When the training corpus has no provenance: what to do first
When a training corpus arrives without provenance records, stop feeding it into new training runs at once. Freeze the current version, isolate the segments with unknown origin, and open a source audit before any further training, fine-tuning, or distribution. Ownership of the resulting model, code, and documentation cannot be confirmed until each segment is traced to a licence, a public-domain source, or an internal record.
Legal basis
Two established acts define the rights most often at stake. Directive 2009/24/EC protects computer programs as original works, which covers code generated with, or produced alongside, a trained model. Directive 96/9/EC protects databases, giving the maker of a substantial collection a sui generis right against unauthorised extraction and reuse.
A training corpus, if it qualifies as a substantial compilation, can trigger that second right even where no single item inside it is separately protected. Provenance is what tells you whether your organisation is the maker of the compilation, a licensee of someone else's, or a party exposed to a claim under either directive.
What changes without provenance
Without provenance, three practical things become uncertain at once. You cannot confirm that the corpus was lawfully assembled, you cannot warrant to a customer that the deliverable is free of third-party rights, and you cannot price a licence for the model's output with any confidence.
The exposure is not theoretical. A rights holder who traces a fragment of the corpus back to a protected database can seek an injunction against further use, independent of whether your organisation acted in good faith. Good faith affects remedies; it does not remove the underlying claim.
What to do first
Work through the audit in a fixed order. Skipping steps to save time usually means repeating them later, with less evidence available.
- Freeze the corpus version currently in use and record its checksum before anything else changes.
- Map every segment to a known source: internal generation, a signed licence, a public-domain register, or "unknown".
- Pull the contracts covering any purchased or vendor-supplied segments, including warranty and indemnity clauses.
- Check whether any segment carries personal data that has already moved to a processor or subprocessor; the pattern is set out in Personal data flows to a new subprocessor technology contract and applies here without change.
- Rank each "unknown" segment by exposure, weighing size, distinctiveness, and how visible the output already is to third parties.
- Decide, segment by segment, whether to remove, retrain around, or seek a retroactive licence.
| Source of the segment | Typical risk | First action |
|---|---|---|
| Internal generation with logs | Low | Confirm authorship internally, close the record |
| Licensed dataset with a contract on file | Low to medium | Check licence scope against actual use |
| Purchased dataset, contract missing | Medium | Request the contract, assess indemnity cover |
| Scraped or otherwise unsourced | High | Freeze use, run a full source audit |
Who decides, and on what timeline
Internally, the decision to freeze, retrain, or licence sits with whoever holds the IP mandate — typically the legal or compliance lead, not the engineering team that built the corpus. That person needs the source map before signing off on any next release.
No fixed statutory clock starts merely because provenance is missing. A deadline begins running only once an event triggers it: a demand letter, a licence termination notice, or a claim under the sui generis right in Directive 96/9/EC. From that date, the response window is set by the claim itself, not by an internal calendar.
Externally, disputes over the corpus are decided by the competent national court where the claim is brought, applying the implementing law behind the two directives above. This page does not set that forum; it only tells you what to gather before you are asked for it.
Branches: three starting points
Different origins call for different first moves.
If the corpus was built internally, with logs of generation or collection, reconstruction is usually fast: pull the logs, confirm authorship, and close the segment as low risk.
If the corpus was supplied by a related party, the same scrutiny used for affiliated structures elsewhere applies — the logic is close to what is set out in When the issuer and the platform are related explained, where a related supplier's paperwork needs the same independent check as a third party's.
If the corpus was scraped or assembled from unclear web sources, treat every segment as unresolved until traced. This branch carries the highest exposure under the sui generis database right and the slowest resolution.
Sources and limits
This page identifies the two directives that govern the underlying rights and sets out the order of an audit. It does not determine whether a specific corpus in fact qualifies as a protected database, and it does not calculate exposure for a particular deliverable — that requires the file itself.
Where the provenance gap leads to a dispute with a counterparty based in the Netherlands, and recovery or verification becomes relevant, our report on Asset in Netherlands in practice sets out what can be established about a party's holdings.
Where the counterparty's standing itself needs checking before a claim is brought, Debtor in Netherlands and the evidence covers what evidence is typically available.
Sources cited above: Directive 2009/24/EC and Directive 96/9/EC, both currently in force. This page was last checked on 12 May 2025.
Related reading
A related audit habit applies when open source turns up unannounced inside a finished deliverable — see When open source is found inside a deliverable in a client's codebase.
If your corpus lacks provenance records, start the audit before the next training run, not after a claim arrives.
FAQ
What is the first step when the training corpus has no provenance?
Freeze the version in use, record its checksum, and stop feeding it into further training until each segment is mapped to a source.
Which deadline is already running?
None starts automatically. A deadline begins only once an event triggers it — a demand letter, a licence termination, or a claim under the sui generis right in Directive 96/9/EC — and the response window follows from that event, not from the discovery of the gap itself.
Who decides, and on what evidence?
Internally, the person holding the IP mandate decides, based on the source map, licence records, and vendor correspondence. Externally, a dispute is decided by the competent national court, applying the law implementing the two directives above, on the evidence each side presents.