A dated record of what the rights chain supports.
We review a dataset's license chain and source documentation against the use you intend, and return an exception file: what the records support, what they contradict, what is missing, and who needs to do what next. Every item cites the document and clause it rests on.
“Licensed: yes” is not a rights chain.
A license that says a dataset may be used commercially does not say whether it may be used to train a model, whether the provider had the right to grant that, or whether the original collector's terms allowed the collection in the first place. Each link in that chain is a separate document, and any one of them can narrow what the next one grants.
Teams training or grounding AI models increasingly have to show that chain, not assert it. In the EU, providers of general-purpose AI models must publish a summary of the content used for training, and copyright questions about training data are being tested in court. A review that names each document, the clause relied on and the date it was checked holds up better in that conversation than a yes/no field.
The review is scoped to one intended use at a time, because rights differ by use. Training, retrieval at inference time, internal analytics and redistribution are separate questions with separate answers, and they are the same uses shown in every listing's permitted-use table.
Two audiences, one method
Data providers
We review the license chain behind your listing. If it holds up, the listing shows “Provenance reviewed” with the month, and its permitted-use table is filled in with the clause behind each answer.
Data buyers
For a dataset on Data.co or one you are sourcing elsewhere, we map the rights to your one intended use and return an exception file your counsel can work from.
Counsel and compliance
Findings come in a fixed shape with citations, not a narrative. Your team makes the legal call; the file shows what the documents do and do not say.
From intake to a dated exception file
Every review follows the same sequence, whether the dataset is listed on Data.co or not.
- 1
Intake
You state the one intended use and list the documents you hold. We confirm scope in writing before reviewing anything, including what is out of scope.
- 2
Source inventory
We list every document in the chain (licenses and amendments, data processing agreements, source terms of service, consent records, assignments) and the public sources we check, each with its version or retrieval date.
- 3
Rights mapping per use
For the intended use, we trace what each document grants, restricts or conditions, and record the clause relied on.
- 4
Contradiction and gap check
We look for documents that disagree with each other and for links in the chain with no supporting record. Each one becomes a finding with a severity.
- 5
Exception file delivered
A dated file lists every finding with the document and clause cited, its status, severity, follow-up owner and due date, so nothing rests on an unstated assumption.
- 6
Re-review on change
A review is point-in-time. When a license is amended, renewed or reaches its expiry date, the dataset is re-reviewed. A listing's badge lapses after 12 months or at license expiry, whichever comes first.
Every finding cites its evidence
This is the shape of an exception file, shortened to six findings. Everything in it is invented: the dataset, the documents, the clauses and the dates. A real file also lists every document reviewed and the conclusion for each use.
# Rights and provenance review: exception file
SYNTHETIC SAMPLE. Not a real dataset or review.
Dataset: Example regional news archive
Intended use: AI model training
Review date: 2026-10-14
F1 Supported · low
Archive license grants model training on all articles.
Evidence: Archive license v3, §4.2
Follow-up: none
F2 Contradicted · high
Publisher site terms bar automated collection.
Evidence: example-news.test/terms §7, retrieved 2026-10-02
Follow-up: Vendor · written publisher permission · due 2026-11-15
F3 Not evidenced · high
No assignment from the original photographers.
Evidence: none provided
Follow-up: Vendor · chain of title for photo set · due 2026-11-15
F4 Not evidenced · medium
No consent records for reader comments (personal data).
Evidence: DPA v1 §2 covers articles only
Follow-up: Buyer counsel · exclude comments or obtain consent · due 2026-11-30
F5 Supported · medium
License term runs to 2028-06-30; renewal is not automatic.
Evidence: Archive license v3, §11.1
Follow-up: Data.co · re-review before 2028-06-30
F6 Supported · low
Attribution required in model documentation.
Evidence: Archive license v3, §6.3
Follow-up: Buyer · add attribution to the model cardSynthetic sample · invented dataset, documents and findings
What each finding records
Findings use one fixed shape so they can be tracked and closed, and so two reviews of the same dataset can be compared.
| Field | Contents |
|---|---|
| Item | A stable id within the review, such as F3. |
| Finding | One statement about the rights chain, tied to a use or to the chain as a whole. |
| Status | Supported (a record backs it), contradicted (a record says otherwise) or not evidenced (no record seen). |
| Evidence cited | The document, version and clause, page or URL the finding rests on, with a retrieval date for web sources. |
| Severity | High, medium or low: how much the finding limits the intended use while it stays open. |
| Follow-up, owner, due | What closes the finding, who does it (provider, buyer, counsel or Data.co) and by when. |
What you receive
- A dated exception file, every item citing a document or clause
- The list of documents and public sources reviewed
- A conclusion for the intended use, with any conditions
- Follow-ups with an owner and a due date
- For providers: a filled, sourced permitted-use table on the listing
- For providers: the “Provenance reviewed” badge when the chain holds up
- A re-review when the license changes or expires
What the review is, and what it is not
It is not legal advice, and Data.co does not act as your lawyer. The exception file is input for your counsel, who makes the legal call on whether and how to proceed.
We review the documents you provide plus public sources such as published terms and license texts. We do not audit a provider's systems or contact third parties unless the agreed scope says so, and a finding is only as complete as the records behind it.
A gap means “not evidenced”, not “infringing”. A missing consent record or assignment tells you a link in the chain is undocumented; it does not tell you anyone's rights were violated.
Reviews are point-in-time and dated. A license amendment, a change to a source's terms or an expiry after the review date is not covered until the dataset is re-reviewed.
Request a rights and provenance review
Tell us the dataset, the one use you need to rely on, and which documents you have. We confirm scope and the documents we need before any review starts.
Questions buyers ask
Can you review a dataset that is not listed on Data.co?
Yes, for buyers. Describe the dataset and the documents you hold. The result is an exception file; badges apply only to Data.co listings.
Why only one intended use per review?
Rights differ by use. A license can allow internal analytics and say nothing about model training. Scoping to one use keeps each finding specific. If you need more uses, we scope them as separate questions in the same engagement.
What does the “Provenance reviewed” badge mean?
That we reviewed the listing's license chain and source records, found no contradicted finding and no open high-severity gap, and filled in the permitted-use table from the documents. It shows the month of the review and lapses after 12 months or when the license expires.
Does a badge mean the dataset is cleared for any use?
No. It covers the uses recorded in the listing's permitted-use table, as of the review month. Check the table for your use and its conditions, and have your counsel review anything you will rely on.
What happens if a provider's review does not hold up?
The provider gets the exception file with the follow-ups needed. The listing shows no badge until a later review holds up.
Do you use AI to do the review?
Not today. Every finding is recorded by a reviewer. If we add model-assisted extraction, it will only propose findings with the exact clause quoted, a reviewer will accept or reject each one, and no badge will be granted from model output.
Need to rely on a dataset for one specific use?
Tell us the dataset, the use and the documents you have. We will confirm scope before reviewing anything.