Rights and contracts
Can a buyer publish examples from your private dataset?
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
A buyer may publish examples from your private dataset only if the license grants that tier of publication, and the safe default is no. Training or evaluation rights do not include public disclosure. Define tiers from aggregate scores to verbatim items and benchmark releases, require written approval of the exact items, and block anything that reveals trade secrets, personal data or third-party confidences.
Evaluation split
AI dataset benchmark
Key takeaways
- Publication is the one derived use that cannot be undone, so it needs its own grant.
- Name tiers, from aggregate scores to verbatim excerpts and benchmark releases, rather than treating publication as one act.
- Approve exact items in writing, with silence meaning refusal.
- Published items may end up in future training data and can end trade secret protection for what they reveal.
- Treat attribution as a separate approval from the content of any example.
Can a buyer publish examples from your private dataset?
Only if the license allows it, and the safest default is that it does not. A license to train or evaluate models on your records says nothing clear about printing those records in a paper, a model card, a sales demo or a public benchmark. Publication should be a separate right with its own tiers, its own approval step and its own confidentiality checks.
Publication differs from every other use because it cannot be reversed. Records inside a buyer's systems can be deleted at termination; records on a website, in a paper appendix or in a public benchmark cannot be recalled, and copies may be collected into other datasets and models.
This page gives company legal teams a checklist for deciding publication requests: what counts as publishing, which items block a request, how approval should work and what to do when a buyer wants to turn your records into a benchmark. It is general information, not legal advice; counsel should apply it to your agreement and records.
What counts as publishing for AI dataset benchmark publication rights?
Publishing is any disclosure of your records, or of material derived from them, to people outside the parties and their bound contractors. It ranges from harmless to irreversible, so the license should name tiers rather than treat publication as one act.
At a glance
What counts as publishing for AI dataset benchmark publication rights
The tier names below are this page's own scale, not an industry standard; use whatever names fit your agreement, but keep the distinctions. A license that allows only Tier 0 and Tier 1 without approval, and requires written approval for everything above, is a reasonable supplier-protective starting position.
A workable set of tiers, from least to most exposure:
- Tier 0, aggregate scores: a model's accuracy on a private evaluation set, with no items shown and no description that identifies you.
- Tier 1, described examples: a prose description of the kind of task, written so no individual record can be reconstructed.
- Tier 2, paraphrased or synthetic examples: new items written to resemble yours, checked so they do not reproduce rare values.
- Tier 3, redacted excerpts: real records with names, numbers and free text removed or replaced, shown in a paper, model card or demo.
- Tier 4, verbatim excerpts: real records shown as they are, even a few.
- Tier 5, item release: a public or gated benchmark containing test items built from your records.
- Attribution, at any tier: naming your company as the source, which can reveal a business relationship even when no record is shown.
Why publication is the hardest derived use to undo
Publication can end the protection your records rely on. Under the federal trade secret definition, information is a trade secret only if its owner has taken reasonable measures to keep it secret and it derives value from not being generally known. A verbatim excerpt of a pricing rule or an underwriting note in a public paper is, from then on, generally known.
It can also feed future training. The BIG-bench project tells task authors that anything accessible online and not marked with its canary string should be considered potentially part of the training data of future models, and asks them to include the canary string when they publish any part of a task elsewhere. Research has shown that models can memorize and emit training sequences, including ones that appeared in only one document, so a published example can resurface in other models' outputs.
Personal data adds a separate test. The European Data Protection Board's Opinion 28/2024 stresses that personal data in training material can persist and be extracted, and that claims of anonymity need evidence. A redacted excerpt that still contains a rare job title, location and date may identify a person.
Transparency duties are not a reason to publish records. The EU AI Act requires providers of general-purpose AI models to publish a sufficiently detailed summary about the content used for training, and its recitals describe that summary as generally comprehensive rather than technically detailed. A buyer can meet such a duty with a description of sources at an agreed level, not with your records. Agree that description level in the license.
The publication approval and confidentiality issue checklist
The checklist below is the asset. Run it on every publication request above Tier 0, and record the answer to each row. Any row in the last column that applies blocks publication until it is resolved.
The rows are written for a supplier's legal team working with the record owner and a technical reviewer.
Two rows deserve the most care. The trade secret row needs someone who knows the business, because a harmless-looking item can expose a pricing rule, an approval threshold or a supplier relationship that only an insider would recognize. The rare-value row needs someone who can search the source records, because a value that appears once in your data is much easier to tie back to a person or a customer than a common one.
Keep the completed checklist with the approval. If a question arises later, the record shows which items were reviewed, by whom, and why each was approved or refused.
| Check | Question to answer | Who answers | Blocks publication if |
|---|---|---|---|
| License grant | Does the agreement allow this tier of publication at all? | Counsel | The tier is not granted or the agreement is silent |
| Exact items | Has the buyer supplied every item, excerpt or screenshot it will publish? | Buyer, checked by record owner | Anything is described but not provided |
| Trade secrets | Does any item reveal pricing, margins, methods, rules or customer lists? | Record owner | Yes, unless removed and rechecked |
| Third-party confidentiality | Is any content held under confidentiality to a customer, supplier or employee? | Counsel | Yes, unless that party has consented |
| Personal data | Could any item identify a person directly or in combination? | Privacy reviewer | Yes, or the reviewer cannot rule it out |
| Rare values | Do items contain unique names, codes, amounts or phrases? | Technical reviewer | Yes, for verbatim or redacted tiers |
| Attribution | Will your company be named or identifiable as the source? | Counsel and communications | Naming was not separately approved |
| Contamination | Will publishing items weaken a private evaluation you rely on? | Technical reviewer | Items come from an active held-out set |
| Venue and format | Paper, model card, demo, website or benchmark, and is access gated? | Buyer | Venue or access differs from what was approved |
| Record of decision | Is the approval written, dated and tied to the exact items? | Counsel | Approval was oral or covers unnamed items |
How should the approval process work?
Approval should be written, item-specific and time-boxed. The buyer submits the exact material, you review it against the checklist within an agreed period, and silence means refusal rather than consent. Deemed approval is convenient for the buyer and dangerous for you, because a missed email becomes a permanent disclosure.
Agree the review period in the license, long enough for the record owner and a privacy reviewer to look at every item, and agree that a buyer's conference or launch deadline does not shorten it. Deadline pressure is an easy way for items to be approved before anyone has fully reviewed them.
A practical process has these steps:
- The buyer submits a request naming the tier, the venue, the audience, the publication date and every item it intends to show.
- The record owner and a technical reviewer run the checklist and mark each item approved, approved with edits or refused.
- Counsel confirms the license grant, third-party confidentiality and any attribution question.
- Approval is issued in writing, lists the approved items by identifier and states that changes need a new request.
- The buyer confirms after publication what was published and where, so the record matches reality.
- Approvals are logged so later requests can reuse decisions about the same items instead of starting over.
What if the buyer wants to release a benchmark built from your records?
Treat a benchmark release as the highest tier, because it publishes many items at once and invites wide reuse. The default should be no public item release; if the buyer needs others to reproduce results, prefer a private evaluation the buyer runs on request, or gated access under terms you approve.
Gated distribution is one option some buyers suggest. On the Hugging Face Hub, for example, dataset authors can enable access requests so users must share their contact information to access the files, and access is granted to individual users rather than to whole organizations. Gating limits who downloads items, but it does not stop an approved user from copying them, so it reduces exposure rather than removing it.
Consider also what a benchmark does to your own evaluation. Once items are public, they can no longer test whether a model learned a skill or simply saw the answers, so a supplier that uses the same records to evaluate other buyers' models loses that option for the published items.
If you do approve a release, ask for:
- Only synthetic, paraphrased or fully redacted items, never verbatim records.
- A dataset card describing contents, intended use and limits; the Hugging Face guidance describes dataset cards as the place to explain how a dataset should be used.
- A canary string or similar marker so model developers can filter the items out of training corpora.
- A held-out private split that is never published, so you keep a clean evaluation.
- A takedown commitment for any item later found to reveal confidential or personal information.
Worked example: a claims-handling benchmark and one exception
This example is invented for illustration. A commercial insurance agency licenses de-identified claims notes and adjuster decisions to a developer for evaluation of a claims-handling assistant. Six months later the developer asks to publish a paper with twelve example items and to release a two-hundred-item benchmark built from the same notes.
Running the checklist on the paper request: the license allows Tier 2 but not Tier 3 or 4, so redacted real excerpts are out. The developer rewrites the twelve items as synthetic examples. The technical reviewer finds two that still contain a rare combination of vehicle type, town and loss date, and they are replaced. The agency declines attribution. Approval is issued for ten listed items, Tier 2, one named venue.
The benchmark request fails the contamination row, because the items come from the agency's active held-out set, and the license grant row, because item release was never granted. The agency offers a private evaluation run instead, with aggregate scores publishable at Tier 0.
The exception is aggregate scores. The developer may report its assistant's accuracy on the private set without further approval, because the license grants Tier 0 and no item or source is disclosed. If the paper described the set so specifically that readers could identify the agency, attribution would apply and approval would be needed again.
The limit of the checklist is that it governs the parties' contract. It does not settle whether a given redaction meets privacy law in a particular jurisdiction, or whether a regulator would view a disclosure differently; those questions go to counsel and the privacy reviewer.
When should a publication request stop or go to specialists?
Stop when the buyer will not show the exact items before publication. Approving a description of examples, rather than the examples, is approving something you have not seen.
Commitments in public materials also bind the publisher. The FTC has warned model-as-a-service companies that they must abide by their commitments to customers regardless of how or where the commitment was made, which is one reason buyers should not publish anything that conflicts with what they promised data providers.
Narrow the scope when only some items are a problem. Approving eight clean items and refusing two is a better outcome for both sides than refusing the whole request, and it keeps the buyer coming back through the process instead of around it.
Request specialist review when:
- Items contain personal data, health, financial or employee information, even after redaction.
- Records include customer or supplier material held under confidentiality obligations.
- The buyer asks for attribution, a case study or a joint announcement.
- The request is a public benchmark release or an open dataset.
- The publication would describe your methods, pricing or rules in enough detail to help a competitor.
- The buyer cites a legal transparency duty as the reason to publish records.
Publication tiers as a term of a SourceX license
SourceX licenses business data directly from companies, with rights review and contracting along the way. Publication tiers and the item-approval duty should be written into the agreement the company signs after approving scope and terms and before any records go out; that same agreement limits how SourceX itself may use, sublicense and distribute the de-identified dataset.
A short company assessment, with no data uploaded, shows whether the records could qualify.
Frequently asked questions
Can a buyer publish examples from my data if the license only allows training and evaluation?
Not safely. Training and evaluation rights do not clearly include disclosure to the public, so a buyer relying on silence is taking a risk. A license should state which publication tiers are allowed and require written approval of the exact items above aggregate scores.
Can a buyer publish benchmark scores on my private evaluation data?
Aggregate scores that show no items and do not identify you are the lowest-risk form of publication, and a license can allow them while still restricting every higher tier. Spell out the permission, and require approval again if the description would make your company identifiable.
Does redacting names make a published example safe?
Not always. A combination of rare details such as a job title, location and date can identify a person or reveal a customer, and business rules or prices can remain visible. Run the full checklist, including a rare-value check, on every redacted item.
Should I let a buyer name my company as the data source?
Treat attribution as its own approval. Naming can reveal a business relationship, invite questions from customers and conflict with confidentiality duties to others, so decide it separately from the content of any example.
How can a published benchmark avoid ending up in AI training data?
It cannot be fully prevented once items are public. Gated access, synthetic or redacted items, a canary string that developers can filter on and a private held-out split all reduce the risk. The BIG-bench project, for example, asks authors to include its canary string when publishing task content.
Sources
- Under 18 U.S.C. 1839(3), information is a trade secret only if the owner has taken reasonable measures to keep it secret and it derives independent economic value from not being generally known. Source
- BIG-bench advises that anything accessible online and not marked with its canary string should be considered potentially part of future training data, and asks authors to include the canary string when publishing task content elsewhere. Source
- Carlini et al. (USENIX Security 2021) showed that training examples can be extracted from a language model by querying it, even sequences present in just one training document. Source
- EDPB Opinion 28/2024 says personal data used in training may persist in a model and be extracted, and that a mere assertion of anonymity is not enough. Source
- The EU AI Act requires providers of general-purpose AI models to publish a sufficiently detailed summary of training content, described in the recitals as generally comprehensive rather than technically detailed. Source
- On the Hugging Face Hub, dataset authors can enable access requests so users must share contact information to access files, and access is granted to individual users rather than organizations. Source
- Hugging Face describes dataset cards as documentation that helps users understand a dataset's contents and how it should be used. Source
- The FTC says model-as-a-service companies must abide by their commitments to customers regardless of how or where the commitment was made. Source
See if your company qualifies
A short company assessment. No data uploads are needed.