Provenance review of a training data set: what the decision maker looks at
A provenance review checks where every item in a training data set came from, who had the right to use it, and whether that right still holds at deployment. The reviewer looks for a documented chain from source to model, not a general assurance the data is clean. Gaps in that chain are the finding.
Who runs the review and who decides
Most organisations assign the review to in-house intellectual property counsel or a designated compliance lead, not to the engineering team that built the model. Engineering supplies facts about sources and pipelines; it does not sign off on the legal risk attached to them. The decision maker weighs those facts against the organisation's own risk tolerance, and either clears the data set, orders remediation, or escalates the file to outside counsel.
| Stage | Decision maker | Typical timeframe |
|---|---|---|
| Intake and triage | Compliance lead or in-house IP counsel | 2–3 business days |
| Source mapping | Designated data steward, reporting to legal | 1–2 weeks |
| Gap and risk assessment | IP counsel, with input from engineering | 1–2 weeks |
| Remediation or sign-off | Head of legal or a delegated committee | Depends on the findings |
Stages, in the order they actually run
- Freeze the current version of the data set before anyone edits it further.
- List every distinct source: internal collection, licensed feed, scraped content, public archive, third-party vendor.
- Match each source to a licence, a consent record, or a stated public-domain status.
- Flag any source with no matching record as an open item, not as an acceptable risk.
- Score each open item by how central it is to the trained model's output.
- Route the scored list to the decision maker for a documented decision on each item.
Documents the reviewer will ask for
- Acquisition logs showing when and how each source entered the pipeline.
- Licence terms or vendor contracts covering the supplied content.
- Consent or opt-out records where the data touches identifiable people.
- Any prior legal opinion already issued about a specific source.
- A current model card or equivalent internal note describing what the data set feeds.
Where the trail commonly breaks
The most frequent gap sits at the point where data arrived through an intermediary rather than directly from the original rights holder. A vendor may pass on a data set under terms that the buyer never read in full, in the same way that a distribution contract can bind a party to conditions it never saw, as set out in A reseller signs terms you never saw technology contract. Once that pattern appears in a training data set, the reviewer treats the whole batch supplied by that intermediary as unverified until the underlying terms surface.
The point of no return
Before a model is trained on the data set, an unresolved item can simply be removed. After training, removal means retraining, which most teams delay or skip. The decision maker treats the training run itself as the point of no return, and pushes for sign-off before that run starts rather than after. A model already in production with unresolved provenance items carries a standing liability that does not shrink on its own.
Two forks in the process
If every source resolves to a clear licence or a public-domain status, the reviewer closes the file with a short memo and no further steps are needed. If one or more sources remain unresolved, the file moves to remediation, which can mean removing the source, seeking a retroactive licence, or accepting a documented residual risk at a named level of authority.
A comparable staged pattern appears outside intellectual property as well: some regulated activities carry their own notification duty to a supervisor before a launch date, which runs on a similar logic of stages and deadlines, as described in How notification to the supervisor token issuance runs in. The provenance review does not trigger that duty, but a team already running one staged process tends to recognise the shape of the other.
What this review does not cover
This review confirms that the organisation may use the data set it holds. It does not establish what assets a counterparty holds if a dispute over that data set turns into a claim for damages, which is a separate factual question answered through a targeted report such as Asset in Sweden: what can be established. It also does not locate a debtor who has stopped responding once a claim is filed, a question covered instead in Debtor in Switzerland in detail.
Related reading
A provenance review of the training data itself is one half of the picture; the other half concerns what the trained model produces. That question, and how a decision maker treats open source components inside a delivered work, is covered in Open source compliance in a deliverable: what the decision.
Verified
This page was checked on 12 June 2025. It describes an internal governance procedure set by the organisation running the review, not a procedure fixed by a specific statutory provision, so no external legal citation is attached to it.
FAQ
How long does a provenance review of a training data set take in practice?
A first pass through a moderate-sized data set typically runs two to four weeks from freeze to a documented decision. Larger or more fragmented sources extend that timeline, mainly at the source-mapping stage.
Which documents are needed to start?
The reviewer needs acquisition logs, any licence or vendor contract on file, and consent records for content touching identifiable people. Without at least the acquisition log, the review cannot begin.
What happens if the other side does not respond?
If a vendor or original source does not respond to a request for licence confirmation, the reviewer records the item as unresolved and applies the same treatment as an outright gap: remediation, retroactive licensing, or a documented decision to accept the residual risk.