industry settled its access terms.
When software makes music-like output cheap, legally usable catalogs become scarce. The same shift appears in community data licensing: abundant raw material increases the premium on permission a buyer can verify.
Labels now package recording files with authorization to train, controls over generation, distribution access, and monetization eligibility. Stability’s financing puts capital behind that package instead of treating copyright as a bill that arrives after deployment.
Streaming’s unit of account is too coarse
Streaming trained the music business to license catalogs at platform scale and allocate revenue from a countable event. Generative music breaks that event into smaller decisions.
Spotify and UMG plan a paid Premium add-on that would let users create AI covers and remixes from participating UMG artists, although they announced no launch date. Spotify must distinguish participating from nonparticipating artists, authorized transformations from excluded ones, and generated use from ordinary playback.
Platforms may now assign permission and payment at the level of a model, prompt, stem, imitation risk, or completed work. Each can carry a different economic claim. A YouTube-style revenue split offers a starting template, not a complete ledger.
Adobe, Getty Images, Stability AI, and YouTube pledged to share generative-AI profits with creators but disclosed no payout amounts. Accurate labels can still produce unfair payments.
Sony Music’s 700 letters looked like perimeter defense when they arrived. The deals and investments that followed recast them as the first entries in a counterparty ledger. Sony Music’s 700 warnings marked 700 catalog doors awaiting a price.