Rights and contracts
Can a buyer generate synthetic data from your licensed records?
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
A buyer may generate synthetic data from your licensed records only if the license grants it, because a training grant does not settle whether a generator can be kept or its output shared. Specify each derivative by name: who may receive it, whether it may leave the buyer, what evaluation is required first, and what is deleted or survives at termination. Silence invites a dispute.
License terms
AI dataset synthetic
Key takeaways
- Fitting a generator, keeping it and sharing what it samples are three separate permissions; a training grant settles none of them.
- Treat anything that can reproduce or substitute for your records as a derivative that inherits the source terms unless you grant more on purpose.
- Published licenses disagree sharply on derived outputs, from no restrictions on results to explicit limits on using outputs to train other models.
- Require a written evaluation report before any synthetic set reaches customers or the public.
- Ask for a lineage register so every derivative can be found, deleted and certified at termination.
Can a buyer generate synthetic data from your licensed records?
Only if the license says so. A grant to train models on your records does not by itself say whether the buyer may fit a generator to those records, keep the generator, sample new records from it, or pass those records to anyone else. Treat each of those steps as a separate permission, and decide it in writing before the records leave your systems.
The practical default for a supplier is simple: anything the buyer computes from your records that can still reproduce them, describe them record by record, or substitute for them in someone else's training run is a derivative, and it inherits the permitted use, confidentiality, deletion and audit terms of the source records. The license can then grant wider rights for specific derivatives on purpose, rather than by silence.
This page gives company legal teams a map for making those decisions one derivative at a time. It covers what the buyer can make, who may receive each output, what survives termination, and what evidence the buyer should hand back. It is general information for a negotiation, not legal advice; counsel should apply it to your agreement and the records involved.
Which derivatives can a buyer make from licensed records?
A buyer can make at least eight distinct things from licensed records, and they carry very different amounts of the source. Synthetic data tools are built to learn from the real data: the open-source Synthetic Data Vault, for example, describes its synthesizers as objects that learn patterns from real data and replicate them to generate synthetic data. That is exactly why the output is not automatically free of the input.
The table separates the derivatives a license should name. The default position in the last column is a starting proposal for a supplier, not a market standard; buyers will negotiate it, and the right answer depends on how sensitive and how rare the records are.
| Derivative | What it is | What it can carry from your records | Default to propose |
|---|---|---|---|
| Trained model weights | The model the license was signed to build | Memorized fragments and patterns that can sometimes be extracted by querying | Licensed use, governed by the model terms you already negotiated |
| Generator model | A model fitted specifically to reproduce the distribution of your records | Close copies of rare or unique records unless privacy controls are applied | Internal use only; no transfer, no hosting for third parties |
| Synthetic records | New rows or documents sampled from a generator | Your field structure, value ranges, correlations and sometimes near-copies | Same terms as source records unless an evaluation report clears a wider use |
| Distilled or student models | A smaller model trained on outputs of the licensed model | Behavior learned from your records, one step removed | Same scope as the licensed model, including any field-of-use limits |
| Labels and annotations | Human or machine labels attached to your records | Your content plus new judgments about it | Usable only alongside licensed records; deleted with them |
| Embeddings and search indexes | Vector or keyword representations of records | Enough to retrieve or approximate source text | Internal use; deleted on termination |
| Evaluation items | Test cases built from your records | Verbatim or lightly edited records | Internal evaluation only; publication needs separate approval |
| Aggregate statistics | Counts, rates and distributions across many records | Business facts about your operations, not individual records | Permitted if thresholds prevent small groups and the business facts are not confidential |
Why a training grant does not settle synthetic generation
Because licenses that address derived outputs draw the line in different places, silence leaves the answer to argument. Three published licenses show how far apart the positions can be.
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 models' insights, and states that it imposes no restriction or obligations on the use, modification or sharing of Results. Under a clause like that, a buyer could argue that a generator and everything it samples are unrestricted Results.
The Open Database License takes the opposite approach for databases. It defines a Derivative Database as a database based on the original, including extracting or re-using a substantial part of its contents in a new database, and separately defines a Produced Work as a work resulting from using a substantial part of the contents. The two categories carry different obligations, which shows that a license can treat a derived table differently from an output such as a chart or a report.
Model licenses also regulate outputs directly. The Llama 2 community license, for example, says the licensee will not use the Llama materials or any output or results of them to improve any other large language model. That is a contractual restriction on using outputs to train something else, which is the same kind of term a data supplier may want for synthetic records.
The lesson for a supplier is not to copy any of these texts. It is that each one decided, explicitly, whether derived outputs travel with the source terms. Your data license should make the same decision for each derivative you care about, in plain words, and define the derivative by what it is and how it was made rather than by copyright categories alone.
The derivative use and downstream rights map
The map below is the asset to attach to a term sheet or a data schedule. Each row is a derivative, and each column is a recipient or a point in time. Fill every cell with one of four answers: permitted, permitted with conditions, needs written approval, or prohibited. Blank cells are where disputes start.
The entries shown are an illustrative supplier-protective starting position for a confidential operational dataset. A less sensitive dataset, or a higher price for broader rights, can move cells toward permitted.
| Derivative | Buyer internal use | Contractors and affiliates | Customers or public | After termination |
|---|---|---|---|---|
| Generator model | Permitted for licensed purpose | Contractors under written duties only | Prohibited | Delete and certify |
| Synthetic records | Permitted for licensed purpose | Contractors under written duties only | Needs written approval plus evaluation report | Delete unless approval said otherwise |
| Trained model weights | Permitted | Per model clause | Per model clause, never with training data attached | Per survival clause in the agreement |
| Distilled or student models | Permitted within licensed field | Same as licensed model | Same as licensed model | Same as licensed model |
| Labels and annotations | Permitted with records | Contractors doing the labeling only | Prohibited | Delete with records |
| Embeddings and indexes | Permitted | Hosting providers only | Prohibited | Delete and certify |
| Evaluation items | Permitted | Evaluation contractors only | Needs written approval per item | Delete or return |
| Aggregate statistics | Permitted | Permitted | Permitted above an agreed minimum group size | May be kept |
Which buyer requests need a written yes or no before signing?
Any use that moves a derivative outside the buyer's own control, or that lets it replace your records in someone else's training run, needs an explicit decision. Assumptions are tolerable for internal processing steps; they are not tolerable for anything that leaves the building.
Put these questions to the buyer in writing and record the answer in the agreement or its data schedule:
- Will the buyer keep a generator model fitted to your records, and if so, who can query it and for how long?
- Will synthetic records be shared with a customer, sold, contributed to a data marketplace or released openly?
- Will synthetic records be mixed with other suppliers' data so that your share can no longer be separated or deleted?
- Will synthetic or distilled outputs be used to train models for a field of use or a market the license excludes?
- Will any derivative be used to build a public benchmark, a demo or a paper appendix?
- Will the buyer run re-identification or membership tests, and will you see the results?
- What happens to each derivative when the license ends, is terminated for breach, or the buyer changes control?
How do you check whether synthetic output still carries your records?
Ask for an evaluation before any synthetic set moves beyond internal use, because generated data can still reproduce what it learned from. Research on large language models has shown that an adversary can query a model trained on private data and recover individual training examples, including sequences that appeared in only one training document. A generator fitted closely to a small or unusual dataset raises the same concern.
European regulators take the same view of trained models. The European Data Protection Board's Opinion 28/2024 says a model trained on personal data cannot always be treated as anonymous, that the likelihood of extracting personal data directly or through queries must be insignificant, and that a mere assertion of anonymity is not enough. Your records may be business records rather than consumer data, but they often contain employee names, customer contacts and free text.
Tooling for this kind of evaluation exists. The Synthetic Data Vault includes features to compare synthetic data with the real data and produce a quality report, and NIST's SDNist package evaluates the utility and privacy of a deidentified dataset and produces a summary report. A license does not need to mandate a tool, but it can require a written report covering:
- Exact and near-duplicate matches between synthetic and source records, with the matching rule stated.
- Rare-value checks: unique customer names, addresses, part numbers, prices or free-text phrases that appear in the output.
- A membership or extraction test appropriate to the generator, with its method and limits described.
- Utility measures the buyer relies on, so you can see what the synthetic set is meant to replace.
- The generator configuration, version and training data snapshot the report applies to.
What evidence and audit duties belong in the license?
Write evidence duties for every derivative you allow, because deletion and audit rights only work if the buyer can show what it made. A derivative nobody recorded cannot be deleted on termination with any confidence.
Regulators have shown that derived models can be reached when the underlying data was obtained unlawfully: the FTC notes it has required businesses that unlawfully obtained consumer data to delete products, including models and algorithms, developed in whole or in part from that data. That is a consumer protection remedy, not a contract term, but it illustrates that lineage from data to derivative is something parties should be able to prove.
A workable evidence package asks the buyer to keep and produce on request:
- A lineage register listing every generator, synthetic set, distilled model, index and evaluation set built from your records, with dates and owners.
- The evaluation report for each synthetic set that left internal use, tied to the generator version.
- Notices passed to any approved recipient stating the use limits that travel with the derivative.
- Deletion or return certificates naming each derivative destroyed at termination.
- Confirmation of which models trained on synthetic records, so model survival terms can be applied consistently.
Worked example: work orders, a synthetic augmentation set and one exception
The company and request below are invented to show the method. A commercial field service company licenses three years of work orders, technician notes and parts usage to a developer building a dispatch assistant. Midway through, the buyer asks to fit a generator to the work orders, sample several times as many synthetic work orders to improve rare fault coverage, and later license a synthetic benchmark to its own customers.
Walking the map produces three different answers. Fitting the generator and training on synthetic work orders internally is permitted for the licensed purpose, on condition that the generator stays internal and is deleted at termination. Licensing the synthetic benchmark to customers moves synthetic records to the customers column, so it needs written approval and an evaluation report; the supplier asks for duplicate and rare-value checks on technician notes and customer site names before deciding. Using synthetic work orders to train a model for a separate building-security product falls outside the licensed field, so it is prohibited.
The exception is aggregate statistics. The buyer also wants to publish average repair times by equipment category in a white paper. If the categories are broad, the minimum group size is met and the figures do not reveal confidential pricing or customer identities, the map permits it without a separate approval. If a category contains one large customer, the figure is effectively that customer's data and should be suppressed or merged.
The limit of the recommendation is that the map handles rights, not quality or privacy law. An approved synthetic set can still be unfit for its purpose, and a set that passes a duplicate check can still raise privacy questions that counsel and a privacy reviewer need to assess.
When should a synthetic data request pause the deal?
Stop and narrow the scope when the buyer cannot say which derivatives it will make, or when it insists that all outputs are unrestricted results. Agreeing to an undefined derivative right is the same as agreeing to an undefined license.
Request specialist review before signing in these situations:
- The records include personal data, employee information or regulated categories, so privacy counsel must assess whether synthetic output is still personal data.
- Your records contain customer information you hold under confidentiality obligations to others, which may limit any derivative sharing.
- The buyer plans open release of a generator, synthetic set or model trained on them.
- The buyer is a competitor, or plans to serve your competitors with outputs built from your records.
- The dataset is small, highly specific or contains rare values that a generator is likely to reproduce.
- The agreement crosses jurisdictions, so the treatment of derived data and anonymity may differ.
The derivative map as a schedule to the license
SourceX licenses business data directly from companies and runs the rights review and contracting, which makes the signed agreement the natural home for the derivative map. The company approves scope and terms and signs before delivery, and SourceX's own use, sublicensing and distribution of the de-identified dataset must then stay within what was signed, so record each permitted and prohibited derivative there in plain words.
Whether the records could be licensed at all is checked through a short company assessment; nothing is uploaded for it.
Who decides whether licensed records may seed synthetic examples?
The supplier decides, through the person who owns the records and with counsel, one record class at a time and against the signed license first. Where the license names synthetic generation for a class of records, the request is checked against that grant; where the license is silent or rules it out, the working answer for that class is no until a signed amendment changes it. Whether the buyer can technically run a generator on records it already holds plays no part in the test.
A second layer sits under the license. Some licensed records carry rights the supplier never held outright, such as messages its customers wrote or files a vendor supplied, and the supplier can release those as seeds only as far as the terms behind them allow. Generation adds two problems that a plain training review can pass over: excluded text hides inside permitted entries, and rare values can come back out of a generator verbatim. So the decision has to work at the level of entries, spans and values, not tables or tickets. The derivative map above settles what a buyer may make; this part settles which inputs the supplier may release for making it, how each exclusion rides along with the delivery, and what the decision record holds.
Which records can be released as seeds, and at what level must an exclusion work?
Authorship usually sits on each message or entry, not on the ticket or the table, so that is the level at which a seed decision can be enforced. osTicket, an open-source support ticket system, shows the pattern. At commit 8d38b06 (June 17, 2026), its thread_entry table has 19 columns and stores a one-letter type on every entry: M for messages, R for staff responses, N for notes. Emailed messages from the ticket owner and from collaborators copied on the ticket are both filed as type M, so the type cannot tell them apart; a flag bit, 0x0020, commented "Message from collaborator", or a comparison of the entry's user_id with the ticket owner's can. An agent who answers by email is filed as a note.
Quoted history is the second trap. With its default settings, osTicket cuts quoted text from an emailed message only when the separator "-- reply above this line --" appears in the body with text above it; otherwise whatever earlier messages the sender's mail client quoted stay inside the entry. An exclusion written as "no customer text" therefore cannot be enforced by entry type alone. This reading describes what the software can store, not any company's data.
The matrix applies that to a licensed set. Its default decisions are editable assumptions for a confidential operational dataset, not measured results or legal conclusions; counsel replaces each one with a reading of the actual documents:
| Record class | Governing document | Evidence that settles it | Default seed decision (editable) | How the exclusion travels into delivery |
|---|---|---|---|---|
| Staff responses and internal notes | The license grant, plus staff privacy notices and workplace monitoring policy | Redaction record for staff names and signatures, including agent email replies filed as notes | Permit once names and signatures are removed | Redaction log keyed to entry IDs, delivered with the data |
| Messages written by customers | The customer agreement or terms of service and any data processing addendum | Counsel's reading of the data-use clause; restrictions or opt-outs on file | Exclude unless the agreement allows AI derivative use | A seed flag on each entry in the delivery manifest plus a list of excluded entry IDs |
| Messages from people copied on the thread | Usually none between the supplier and the author | The collaborator flag, or the entry's author compared with the ticket owner, checked on every entry | Exclude | The same seed flag, set from the collaborator test rather than the entry type |
| Quoted or forwarded text inside an entry | Whatever governs the original author | Separator and quote-block detection on each body, with a sampled miss rate | Strip the quoted span and keep the author's own text | Span removed before delivery; rule and sample result in the review log |
| Attachments from vendors or customers: manuals, drawings, price lists | The vendor or customer terms the file arrived under | File origin traced through the attachment-to-entry link | Exclude from seeding unless those terms allow it | Withhold the file, or ship it in a partition marked not for seeding |
| Rare values: one-customer part numbers, unique fault codes, unusual phrases | Customer confidentiality terms and the license's own confidentiality clause | A count of how many records carry each value across the licensed set | Mask values seen fewer times than an agreed floor before any seeding | Masked tokens in the seed copy; the floor and the masked count in the review log |
| Records under deletion or opt-out requests | The request itself, the privacy notice and any customer termination clause | A dated register of requests matched to record IDs | Exclude, and re-check before every generation run | A versioned excluded-ID list the buyer re-pulls before each run |
If the buyer already holds the records and the names are gone, may it generate?
Not on that basis. Holding the records proves the buyer can run a generator, and redaction lowers identifiability; neither shows that whoever controls the records agreed to the use. Official texts in privacy law and database law both treat a change of form as leaving the underlying question open.
On privacy, the UK Information Commissioner's Office says pseudonymisation "is effectively only a security measure. It does not change the status of the data as personal data." Its current pseudonymisation guidance makes the same point (paraphrase): pseudonymising reduces risk and improves security but does not move data outside the law. Stripping names from seed records is a safeguard to log, not a permission to cite.
On databases, the Court of Justice held in Directmedia Publishing (C-304/07, judgment of 9 October 2008) that a transfer following on-screen consultation and "an individual assessment" of the material "is capable of constituting an 'extraction'" where it takes a substantial part of the protected contents, judged qualitatively or quantitatively. In that case (paraphrase), the publisher had used a list as a guide, leaving out some entries and adding others. The ruling says nothing about synthetic generation; whether the EU right protects a given set of records at all, and whether seeding a generator counts as an extraction, are questions for counsel in each jurisdiction.
So a buyer who says the records are already on its servers and the names are gone has described two safeguards, not a grant. The review log should record those safeguards in a field of their own, apart from the permission.
What goes in the review log for one synthetic-example request?
One entry per request, completed before the answer goes back, with every field filled or marked not applicable. It is the document that later shows who decided, on which paper, on what evidence, and which records stayed out and by what means.
The log also matters if the records are offered for licensing again. SourceX's rights review confirms provenance, supplier permissions and the intended use, and data already licensed for AI training is not eligible for a SourceX license: the supplier confirms it has not been licensed for AI training before. Records a buyer already holds under a training license, synthetic rights included, cannot pass that confirmation, and a log naming which entries went where is how a supplier separates them. For a first read on fit, the estimator on the SourceX homepage asks a few questions with no uploads; its estimate is indicative, not an offer or a promised payout.
The fields, with an invented entry for a support-ticket license:
| Log field | What it must say | Illustrative entry (invented) |
|---|---|---|
| Request | The buyer's own words, the date received and the use it names | Generate synthetic troubleshooting dialogues from licensed tickets for internal fine-tuning; received 2026-10-21 |
| Records in scope | Record classes and entry IDs or manifest ranges, with counts | Message, response and note entries from the licensed export; attachments listed separately |
| Transformation | Method, generator, where it runs, and whether records leave the buyer | Few-shot prompting on a self-hosted model; no third-party API |
| Outputs and recipients | What will be produced and who receives it | Synthetic dialogues for internal training only; no customer or public release |
| Governing documents | The license clause, then every upstream document checked, with version dates | License permitted-use clause; customer terms of service in force when the tickets were written; vendor manual license |
| Evidence considered | What was read or tested, by whom, with results | Counsel note on customer terms; collaborator-flag count; quoted-span check on a sample of entries; rare-value counts; deletion register |
| Decision by record class | Permit, permit with conditions, exclude, or refer to counsel | Staff responses permitted after redaction; customer and copied-party messages excluded; vendor attachments excluded; rare values masked below the floor; open deletion requests excluded |
| Exclusion carrier and check | How each exclusion travels and how compliance will be tested | Seed flag in the manifest; the buyer's run log lists seed entry IDs for a spot check against the excluded list |
| Approvers | Roles that signed off, each with a date | Data owner, outside counsel and privacy lead, each dated |
| Re-review trigger | What reopens the decision | A new generator, any third-party API, a new recipient or a new record class |
Sources for this section
- osTicket's installer schema at commit 8d38b06 (June 17, 2026) defines a thread_entry table of 19 columns, including staff_id, user_id, a one-character type, flags, body and recipients, and a collaborator role comment reading: M => (message) clients, N => (note) 3rd-Party, R => (reply) external authority. Source
- osTicket's thread code at the same commit gives message, response and note entries the types M, R and N; files emails from collaborators and from the ticket owner as type M, with collaborators marked by flag 0x0020 (Message from collaborator); files emails from staff as notes; links each attachment to its entry (attachment type H); and strips quoted text only when the configured separator is found in the body with text above it. Source
- osTicket's default English install settings turn quoted-reply stripping on and set the reply separator to '-- reply above this line --'. Source
- The ICO says pseudonymisation is effectively only a security measure and does not change the status of the data as personal data. Source
- The ICO's pseudonymisation guidance (paraphrase) describes pseudonymisation as a way to reduce risk and improve security, not a way of transforming personal data so that the law no longer applies. Source
- In Directmedia Publishing (C-304/07, 9 October 2008) the Court of Justice held that a transfer from a protected database to another after on-screen consultation and an individual assessment is capable of constituting an extraction where it amounts to a substantial part of the contents, evaluated qualitatively or quantitatively. Source
Frequently asked questions
Does a license to train AI models let the buyer create synthetic data?
Not clearly. A training grant covers building models, but fitting a generator, keeping it and sharing what it produces are separate acts that the license should address by name. If the agreement is silent, the parties are left to argue about intent, so decide it before signing.
Is synthetic data generated from my records still my data?
Ownership is mostly set by the contract, not by who ran the generator. A supplier-protective license treats synthetic records as derived data that inherit the source terms unless a wider use is approved. Whether synthetic output is still personal data is a separate privacy question for counsel.
Can a buyer sell or share synthetic data made from my records?
Only if the license allows it. A sensible position is to require written approval and an evaluation report before any synthetic set reaches customers, marketplaces or the public, because generated records can still reproduce rare or unique source values.
What should happen to synthetic data and generator models when the license ends?
Spell it out per derivative. Generators, indexes and labels usually should be deleted and certified, while trained model survival follows the model clause you negotiated. Synthetic sets that were approved for wider use need their own rule, because copies may already sit with third parties.
How do I know whether synthetic data leaks my original records?
Ask for a written evaluation covering exact and near-duplicate matches, rare-value checks and an extraction or membership test suited to the generator. Open tools such as the Synthetic Data Vault and NIST's SDNist produce quality and privacy reports, though no single test proves that nothing leaks.
Sources
- The Synthetic Data Vault describes its synthesizers as objects that learn patterns from real data and replicate them to generate synthetic data, and includes tools to compare synthetic data with real data in a quality report. Source
- NIST's SDNist package evaluates the utility and privacy of a deidentified dataset and generates a summary quality report. Source
- CDLA-Permissive-2.0 defines Results as any outcome of computational analysis of the data, including machine learning models, and imposes no restriction or obligations on Results. Source
- The Open Database License defines Derivative Database and Produced Work as separate categories with different obligations. Source
- The Llama 2 community license states that the licensee will not use the Llama materials or any output or results of them to improve any other large language model. Source
- Carlini et al. (USENIX Security 2021) showed that an adversary can query a language model trained on private data to recover individual training examples, even sequences that appeared in a single training document. Source
- EDPB Opinion 28/2024 says AI models trained on personal data cannot always be considered anonymous, extraction must be insignificant, and a mere assertion of anonymity is not enough. Source
- The FTC notes it has required businesses that unlawfully obtained consumer data to delete products, including models and algorithms, developed in whole or in part using that data. Source
See if your company qualifies
A short company assessment. No data uploads are needed.