Skip to main content

7. Licence, terms, and privacy

Question: Under what licence and terms is a dataset released for model training?

Legal review

The licences and terms in this chapter are examples of practice. They are not legal advice. The organization's legal office and its disclosure control unit confirm the licence, the terms, and the release of any dataset under the organization's statistical law and data protection rules.

Rationale​

A licence says what a user may do with a dataset. For training data the question that matters is whether training a model, and distributing the model, is allowed and under what conditions. The open licences statistical organizations already use (Creative Commons Attribution 4.0 is the World Bank's default) permit copying, transformation, and redistribution with attribution, and a model is a transformation. Creative Commons has said that its licences do not restrict uses that copyright law already allows, such as text and data mining where a law permits it, and do not add rights where it does not. The practical consequence is that a statistical organization states, in plain words beside the licence, that training is permitted and how attribution is to be given by a model, so that no user has to guess.

The other question is privacy. A training dataset is built from records, and records can identify people. The rules of the microdata access policy apply unchanged: a dataset is built from a public use file, from aggregates, or from synthetic data, and the disclosure control unit reviews it before release. Free text is the part that needs most care, because a job description can name an employer or a place. The recipes of this chapter choose the licence, write the terms, make both machine-readable, and record the privacy review.

Target state​

  • Every dataset carries an open licence, stated as a URL in the record, the card, and the loading record, with a plain statement that training and evaluation are permitted.
  • The terms state the attribution form for publications and model cards, and prohibit re-identification.
  • The disclosure control unit has reviewed the dataset, including its free text, and the review is recorded in the card.

Maturity levels​

LevelSteps
FoundationalChoose the licence; write the terms; record the privacy review in the card.
AI-readyPut the licence URL in the Croissant record and the catalog record; add the free-text screening to the release checks; state the attribution form for model cards.
AI-nativeTrack models that cite the dataset (chapter 8) and review the terms yearly against what the organization sees in use.
Running example

The occupation-coding dataset is published under CC BY 4.0 with a statement that training and evaluation are permitted, the attribution form to use in a model card, and the prohibition of re-identification. Its job descriptions were screened for names, employers, and places before release, and the review is recorded in the card.

Recipes​

Recipe

7.1 Choose the licence and write the terms

Level
Foundational
Skills
Dissemination lead; legal
Indicative time
Half a day

Result: the dataset has a licence and a short terms text that says, in words a model builder reads, what the licence permits and what the organization asks in return.

  1. Choose the licence. CC BY 4.0 is the default for the organization's open data, and a training dataset is open data. A licence that restricts commercial use or derivatives restricts models too, and the organization chooses it only on purpose.
  2. Write the terms, in the card's licence section:
    • training and evaluating models on the data is permitted;
    • attribution is given in the form of the citation, in any publication, model card, or product that used the data;
    • re-identification of any person or enterprise is prohibited, and so is linking the data to other data for that purpose;
    • the organization asks that models trained on the data cite it in their model card.
  3. Where the organization's legal counsel wants a specific statement on machine-learning use beyond the licence, write it as a separate terms document and link it. The licence itself stays a standard one that tools recognize.
  4. Record the decision in the AI-use policy of the organization (dissemination cookbook, chapter 8).

Worked example. The licence section of the running example's card reads: "This dataset is published under the Creative Commons Attribution 4.0 licence. Training and evaluating models on it is permitted. Please attribute it as 'National Statistics Office (example), LFS occupation coding, version 1.0.0' in any publication, model card, or product that used it. Any attempt to identify a respondent or an employer, including by linking these records to other data, is prohibited. Models trained on this dataset are asked to cite it in their model card." Legal counsel reviewed the four sentences and asked for no separate terms document, and the decision is recorded in the AI-use policy under datasets for model training.

What this does not do. A licence grants rights it has, and the terms ask for behaviour. Neither is enforced by the files. The use tracking of chapter 8 shows whether attribution happens, and the organization's monitoring of answers shows what models say.

Recipe

7.2 Make the licence and the terms machine-readable

Level
AI-ready
Skills
Catalog administrator; developer
Indicative time
An hour

Result: every system that reads the dataset's records reads the same licence, in the Croissant record, the card's header, the catalog record, and the dataset page.

  1. In the Croissant record, the license property is the licence's URL (https://creativecommons.org/licenses/by/4.0/). The check of chapter 3 warns when it is a name.
  2. In the card's header, the licence identifier repositories use (cc-by-4.0). In the card's text, the full statement of recipe 7.1.
  3. In the catalog record, the licence field of the World Bank schema, and in the schema.org markup of the dataset page, the license property with the same URL (dissemination cookbook, chapter 5).
  4. Where the terms go beyond the licence, give them a stable URL and link it from the card and the record's description.

Worked example. The Croissant record's license property holds https://creativecommons.org/licenses/by/4.0/, and the check of chapter 3 reports no warning. The card's header holds license: cc-by-4.0, which the public repository reads and shows as a filter. The catalog record's licence field holds the same URL, and the dataset page's schema.org block carries it as the license property. A developer who fetches the record, the page, or the repository listing gets the same answer, and a search for datasets under CC BY 4.0 finds the dataset in all three places.

What this does not do. Machine-readable licensing tells tools what the licence is. It does not tell them what the terms ask; the terms are read by people, which is why the card carries them in plain words.

Recipe

7.3 Review the dataset for disclosure before release

Level
Foundational
Skills
Disclosure control unit; dataset owner
Indicative time
One to three days per dataset

Result: a recorded review confirms that the dataset can be published under the chosen tier, the free text has been screened, and the treatment of the source records is stated in the card.

  1. Establish the source. A dataset built from a public use file inherits its disclosure control (access chapter of the microdata cookbook). A dataset built from the full file is a new product and gets a new review, with the key variables, the uniqueness measures, and the recoding the unit requires.
  2. Screen the free text. Search every text field for names, employers, addresses, phone numbers, and places below the published geography, with a list of patterns and a sample read by a person. Remove or generalize what is found ("works at [employer]" becomes "works at a bank"), and count the records changed.
  3. Decide whether the dataset can be fully open. Where it cannot, publish a synthetic dataset built as the synthetic data cookbook describes, labelled as synthetic, or publish under a registered-user tier with the terms of the access policy.
  4. Record the review in the card's preprocessing and licence sections: what was removed, what was generalized, how many records changed, who reviewed, and when.

Worked example. The source is the public use file of the 2025 Q2 survey, so the records already have direct identifiers removed and geography coarsened to the region. The free-text screening runs a list of patterns for capitalized names, company suffixes, phone numbers, and place names below the region over the 600 job descriptions and flags 41. A reviewer reads all 41 and the 60 unflagged descriptions drawn at random, finds 23 employer names such as "at the Central Hospital" and 4 place names, and generalizes them to "at a hospital" and removes the places; 27 records change. The disclosure control unit's head reviews the screened file, approves open publication, and the card's preprocessing section records the counts, the reviewer, and the date.

What this does not do. A review makes a release defensible under a stated threshold. The terms' prohibition of re-identification and the tier's conditions remain part of the protection, and a dataset of free text is reviewed again when a new version adds records.

Implementation options​

  • Licences: CC BY 4.0 (the default); CC0 where attribution is not required; a registered-user tier with terms where the data cannot be fully open. Licences written for AI artefacts (the RAIL family) restrict uses of models and are a choice for models, seldom for data.
  • Terms: a paragraph in the card; a separate terms page linked from the card and the record.
  • Privacy: public use file as source; synthetic data; free-text screening by patterns and sample reading; a registered tier.

World Bank examples​

  • The World Bank's terms of use for datasets default to CC BY 4.0, with the licence in the schema's license field and on every page.
  • The microdata cookbook's access chapter gives the tiers, the disclosure control workflow, and the rules for automated and AI clients that a training dataset inherits.

Resources​

Common mistakes​

  • A licence as a name only. Tools cannot resolve it; the check warns.
  • Terms that forbid training. They defeat the purpose of the dataset; a restriction the organization wants is written as a condition, with legal advice.
  • Free text unscreened. A job description that names the employer identifies the respondent.
  • The full file as source "because it is anonymized". A new product gets a new review.

Verification​

  • Licence. The same URL appears in the Croissant record, the card header, the catalog record, and the page markup.
  • Terms. The four statements of recipe 7.1 are present in the card.
  • Review. The review is recorded in the card with the reviewer, the date, and the count of records changed.

Checklist​

0 of 5 done

Ticks are saved in this browser only.