Your Music Is Being Used to Train AI. Here's What You Can Actually Do About It.
AI companies trained on tens of millions of recordings — including yours — without consent. Here's what the lawsuits mean, what opt-outs exist, and the steps you can take to protect your catalog now.
The Scale of the Problem
Suno, one of the most widely used AI music generators, admitted in federal litigation that its training data includes "essentially all music files of reasonable quality that are accessible on the open internet" — constructed by ingesting tens of millions of recordings gathered from publicly available sources. That admission covers copyrighted works owned by major labels. It almost certainly covers yours too.
The scope across the industry is staggering. Google's MuLan dataset, used to train its MusicLM model, contains 44 million music recordings totaling 370,000 hours of audio sourced from YouTube. Google's MusicCaps evaluation dataset drew on 5,521 music clips from YouTube — with artist, song, and album metadata conspicuously stripped out, a choice critics argue was designed to obscure each clip's copyright status.
What makes independent and back-catalog recordings especially attractive for training is not popularity — it's depth and genre diversity. A model's musical range depends on how varied its training data is, not how recent. That means a lo-fi indie recording from 2009 or an obscure jazz session from a decade ago carries genuine value to an AI developer, regardless of whether it ever charted. Independent musicians and producers are not collateral damage in this story. Their catalogs are part of the foundation these systems are built on — and so far, without compensation or consent.
Is Training on Your Music Copyright Infringement?
The honest answer is: probably, in many cases — but courts are still working through precisely when and how. The U.S. Copyright Office addressed this directly in its May 2025 report on generative AI, concluding that AI training cannot automatically qualify as fair use. Where a model is trained to produce content that "shares the purpose of appealing to a particular audience" as the original works, the Office found the use is "at best, modestly transformative" — a weak fair-use posture. AI companies have argued their models learn from music the way a musician learns by listening. The Copyright Office rejected that analogy outright: AI creates perfect copies of recordings during training, not the imperfect impressions a human listener retains.
The market-harm analysis cuts especially hard against AI music generators. The Copyright Office flagged three distinct harm theories: direct verbatim reproduction in outputs, genre dilution as AI-generated tracks flood streaming platforms, and lost licensing revenue from a viable training-data market being bypassed entirely. That last point matters for independent artists — if AI developers had to license training data the way they license sync rights, your catalog would have market value. The Office's position suggests that bypassing that market is not a free pass. Music also carries a structural wrinkle: copyright in a sound recording and copyright in the underlying composition are separate rights. Courts have held that a new recording simulating an original does not automatically infringe the original recording — a nuance that may limit some label claims while leaving composition-infringement arguments intact.
The Lawsuits: Where Things Stand
In June 2024, the Recording Industry Association of America filed copyright suits against Suno in federal court in Massachusetts and against Udio in the Southern District of New York, alleging both companies copied "decades worth of the world's most popular sound recordings" without license on a "massive scale." These were not demand letters or licensing disputes — they were direct infringement claims filed by the major labels in two federal courts simultaneously.
By late 2025, partial settlements had been reached: Universal Music Group settled with Udio in October 2025 and Warner settled with Suno in November 2025. But neither lawsuit is fully resolved — Sony continues active litigation in both cases as of April 2026. The UMG-Udio settlement is structured as artist opt-in licensing, meaning artists must affirmatively authorize use of their catalog rather than being opted in by default. That structure is meaningful: it treats artist authorization as a prerequisite, not an afterthought.
What those settlements did not do is answer the legal question. Label deals with AI companies are business arrangements — they exchange licensing fees for liability releases without any court ruling on whether the original training was lawful. As a result, fair use remains completely unresolved for anyone not party to those deals. Independent musicians whose recordings were ingested get nothing from a Sony-Suno licensing agreement — their claims were not settled, and their rights were not addressed.
Sony is pressing forward specifically to force a court ruling, and the Suno case in Massachusetts is expected to produce the most consequential AI music copyright decision to date, with a hearing anticipated for summer 2026. Separately, independent musicians filed their own class actions against Suno and Udio in 2025, including novel DMCA claims — a sign that the fight has expanded well beyond the major labels. If you want to understand where your intellectual property rights stand in this landscape, the answer depends on what courts do next, not what labels already agreed to.
Opt-Out Options: What Exists and Whether It Works
The short answer is: the options are limited, and none of them protect music that's already been ingested. Spotify's terms grant the platform a sweeping worldwide license to uploaded content — including the right to "create derivative works" — to power its own AI features. There is no opt-out mechanism available to artists. You agreed to those terms when you distributed through the platform.
Elsewhere, metadata-based opt-out tags exist in theory. In practice, research published in the Berkeley Technology Law Journal found that these tags are easily stripped from files, incompatible with some audio formats, and don't bind AI companies that acquired files before the tag was added. The same analysis applied to robots.txt domain-level exclusions — Perplexity AI violated robots.txt directives outright, demonstrating that a decades-old web standard carries no legal enforcement mechanism against determined actors.
The deeper problem is technical, not procedural. Once a model has trained on a recording, that data cannot be removed from its learned weights without retraining the entire model from scratch — something no AI company will voluntarily undertake. By the time most musicians even learn their work was included in a training set, opting out only affects future training runs. The past ingestion is already baked in.
One signal worth watching: the UMG-Udio settlement used an opt-in structure, requiring affirmative artist consent before any work could be licensed for AI training. That's the inverse of every opt-out regime currently on the market, and it may be where the industry is heading under litigation pressure. For independent musicians choosing which AI tools to support, Fairly Trained certifies models that obtained consent for all training data and did not rely on fair use exceptions — a meaningful signal, even if it can't undo past harm.
What You Can Do Right Now
The legal landscape is still shifting, but a few concrete steps reduce your exposure today. The most important is registering your copyrights with the U.S. Copyright Office. Registration is a federal prerequisite to filing an infringement lawsuit, and timing matters: register before an infringement occurs — or within three months of publication — and you're eligible for statutory damages between $750 and $30,000 per infringed work, plus attorney's fees. Without it, you're limited to actual damages, which are notoriously hard to prove against a company whose training pipeline ingested millions of files.
Registration is also becoming a prerequisite for the next generation of transparency rights. The proposed TRAIN Act (S. 2455) would allow copyright holders to obtain an administrative subpoena demanding disclosure of whether a specific AI model trained on their registered works — and failure to comply would create a rebuttable presumption that the developer made unauthorized copies. California's AB 412, backed by SAG-AFTRA, would impose a parallel disclosure obligation at the state level. Neither law is in force yet, but both are tied to registration status. Musicians who haven't registered are locked out of both mechanisms before the fight even starts.
On the PRO side, ASCAP, BMI, and SOCAN now accept registrations for compositions with AI-assisted elements under aligned policies adopted in October 2025 — register your catalog under those policies before a dispute forces the question. And when evaluating distributors or sync platforms, look for contract language reflecting the consent standard in the UMG-Udio and Warner-Suno settlements: artists retaining full control over whether their name, likeness, voice, and compositions can be used by AI systems.
Collective action matters too. The Artist Rights Alliance's April 2024 open letter — signed by Billie Eilish, Nicki Minaj, Stevie Wonder, and more than 200 others — put direct pressure on AI developers and platforms not to devalue music. Supporting organizations actively lobbying on these issues is one of the few levers independent musicians have while litigation and legislation catch up to the technology.
If you want a legal review of your existing catalog registrations or help structuring AI-use language in your contracts, Promise Legal's IP practice works with independent artists and producers on exactly these questions.
Six Steps to Protect Your Catalog Now
- Register your recordings and underlying compositions with the U.S. Copyright Office — registration is a prerequisite to filing an infringement lawsuit and strengthens any licensing negotiation.
- Verify your works are registered with your performing rights organization (ASCAP, BMI, or SESAC) and that metadata is current and complete.
- Read your distributor's current terms of service with AI specifically in mind — look for clauses permitting sublicensing to third-party AI developers.
- Add explicit "no AI training" language to any new sync, licensing, or collaboration agreements you sign going forward.
- Support legislative efforts such as the TRAIN Act and California AB 412 by contacting your representatives, and consider joining the Artist Rights Alliance.
- If you have reason to believe your music was already ingested without authorization, consult an IP attorney to evaluate your options before the legal landscape shifts further.