Data Provenance
Knowing where your data came from and what you're allowed to do with it — and the licence field on the dataset you're using is probably wrong.
Reviewed July 13, 2026Stable
When not to use it
- (It's a practice. The question is when to distrust a claim.)*
- Trusting the licence tag on an aggregated dataset. It describes the collection, not the restrictive source inside it.
- Assuming a derivation chain carried its terms. Each hop is a chance for the field to be retyped wrong.
- Treating web-scraped as permissive. It's unlicensed, which is the unresolved question, not an answer.
- Assuming you can remove data later. You can't. Unlearning doesn't give the guarantee you'd need.
Reach for something else instead
- Provenance-traced corpora — collections built with the chain intact. Newer than what you're using.
- Licensed data — expensive, clean, and it favours the largest labs.
- Public domain — clean, smaller, weaker.
- Datasheets — if you're publishing, this is the template.
Read more on the blog
- Do AI detectors work? Accuracy, bias, and false positivesStudents are accused, applicants are rejected, and articles are dismissed on the strength of a percentage produced by an AI detector. Those tools do not measure who wrote something. They measure how predictable it is, and the difference between those two things has real consequences for real people.
- Open weights vs open source AI: what's actually releasedAlmost every model marketed as open source AI is not open source. You get the finished weights under a licence, not the training data or the code that produced them. The difference decides what you can actually do, and regulators have now attached legal exemptions to the word, which makes the definition worth money.
- AI and copyright: the three questions people confuseWhether training on protected work is lawful, whether AI output can be owned or infringes, and whether any of it is fair to creators are three separate questions with different rules and different answers. Most of the argument consists of people answering different ones at each other.
Further reading
- Longpre et al. (2023), The Data Provenance Initiative: A Large Scale Audit of Dataset Licensing & Attribution in AI — the audit; most licence tags on popular datasets are wrong or missing.
- Gebru et al. (2018), Datasheets for Datasets — the template; the good sections are the ones people skip.
- Bourtoule et al. (2021), Machine Unlearning — why you can't take it back, and what the approximate methods don't guarantee.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Reading the licence field as a fact. It's a claim someone typed, and the audit says it's often wrong.
- Missing that aggregation launders terms. One tag over 50 sources hides the restrictive one.
- Forgetting synthetic data carries the generating model's terms. Distillation is usually explicitly prohibited.
- Planning to remove problematic data later. Retraining is the only clean answer, and at scale that's not a plan.
At a glance
FieldSafety & Ethics
The findingmost licence tags on popular datasets are wrong or missing (Longpre et al.)
Whyprovenance is transitive and nothing enforces it; each hop drops it
Why it's consequentialyou cannot unlearn; retraining is the only clean fix
What to checkthe sources, not the tag
DifficultyBeginner
Often compared with
The licence tag vs. the actual sources — one is a claim someone typed at the last hop; the other is the thing that binds you. The audit says they frequently differ.
Where this sits
4 concepts come first. Understanding it opens up 2 more.
3Levelsteps in
4Needs firstconcepts
2Opens up1% of 310
1Areastays here
Learn these firstCopyright and Training Data
Computed from the prerequisite graph, not assigned. How this works