Skip to content

Rights and contracts

AI dataset purpose expansion: how to assess a buyer's request for a new use

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

Treat a buyer's request for a new data use as a change to the license, not an interpretation of it. Have the buyer describe the use, records, recipients, outputs and dates in writing, test each against the signed grant, your customer promises and any privacy duties, then approve, narrow or decline it in a signed amendment. Anything the agreement does not grant in words stays with the supplier.

License terms

AI dataset purpose expansion

This page covers AI dataset purpose expansion: how to assess a buyer's request for a new use.Illustrative

Key takeaways

  • A new purpose, recipient, output, publication or retention period is a scope change even when the buyer calls it a clarification.
  • A fixed request template forces the buyer to name the records, parties and outputs, which makes a narrow approval possible.
  • Privacy compatibility, contractual rights and promises to your own customers are separate tests; the new use has to pass all three.
  • Partial approvals are normal: approve the part that fits, convert third parties into instructed recipients, and refuse public release of record text.
  • Record every outcome, including a refusal, in writing signed by someone with authority.

How should a supplier assess a buyer's request for a new data use?

Treat it as a request to change the license, not as a question about what the license might already allow. Ask the buyer to put the new use in writing, compare it clause by clause with the signed grant, check whether your own obligations to customers and the people in the records permit it, and then approve, narrow or decline it in a signed amendment.

Buyers rarely send a request labeled scope change. It arrives as a project update: the evaluation set worked well, so the team would like to fine-tune on it; a subsidiary wants a copy; a research group wants to publish a few examples. Each of those is a different use, and each changes who holds your records, for how long and with what result.

The safest working assumption is simple. Anything the agreement does not grant in words stays with you. The work is to find out exactly what the buyer wants to do, what it would produce, and whether you can say yes without breaking a promise you made to someone else.

Does the new use fall outside the grant you already signed?

Often it does, and the way to tell is to read the grant against the request one element at a time rather than asking whether the new use is similar in spirit. A license for AI training data usually fixes six things: the permitted purpose, the parties allowed to use the records, the outputs the buyer may keep, publication, the term, and what happens at the end.

If the request changes any one of those six, it is a scope change even when the buyer calls it a clarification. The table below maps the requests suppliers most often receive to the clause that decides them.

Does the new use fall outside the grant you already signed?
What the buyer asksClause to readUsually inside the grant?What to confirm
Fine-tune a model on records licensed for evaluation onlyPermitted purposeNoWhether training was excluded by name or simply not listed
Use the records for a second product line or model familyPermitted purpose and field of useOnly if the grant names the buyer's models generallyWhether the grant is tied to a named model, product or program
Give an affiliate or partner its own copyParties and permitted recipientsNoWhether the affiliate would use the copy for its own purposes
Publish sample records, benchmark items or model outputs that quote recordsPublication and confidentialityNoWhether any record text, even edited, would become public
Keep a held-out test set after the term endsTerm, return and deletionNoWhich copies survive and who can open them
Generate synthetic data from the records and treat it as the buyer's ownDerived outputs and ownershipDepends on the derivative clauseWhether synthetic outputs are licensed materials or buyer property
Combine the records with another datasetPermitted purpose and re-identification termsRarely addressedWhether the combination makes people or customers easier to identify

Which proposed uses deserve a written decision even when the agreement seems to cover them?

Some requests look covered by broad wording but change the risk enough that you should decide them explicitly and record the answer. Broad grants such as any machine learning purpose were drafted before anyone knew what the buyer would build, and the buyer's reading may not be yours.

Make an explicit decision, in writing, whenever the request involves any of the following.

  • A model that will be offered to the buyer's own customers when the license contemplated internal research.
  • Training for a third party, such as a model the buyer builds under contract for someone else.
  • Any public release: a paper appendix, a leaderboard test set, a demo that shows real records, or a model card that quotes them.
  • A new recipient that would decide for itself how to use the records, rather than acting only on the buyer's instructions.
  • A use that depends on the records still describing identifiable people, customers or employees.
  • A longer term, a perpetual retention right for one copy, or survival of a test set after deletion.
  • Any use your own customer contracts, privacy notices or employee policies would not have led those people to expect.

What should the scope change request template contain?

The template should make the buyer describe the new use precisely enough that you could later prove what was approved. Vague requests produce vague approvals, and a vague approval is read in the buyer's favor once a dispute starts.

Send the template as a fixed form, ask the buyer to complete every row, and keep the completed version with the amendment it led to. The rows below are the minimum; add industry-specific rows where your records carry extra duties.

What should the scope change request template contain?
Template fieldWhat the buyer must stateWhy the supplier needs itAnswer that should stop the review
Current grant referenceAgreement name, date, version and the clause the request changesTies the request to one signed textThe buyer cannot point to a clause
Proposed useThe activity in one sentence, such as fine-tuning a named model for customer supportDefines what an approval coversAny AI purpose or similar open wording
Records affectedWhich delivered files, fields and record counts the new use touchesLets you narrow approval to a subsetAll licensed materials with no reason given
Who will touch the recordsEvery legal entity and contractor, and whether each acts for the buyer or for itselfSeparates recipients from new licenseesA recipient that wants to reuse the records itself
Outputs producedModels, weights, synthetic records, labels, evaluation scores, publicationsDecides what survives the term and what can be shownOutputs that would contain record text
Exposure to othersWhether outputs reach the buyer's customers, the public or a third partyChanges the confidentiality and re-identification riskPublic release of anything derived from single records
DurationStart date, end date and any copy kept after the termPrevents silent retentionIndefinite retention of a test set
SafeguardsAccess controls, re-identification ban, output filtering, deletion methodShows the new use is controlled, not just describedNo change from the original controls despite wider exposure
Evidence offeredAudit rights, logs or a written certificate the buyer will provideLets you check the use laterNo audit or certification at all
SignatoryThe person with authority to bind the buyer to the amendmentAvoids approvals by email from project staffNo named signatory

What does privacy law add when the records describe people?

When the licensed records still contain personal data, the new use has to work under privacy law as well as under the contract, and the two tests are separate. A contract can allow something that privacy law forbids, and privacy law can allow something the contract never granted.

Under the GDPR, personal data must be collected for specified purposes and not further processed in a way incompatible with them. Where the new purpose is not based on consent or a qualifying law, Article 6(4) lists what the controller weighs: any link between the purposes, the context of collection, the nature of the data, the possible consequences for the people involved, and safeguards such as encryption or pseudonymization.

The Court of Justice applied that test in the Digi case, holding that copying customer data into a separate database for testing and error correction is not barred by purpose limitation if the further processing is compatible with the original purposes. The compatibility question arises only when the new purpose differs from the original one, which is exactly the situation a scope change creates.

The UK regulator adds a point suppliers often miss: compatibility is not itself a lawful basis. The ICO says the original basis may not suit the new use, so a lawful basis must be identified for the reuse, and data first collected on consent faces stricter rules. For AI models specifically, the European Data Protection Board says a model trained on personal data cannot be assumed anonymous; that question is assessed case by case.

Keep the distinctions clear when you answer. Access to the records, privacy permission to process them, ownership of them and the contractual right to license them are four different questions. De-identifying the records lowers privacy risk but does not by itself give you or the buyer a right to a new use.

Can promises you made to your own customers block the new use?

Yes, and this is the check most likely to change the answer. Your records were created under customer contracts, privacy notices, employee policies and confidentiality terms, and the buyer's new use has to fit inside those as well as inside your license to the buyer.

The U.S. Federal Trade Commission has warned that a company adopting more permissive data practices, such as using data for AI training, and announcing the change only through a retroactive amendment to its terms or privacy policy may be acting unfairly or deceptively. In a separate post the agency said that commitments made in terms of service and other materials still bind a company, and noted that past orders required deletion of models built on unlawfully obtained data, not only the data itself.

For a supplier, that means a buyer's new use can expose you even when the buyer is the one changing its plans. Before approving, check whether the new use is consistent with what you told the people and businesses behind the records, and whether a past promise would be broken if the model or its outputs became public.

How would a supplier split an illustrative fine-tuning request on an evaluation license?

This example is illustrative and not drawn from a real deal. A field services company licensed a set of closed work orders to a buyer for evaluating a scheduling assistant. The grant named evaluation only, a two-year term, deletion at the end, and no publication.

At a glance

How would a supplier split an illustrative fine-tuning request on an evaluation license

Eight months in, the buyer asked to fine-tune the assistant on the same records, to share a copy with an affiliate that runs the assistant for a different market, and to publish twenty de-identified work orders as examples in a technical report.

The supplier sent the scope change template. The completed form showed that fine-tuning would use about 60 percent of the delivered records, the affiliate would decide for itself how to use its copy, and the published examples would include free-text technician notes. The supplier's customer contracts allowed internal analytics and service improvement but said nothing about third-party publication.

  • Fine-tuning: approved for the named assistant only, on the identified subset, with the original deletion date unchanged and weights treated as buyer property but no record text allowed in outputs.
  • Affiliate copy: declined as framed. The supplier offered instead to let the affiliate act as a permitted recipient working on the buyer's instructions, which the buyer accepted.
  • Publication: refused. Free-text notes could identify customers' sites and technicians, and the supplier's own customer terms did not contemplate public release. The buyer agreed to publish aggregate scores only.
  • Exception: one category of records, work orders for a customer whose contract barred any use beyond service delivery, was removed from the fine-tuning subset entirely and the buyer certified its deletion.

Approve, narrow or decline: how should the decision be recorded?

Record every decision as a signed amendment that names the new use, the records, the parties, the outputs and the dates, even when the answer is a partial yes. An approval given by email or in a call is hard to prove later and is usually read broadly.

Choose one of four outcomes and write the amendment to match it.

Approve, narrow or decline: how should the decision be recorded?
DecisionWhen it fitsWhat the amendment must say
Approve as requestedThe use fits your upstream obligations and the template answers are specificThe new purpose in the buyer's own words, the records covered and the unchanged controls
Approve a narrower versionPart of the use fits, or it fits for some records onlyThe subset, the excluded uses and any added safeguard such as output filtering
Approve with a new recipient roleA third party is needed but should not gain its own rightsThat the party acts only on the buyer's instructions, with flow-down terms and buyer liability
DeclineThe use conflicts with customer promises, privacy duties or your risk appetiteNothing; send a short written refusal so the record shows the use was not granted

Which signals mean a scope request needs counsel and a privacy review first?

Bring in counsel before you reply if any answer on the template is vague, if the use reaches the public or a third party's customers, or if the records still describe identifiable people. Those are the cases where a wrong yes is expensive to unwind, because models and published material cannot easily be recalled.

Also pause when the buyer cites urgency, when the request arrives after the buyer has already started the new use, or when the buyer's ownership has changed since signing. A buyer that was acquired may be asking on behalf of a parent whose plans you never assessed. Counsel should review the specific agreement, the upstream contracts and the jurisdictions involved before anything is signed.

Where does a buyer's scope request fit in a SourceX licensing deal?

SourceX deals start from the same principle this page applies: the supplying company approves the scope and terms and signs before any record changes hands, and what SourceX may then do with the de-identified dataset, including using, sublicensing and distributing it, is bounded by that signed text. So when a wider use comes up later, it is answered against the agreement, and the request template above is a workable way to put the question.

A company still deciding whether its records could be licensed at all can begin with a short company assessment that takes no uploaded data.

Frequently asked questions

Can a buyer use licensed training data for a new purpose without asking?

Not unless the license grants that purpose in words. Rights the agreement does not grant stay with the supplier, so a new purpose, recipient, output or publication needs a written amendment. Buyers sometimes treat broad wording as permission, which is why the supplier should decide explicitly.

Is fine-tuning a different use from evaluation?

Usually yes. Evaluation measures a model against the records, while fine-tuning changes the model's weights using them, so the records shape a lasting asset. If the license names evaluation only, fine-tuning is a scope change that needs its own decision.

Does de-identification let the buyer expand its use of the records?

No. De-identification reduces privacy risk, but it does not create a contractual right to a new use. The supplier's license, its customer contracts and any confidentiality terms still decide what the buyer may do.

What if the new use is compatible under the GDPR?

Compatibility under Article 6(4) answers only the privacy question, and only where personal data is involved. The ICO notes a lawful basis is still needed for the reuse. The contract must also grant the use, and the supplier's promises to its own customers must allow it.

Should a scope change be priced or documented differently?

It should always be documented as a signed amendment that names the use, the records, the parties, the outputs and the dates. Commercial terms for the change are a separate negotiation, but the written scope is what protects the supplier later.

Sources

  • GDPR Article 5(1)(b) requires purpose limitation, and Article 6(4) lists the compatibility factors: link between purposes, context of collection, nature of the data, possible consequences and safeguards such as encryption or pseudonymisation. Source
  • In Digi (C-77/21) the Court of Justice held that storing previously collected data in a test database is not barred by purpose limitation where the further processing is compatible, and that compatibility arises only when the purposes differ. Source
  • The ICO says compatibility is not itself a lawful basis, the original basis may not suit the new use, and data collected on consent faces more restrictive reuse rules. Source
  • The EDPB's Opinion 28/2024 says AI models trained on personal data cannot in all cases be considered anonymous; anonymity is assessed case by case. Source
  • The FTC said adopting more permissive data practices, such as AI training, through a surreptitious retroactive change to terms or a privacy policy may be unfair or deceptive. Source
  • The FTC said commitments in terms of service and other materials bind companies, and noted past orders required deletion of models built on unlawfully obtained data. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies