AI Music Digest: Catch-Up Edition

Jamendo, the Luxembourg-based music licensing platform owned by Winamp Group, filed two US copyright suits in one week this June over the same body of music. On June 22 it sued Nvidia in the Northern District of California, alleging the chipmaker trained its audio models Fugatto and Audio Flamingo on the MTG-Jamendo Dataset — roughly 55,000 tracks, about 919 hours of audio, released around 2019 for non-commercial research use. On June 29 it sued Suno in federal court in Massachusetts, alleging the same dataset (and individual tracks) were used to train Suno’s models, including its early open-source model Bark. Both complaints run on the same theory: the dataset’s Creative Commons non-commercial license required a paid license for commercial use, and neither company took one. Against Suno, Jamendo sought at least €17.8 million (about $20 million) or statutory damages of up to $150,000 per work if willfulness were found; Winamp Group CEO Alexandre Saboundjian said “innovation and respect for intellectual property must go hand in hand to ensure a sustainable ecosystem.”

Then, six weeks later, the Suno case ended without an answer. On August 13, Jamendo filed a notice of voluntary dismissal — without prejudice, meaning it can refile the identical claims later. Per Music Business Worldwide, the notice gave no reason and made no mention of a settlement, and none of the parties responded to requests for comment. Plagiarism Today noted Jamendo could dismiss unilaterally because Suno had not yet responded to the complaint at all. The Nvidia case remains active in the Northern District of California, and a related commercial claim of roughly $18.2 million is pending against Nvidia in Belgium’s Ghent Enterprise Court.

Why It Matters: These are the only music-AI suits naming Nvidia — extending the litigation from music generators to the infrastructure companies whose research models consumed the same data. And a plaintiff dropping a case without prejudice, silently, is a reminder that this docket moves in ways press releases never explain.

On June 5, the American Federation of Musicians sued Universal Music Group, Warner Records and Atlantic Recording Corp. in the Southern District of New York — not over AI training, but over what the labels did after they settled. The union alleges that when the majors resolved their copyright suits against Suno and Udio in late 2025 by licensing their catalogs to the AI companies, they licensed recordings made by AFM members without notifying the union or paying the musicians who performed on them. That, the complaint says, breaches the “new use” provision of the Sound Recording Labor Agreement, the collective bargaining agreement under which the recordings were made: when a recording is repurposed or licensed into a new format, the label must notify the AFM and pay new-use compensation to the performers.

The case has moved briskly. At a July 21 pre-motion conference, Judge Edgardo Ramos let the union file an amended complaint — filed July 24, swapping parent company Warner Music Group for its subsidiaries Warner Records and Atlantic after Warner argued the parent was not a proper party — and let both majors move to dismiss. Opening briefs were due August 14, and per Music Business Worldwide the briefing schedule runs through September 11 — this Thursday. The AFM has published its own summary of the litigation on its site.

Why It Matters: This is the first labor-side lawsuit over AI settlement proceeds — a new litigation category. The majors spent 2025 establishing that AI companies must pay for catalogs; the AFM is now testing whether the musicians on those recordings are owed a cut of exactly those deals.

In mid-June, The Atlantic published an investigation by Alex Reisner identifying four music datasets circulating among AI developers, holding over 21 million recordings combined, and released searchable databases so any artist can check whether their work appears. Music Business Worldwide reported the investigation on June 16. The four collections: LAION-DISCO-12M, roughly 12 million tracks compiled from YouTube by the German nonprofit LAION; a second dataset of about 9 million tracks; and two smaller collections of about 100,000 tracks each, one of them the Free Music Archive. The contents span Bad Bunny, Nirvana, Taylor Swift, Billie Eilish, Pearl Jam and the Beatles alongside tens of thousands of independent artists. Per the report, Google and Stability AI have used material from the Free Music Archive; each of the four datasets has been downloaded several thousand times, and because AI companies keep training data secret, it is not publicly known who used the rest.

The caveats are part of the story. Presence in a dataset does not by itself prove any company trained on a given track — the investigation maps what circulates in the AI-development community, not who used what. LAION’s collection contains YouTube links and metadata rather than audio files, and LAION warns against commercial deployment. But the searchability changed the conversation: Irish music site Nialler9, testing the tool, found “almost every Irish and international artist you can think of,” and Australasian collecting society APRA AMCOS said the investigation proved AI companies had “plundered the life’s work of Australian and New Zealand songwriters,” with CEO Dean Ormston contrasting the unlicensed taking with the industry’s push for licensing — while, per Complete Music Update, ARIA’s Annabelle Herd dismissed as “nonsense” the tech-sector claim that building music AI lawfully in Australia is impossible.

Why It Matters: Training data was the industry’s largest known unknown. A free, searchable index — however caveated — hands individual artists concrete evidence for the first time, and it is already being cited in the copyright litigation this site tracks daily.

On May 21, Spotify and Universal Music Group announced recorded-music and publishing licensing agreements enabling a Spotify tool that will let fans create AI-generated covers and remixes of songs by participating UMG artists. The companies framed the deal around three principles — consent, credit and compensation: no catalog enters the system unless the rights-holding artist opts in, participation is reversible and granular, and participating artists and songwriters share in the revenue. Spotify co-CEO Alex Norström called the approach “grounded in consent, credit, and compensation for the artists and songwriters”; UMG CEO Sir Lucian Grainge said it would let artists “deepen their fan relationships while also creating additional revenue opportunities.” The tool will be a paid add-on for Premium subscribers, rolling out first in the United States.

The announcement left the practical details open — no price, no launch date, and no word on which UMG artists have opted in. TechCrunch’s reporting placed the deal against the Suno and Udio litigation: where those platforms built first and faced label lawsuits, Spotify secured the licenses before shipping the product — permission first rather than forgiveness later.

Why It Matters: This is the licensed path made concrete: the largest streaming platform and the largest label building the thing fans were already doing illegally, with an opt-in and a revenue share attached. Whether artists actually opt in — and at what price fans will pay — is the open question.

On January 13, Bandcamp announced it would no longer allow music “generated wholly or in substantial part by AI” — the first major music platform to issue a full ban on AI-generated audio rather than a labeling or royalty policy. The policy also prohibits using AI to imitate other artists or styles, and cuts both directions: it bans scraping, text and data mining, and any use of Bandcamp’s catalog to train machine learning or AI models. Enforcement leans on the community — the platform asks users to flag releases that appear substantially AI-generated, and reserves the right to remove music on suspicion, as reported by Hypebot at the time.

The line the policy draws is deliberately rough. “In substantial part” reaches tracks where AI generated the core composition even if a human added production on top, while leaving incidental AI use — mastering, EQ — untouched. Bandcamp acknowledged the boundary will be contested and committed to updating the policy as the technology evolves, positioning itself explicitly as a human-made-music sanctuary against the AI-volume strategies of the larger platforms.

Why It Matters: Eight months on, Bandcamp still anchors one end of the platform spectrum this site tracks — refuse, demonetise, label, or license. Every DSP policy decision since January has effectively been a choice of position between Bandcamp and Spotify.

On August 13, Chinese AI company MiniMax released MiniMax-Music3, an open-weights music generation model that produces complete songs of up to five minutes — composition, arrangement, vocals and production in a single pass — from lyrics and a structured caption describing genre, vocal characteristics and arrangement. The company describes an architecture pairing an 8-billion-parameter global language model with a smaller local model over a flow-matching synthesis stack, and claims improved vocal naturalness and instrument realism; those quality claims are MiniMax’s own and have not been independently benchmarked. The weights are published on Hugging Face under a community license that permits commercial use with conditions, including prominent attribution in products and separate authorization for organizations above a revenue threshold, per the model card and technical press coverage.

What the announcement does not say is equally notable: MiniMax’s release materials contain no information about what the model was trained on — no sources, no size, no licensing status. An open-weights model of this capability, downloadable by anyone, with undisclosed training data, arrives in the same year The Atlantic’s dataset investigation (above) put training-data provenance at the center of the industry’s fight.

Why It Matters: Licensed platforms and lawsuits both assume a company you can negotiate with or sue. Freely downloadable open weights have neither lever — once released, the capability is permanent and jurisdictionless, and the training-data question travels with every copy.