Training Data and Consent

What a model learned shapes what you can do with it, and who should have been asked

Related

Public DomainCopyrightSample Clearance

The problem

Ask what a model was trained on and most product pages trail off into "millions of tracks" and "licensed and public data". For a user this is not gossip. Two practical questions hang on it. Can the output be used commercially, and will what you upload feed the next training round. Both depend on where the data came from and whether anyone agreed.

How it works

Training data comes in layers. Scores and MIDI are symbolic, holding pitch, duration and velocity. Recordings are audio, carrying touch, breath and room. Lyrics and metadata are text. The three carry different rights and require consent from different people. A folk tune in symbolic form is often labelled traditional, but traditional is not a licence, it only means no known rights holder, while a collector, an arranger or a translator may each be a specific person. Audio adds performers and record producers on top. A model learns all of it together, so the output cannot tell you whose playing it absorbed, and you cannot work the provenance back from a result.

Practice

Ask a vendor for four things in writing. The categories and date range of training data, not "publicly available". Whether there is filtering and an opt-out route for rights holders. Whether your inputs are used for training, whether that can be switched off, and what breaks if it is. The scope of the commercial licence, including your territory and distribution channels. Keep your own record too, purchase date, a snapshot of the terms, the reply email, all in the project folder. Terms change, so store the version you agreed to. Rules on data and personality interests differ by jurisdiction, so confirm the specifics with a lawyer.

What to listen for

In the Gymnopedie score, what is missing when only notes remain, touch and pedal are not in the file. In the Chopin nocturne recording, what one score becomes under a pair of hands, which is another layer of data. In the hornpipe credited to a named collector, the specific source behind a tune, and then in the Klezmer tune labelled only traditional, a case where nobody is named to ask.

Sources

• 训练数据的层次划分，依公开模型文档与数据说明核对，未抄原文
• 版权与人格利益对同意的要求，依公开法律评论归纳，未抄文本

Used for fact-checking only; all prose is original.