Rights and contracts
AI dataset derived model rights: what should your data license say?
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
A data license should name every derived use explicitly rather than rely on a derivative works clause: covered and successor models, distilled students, synthetic data and embeddings, evaluation and publication, weight release, affiliates and contractors, and what survives termination. For each, counsel should decide the scope, where it lives in the agreement and what evidence the licensee keeps so the licensor can check it.
License terms
AI dataset derived model
Key takeaways
- Silence about models in a data license can be read as granting almost everything or almost nothing; neither reading is safe.
- Copyright's derivative work definition does not settle whether a trained model is covered, so the contract must define its own terms.
- Distillation, synthetic data, retrieval indexes, publication and weight release need explicit decisions, not assumptions.
- Termination terms must say what happens to data, derived data and models separately, backed by a model register.
- Published model licenses already draft for distillation and flow-down, which makes them useful references.
Which derived model uses should a data license address explicitly?
A data license should say, in its own words, what the licensee may do with every model and artifact built from the licensed records: the first trained model, fine-tuned and successor versions, distilled or compressed students, synthetic data and embeddings, evaluation sets, published or open weights, and use by affiliates and contractors. It should also say what survives termination and what must be deleted.
At a glance
Which derived model uses should a data license address explicitly
Most first drafts license the data and stay quiet about the models. That silence is not neutral. Depending on the wording and the governing law, it can be read as allowing any use of anything built from the data, or as allowing almost nothing beyond the first training run. Either reading surprises somebody, and the surprise usually arrives after a model has shipped.
For a company licensing its records, the practical goal is not to forbid derived models. Training a model is usually the point of the deal. The goal is to make each derived use a decision both sides wrote down, with evidence obligations that let the licensor check it later. The checklist in this guide is built for that conversation with counsel.
- Models trained or fine-tuned on the licensed data, and the versions that follow them.
- Students distilled from those models, and models trained on their outputs.
- Synthetic records, labels, embeddings and indexes generated from the data or from covered models.
- Evaluation sets and benchmarks drawn from the data, and whether results or examples can be published.
- Release of weights to the public or to third parties, and hosted access by customers.
- Who counts as the licensee: affiliates, contractors, cloud providers and acquirers.
- What survives termination, what must be deleted, and what evidence proves either.
Why does a derivative works clause not settle model rights?
Because copyright's derivative work definition was written for works like translations and adaptations, and whether a trained model is one is unsettled; the contract has to define its own terms. U.S. law defines a derivative work as a work based upon one or more preexisting works, with examples such as translations, adaptations and other recast or transformed versions. Applying that to model weights trained on business records is a question courts and commentators still argue about.
Business records add a second problem. Much of what makes a support history or a claims archive valuable is facts, structure and process, which copyright may protect thinly or not at all. A clause that only restricts derivative works in the copyright sense can leave the most valuable uses outside its reach.
Standard data licenses show how much a default can give away. The Community Data License Agreement Permissive 2.0 defines Results as any outcome obtained by computational analysis of the data, including machine learning models, and states that the agreement imposes no restriction or obligations on the use, modification or sharing of Results. That is a deliberate choice for open data sharing, but a company licensing commercial records on a similar template would be granting unrestricted model rights without negotiating them.
The fix is to define the terms in the contract itself: what a Covered Model is, what Derived Data is, and which uses of each are permitted. Copyright law then becomes a backstop, not the main instrument.
How do licensed records turn into derived artifacts?
Each artifact carries the data's influence differently, so each needs its own line in the license. Use this map to decide which artifacts your draft mentions and which it leaves silent on purpose.
Two columns matter most. Whether an artifact can reproduce licensed records decides how much confidentiality and privacy exposure it carries. Whether the licensee can move it to a third party decides how much control you keep. An artifact that scores high on both, such as a fine-tuned model whose weights could be released, deserves the most specific language in the agreement.
Leaving an artifact out can be a valid choice, for example when the licensee will never generate synthetic data. Record that choice in the negotiation notes so a later reader knows the silence was deliberate.
| Derived artifact | How it is made | Can it reproduce licensed records? | Risk if the license is silent |
|---|---|---|---|
| Trained or fine-tuned model | Weights updated on the licensed data | Possible, through memorization | Unclear whether hosting, resale or later versions are allowed |
| Successor model | New version trained from the prior model or the same data | Possible | Licensee may assume rights carry forward indefinitely |
| Distilled student | Trained on a covered model's outputs or internal signals | Less likely, but not excluded | Student may be treated as outside the license entirely |
| Synthetic dataset | Generated from the data or a covered model | Possible for rare or unique records | May be sold or shared as if it were new data |
| Embeddings and search index | Vectors computed from each record | Often linkable back to the source record | Index may outlive the data and be used for retrieval |
| Evaluation set or benchmark | Selected records held out for testing | Yes, often verbatim | Records may be published with results |
| Open or transferred weights | Model files released or delivered to others | Possible | Licensor loses any practical control over use |
What should the counsel reviewed derived use issues checklist include?
It should list each derived use as an issue, the question counsel must answer, where in the agreement the answer lives, the evidence the licensee must keep, and the realistic options. The table below is the asset of this guide. It is a starting point for qualified counsel, not a substitute for their review, and the decision options are examples rather than recommendations.
At a glance
What should the counsel reviewed derived use issues checklist include
| Issue | Question for counsel | Where it lives | Evidence to require | Typical options |
|---|---|---|---|---|
| Covered model definition | Which models count as built from the data? | Definitions | Model register with training data versions | Named model only; models trained in the term; all descendants |
| Successor versions | Do later versions keep the rights, and for how long? | Grant and term | Lineage record per version | Term-limited; named family; perpetual for models trained in term |
| Distillation and output training | May covered models teach other models? | Permitted uses | Teacher and student records | Prohibited; allowed into covered set only; allowed freely |
| Synthetic and derived data | May the licensee generate and keep data from the records? | Derived data clause | Generation logs and similarity checks | Internal only; covered by same terms; prohibited |
| Embeddings and retrieval | May records be indexed and served back? | Permitted uses | Index inventory and deletion path | Training only; retrieval inside product; no retrieval |
| Evaluation and publication | May records or examples appear in papers or benchmarks? | Confidentiality and publication | Pre-publication review log | No publication; aggregate results only; approved excerpts |
| Weight release | May covered models be released or transferred? | Distribution | Release log with recipients | No release; hosted access only; release with notice |
| Affiliates and contractors | Who may exercise the rights? | Parties and sublicensing | Access list and flow-down terms | Licensee only; named affiliates; contractors under written terms |
| Termination and deletion | What happens to data, models and derived data at the end? | Term and termination | Deletion certificate covering each artifact | Delete data only; delete data and derived data; models survive |
| Audit and verification | How does the licensor check any of this? | Audit clause | Records retained for an agreed period | Self-certification; documents on request; independent audit |
Which proposed uses require an explicit scope decision rather than an assumption?
Any use that moves the data's value into an artifact the licensor never sees needs an explicit decision: distillation, synthetic data, retrieval indexes, publication and weight release. These are the uses most likely to be argued both ways if the agreement is silent.
Distillation is the clearest case. Some public model licenses already treat it as producing a derivative. The CreativeML Open RAIL-M license defines Derivatives of the Model to include any model created by transfer of patterns of the weights, parameters, activations or output of the Model, including distillation methods and methods based on synthetic data generated by the Model. Google's Gemma Terms of Use similarly define Model Derivatives to include models built through distillation or synthetic outputs, while stating that Outputs themselves are not Model Derivatives. A data license that ignores distillation is less specific than the model licenses many AI teams already work under.
Retrieval is the quiet one. A license that permits training may say nothing about storing each record as an embedding and returning near-verbatim passages in a product. That is a different use with different privacy and confidentiality exposure, and it should be named.
Publication and weight release are hard to reverse. Once records appear in a paper or weights are downloadable, no contract remedy restores the earlier position, which is why these uses usually call for prior written approval rather than after-the-fact notice.
What happens to derived models when the license ends or data must be deleted?
Only what the agreement says, so it must say what happens to each artifact: the data copies, derived data and the models. A common split is that data and derived data are deleted at termination, while models trained during the term survive with the restrictions that applied during the term. Other deals require retraining without the data, or deletion of specific models.
Regulators have shown that models can be part of a remedy. In its 2021 settlement with the photo app developer Everalbum, the Federal Trade Commission required deletion of models and algorithms the company developed using photos and videos uploaded by its users, in a case about deceptive use of facial recognition. That case concerned consumer protection law, not a commercial license, but it shows that a model can be treated as something built from data rather than as an independent asset.
For a licensor, the practical point is evidence. A deletion clause that covers models is only as useful as the record showing which models were trained on which data version. Ask for a model register in the agreement, and tie each entry to the dataset release it used.
Personal data deserves its own line. De-identification lowers risk but does not by itself create a right to license, and models can retain information from their training data. If the records involve people, the privacy reviewer should decide whether the deletion obligations reach models, not just files.
How do published model licenses handle the same derived use questions?
They answer them explicitly, which makes them useful reference points for drafting even though they license models rather than data. Three patterns are worth noting.
First, naming and notice. The Llama 3.1 Community License Agreement requires a licensee that uses the Llama Materials or any outputs or results of them to create, train, fine tune or otherwise improve an AI model that is distributed or made available to include Llama at the beginning of that model's name. That is a provenance obligation attached to output training, a concept a data license can borrow.
Second, ownership of derivatives. The same agreement states that, as between the licensee and Meta, the licensee owns the derivative works and modifications it makes. A data licensor may want the opposite allocation for certain artifacts, or at least a license back, and should decide that on purpose.
Third, flow-down. The Open RAIL-M license requires its use-based restrictions to be included as an enforceable provision in any agreement governing use or distribution of the model or its derivatives. A data license that permits sublicensing can require the same: restrictions on derived models travel with every downstream grant.
Worked example: a claims archive license with a distillation request
This example is illustrative. A regional insurer agrees to license de-identified claims notes for training a document classification model. The first draft permits training and says nothing else about models. Counsel works through the checklist and the parties agree on four points: covered models are those trained during a three-year term; successor versions trained during the term are covered; synthetic data may be generated for internal testing only; and weights may not be released outside the licensee.
Six months in, the licensee asks to distill the classifier into a smaller model for on-device use by a customer. Because the agreement names distillation as allowed only into the covered set, the student becomes a covered model with the same restrictions, including no release of weights. The customer can use the product, but the student's weights stay with the licensee.
The exception: the licensee also wants to use the classifier's labels to train an unrelated model for a different line of business. The agreement's permitted-use clause limits covered models to claims document classification, so this falls outside the grant. The parties can amend, at a new price if they choose, or decline. Either way, the question is answered by the text rather than by an argument about whether a label is a derivative work.
Which derived use requests should go back to counsel before you answer?
Stop if nobody can say who owns or controls the records, or if the records came to you under terms that limit onward use; derived model rights cannot be broader than the rights you hold. Narrow the scope when the licensee asks for open-ended model rights you cannot price or monitor: limit the covered models, shorten the term or exclude release.
Request specialist review when the records include personal or sensitive data, when the licensee is in another jurisdiction, when customers or partners contributed the records under their own contracts, or when the licensee asks for perpetual or transferable model rights. These questions turn on facts and law that a checklist cannot settle.
This guide is general information, not legal advice. Qualified counsel should review the specific agreement, the source of the records and the proposed uses, with a technical reviewer confirming that the evidence the agreement requires can actually be produced.
Settling derived model terms before a SourceX agreement is signed
In a SourceX license, derived uses are decided in one document: the company signs off on scope and terms before any record is delivered, and that text bounds what SourceX may later do with the de-identified dataset, including sublicensing and distribution. The checklist above is most useful during that pre-signature window.
Each deal's signed agreement also sets exclusivity and term, so review them together with the derived uses. A short assessment, with no records uploaded, is how fit is checked first.
Frequently asked questions
Does a license to train an AI model include the right to distill it?
Not necessarily. If the agreement does not define covered models or mention distillation, the answer depends on how the grant is read under the governing law, and the parties may disagree. Name distillation explicitly and say whether students are covered, restricted or prohibited.
Is a trained model a derivative work of its training data?
The question is unsettled. U.S. law defines a derivative work as a work based upon preexisting works, but whether model weights fit that definition depends on facts such as memorization and on how courts develop the law. Contracts should define their own terms rather than rely on copyright's definition.
Who owns a model trained on licensed data?
Usually the party that trains it, unless the agreement allocates ownership differently or requires a license back. The licensor's control typically comes from use restrictions, distribution limits and termination terms rather than ownership of the model itself.
Should derived models be deleted when the license ends?
It depends on the deal. Many agreements require deletion of data and derived data while letting models trained during the term survive with restrictions; others require deletion or retraining. Whatever is chosen, require a model register so deletion can be verified.
What evidence should a licensor ask for about derived models?
A register of covered models with the dataset version each used, records of distillation and synthetic generation, a log of any release of weights, and a deletion certificate that lists each artifact. Agree the retention period for these records in the audit clause.
Sources
- Title 17 section 101 defines a derivative work as a work based upon one or more preexisting works, such as a translation or adaptation. Source
- The Community Data License Agreement Permissive 2.0 defines Results to include machine learning models and imposes no restriction or obligations on Results. Source
- The CreativeML Open RAIL-M license defines Derivatives of the Model to include distillation and synthetic-data methods and requires its use restrictions to flow down to downstream agreements. Source
- Google's Gemma Terms of Use define Model Derivatives to include models built through distillation or synthetic data Outputs, and state that Outputs are not Model Derivatives. Source
- The FTC's final Everalbum order required deletion of models and algorithms developed using photos and videos uploaded by the app's users. Source
- The Llama 3.1 Community License requires distributed models trained with Llama outputs to carry Llama at the start of their name and states the licensee owns its derivative works as between it and Meta. Source
See if your company qualifies
A short company assessment. No data uploads are needed.