Data rights and provenance
Tracking where training data came from, what licence it carries and whether its owners opted out, and the legal questions still open.
Read first: Pretraining data.
D1 · StoryFor young children and total beginners
Every piece of training data was made by someone. Some of it is free to use, some has rules, and some owners say no.
Provenance means keeping a record of where each piece came from, so people can check it was allowed.
Like: Checking the label on every ingredient: where it came from and whether you may use it.
D2 · PictureFor curious beginners
Provenance is the record of a dataset's sources, licences and processing steps. Many public datasets are re-packaged from others, and licence details often get lost or mislabelled along the way.
Opt-outs let owners signal that they don't want their work used: a robots.txt rule aimed at a crawler, website terms, or machine-readable reservations. Honouring them is partly voluntary and partly a legal duty, depending on the country.
The law is still settling. Courts in several countries are weighing whether training on copyrighted work without permission is allowed, and some regions now require model makers to publish summaries of their training data.
The questions a piece of data should pass
- Source: where did it come from?
- Licence: what does it allow?
- Opt-out: did the owner say no?
- Personal data: remove or protect
- Record: documented for audit
Like: Food labelling: every ingredient should be traceable back to the farm, with its allergens listed.
Where it breaks: Food can be recalled from shelves. Once data has shaped a model's weights there is no reliable way to take it back out, so an opt-out mostly affects future training runs.
D5 · FrontierFor experts
Big legal questions remain open: whether training counts as fair use or falls under text and data mining exceptions, what is owed when a model reproduces training text, and how opt-outs should work at web scale.
Restricting crawlers protects owners but can shrink the open data commons, favouring organisations that already hold data or can pay for licences.