Non-Commercial License and AI Training Data | Elacity
Jamendo says a research-only music dataset trained Nvidia's commercial audio models. A non-commercial license is a request in a README. Terms need to be checked at the key.
AI Training Data Under a Non-Commercial License: The Terms Stayed in the README
If you have ever shared your work for research only, you already know the catch: once the file leaves you, your terms become a sentence someone else may or may not read. A non-commercial license on AI training data protects you as far as you can afford to sue, and not one step further.
That is the question inside the complaint Jamendo filed against Nvidia on October 12 in federal court in Northern California. The music platform says a dataset it released for non-commercial research ended up training commercial audio models. None of it is proven yet: these are the plaintiff's claims, and the court has not ruled.
What the Jamendo complaint says
The dataset is MTG-Jamendo, over 55,000 full tracks tagged for music research. Its README says it is available "solely for non-commercial research and academic use," and that any other use needs Jamendo's prior written authorization.
Jamendo's complaint alleges that Nvidia's own research materials name the dataset as a training source for Fugatto and Audio Flamingo. Nvidia unveiled Fugatto in November 2024 as a model that can modify voices and generate new sounds.
Jamendo is asking for no less than 17.8 million euros. It refiled days after dropping a nearly identical suit in the same court.
Why a non-commercial license only binds the honest
Read the README again and notice what it is: a polite request attached to a download link. The CC BY-NC-SA 4.0 license on its metadata is clear about non-commercial use. Nothing in the file checks it.
That is where the trust boundary of open licensing sits. Every downloader is trusted equally, the student and the product team alike, and the only check arrives later, in court, if the owner notices and can pay to prove it.
Noticing is harder than it sounds. A large-scale audit in Nature Machine Intelligence found that license information on popular dataset hosting sites is often missing or wrong, so terms can fall off a dataset long before anyone trains on it.
The honest complication
Locking data has a real cost. Open datasets are how academic research moves, and a subset of this one ran a public emotion and theme recognition benchmark for three years. A gate that makes every researcher beg for every file would kill what made it useful.
The law is also unsettled. Creative Commons notes that its license conditions only apply where copyright permission is needed, so an exception such as fair use can take the license out of the picture entirely. No technology settles that question, and no lock can read whether a model is research or a product.
So the goal is not to make sharing harder. It is to make the terms travel with the data and get checked at the moment of use, instead of years afterwards.
What changes when the terms live at the key
This is the problem Elacity dDRM is designed around. A dataset, a song or a catalogue can be packaged as a Wealth Capsule: encrypted, with its licence minted as an on-chain token instead of a line in a text file, so the people who made it hold it as property they control.
- The key is split. The key that opens the capsule is designed to be divided across a 2-of-3 quorum of independent machines, each of which re-checks the requester's on-chain rights before releasing its share, so no single machine ever holds the whole key.
- Research and commercial can become different rights. The design lets an owner grant a lab research access on whatever terms they set, including free, while a product team would need a commercial right at the owner's price.
- The data stays sealed. The design keeps content encrypted everywhere except inside the runtime at the moment of use, so there is no copy in the clear sitting on a server, waiting for someone to forget where it came from.
Someone could still claim research terms and build a product anyway. The difference is that they would have had to sign for those terms first, and that choice would sit on an on-chain record they cannot edit.
Beneath it sits ElastOS, the open-source runtime from the Elastos project, which Elacity helps build. It is designed so apps and agents get capabilities, not access control lists: narrow, signed passes the owner can hand out and take back. In the alpha, the runtime already checks those passes before an app can act, and agent support is still in testing. That is what would let a lab the owner has never met get research access on the owner's terms.
For how the split key works in plain terms, read Decentralized DRM, Explained. For the same pattern in fan remixes, see why AI remix licensing terms still live inside the app. More on licensing as infrastructure lives under Commerce Protocol.
Where the enforcement should live
The Jamendo case will be decided on its facts and the law. The structural lesson does not wait for a verdict: when the only enforcement is a lawsuit, the 3,500 or so artists in a dataset depend on someone else deciding to fight.
As Sasha Mitchell, Elacity's Founder, puts it: "The people who create the value should own it. That is the entire reason Elacity exists." Put the terms where the key is. Follow Elacity on X for the next piece on owning what AI learns from.