# Policy on the use of AI in dissemination (template)

**Office:** [name]
**Approved by:** [role]
**Version:** [number], [date]
**Review date:** [date]

## 1. Scope

This policy covers the use of AI systems, including language models,
embedding models, and automated agents, in the preparation, publication, and
dissemination of official statistics by [office]. It applies to staff,
contractors, and systems operated on the office's behalf.

## 2. Principles

AI use in this office follows the Fundamental Principles of Official
Statistics. In particular:

- Published statistics remain the responsibility of the office. AI systems
  propose; staff decide.
- Confidentiality of individual data is maintained in every AI use.
- Methods, including the use of AI, are documented and published.

## 3. Permitted uses

The following uses are permitted, subject to the controls in section 5:

- Drafting metadata (titles, descriptions, keywords) for review by a curator.
- Ranking and retrieval of published catalog records for search.
- Answering user questions from published records and values, with citations.
- Classifying and extracting information from published documents.
- [add]

## 4. Prohibited uses

- Sending confidential, unpublished, or embargoed data to any external AI
  service.
- Publishing AI-generated numbers that have not been verified against the
  official record.
- Publishing AI-generated text as official metadata without named human
  approval.
- Using AI to make decisions about individuals.
- [add]

## 5. Controls

- **Register.** Every AI component is listed in the AI component register
  with a purpose, owner, model, provider, data used, and review date.
- **Approval.** New uses are approved by [role] before deployment.
- **Human review.** Generated metadata and published text require approval by
  a named staff member, recorded with the date.
- **Data classification.** Data are classified as public, internal, or
  confidential. Only public data may be processed by external services.
- **Access.** AI agents use read-only credentials limited to published data.
- **Logging.** Inputs and outputs of production AI components are logged and
  retained for [period].
- **Evaluation.** Components are tested against the evaluation suite before
  deployment and after any change of model or data.
- **Transparency.** The office publishes a description of where AI is used.

## 6. Roles

| Role | Responsibility |
|---|---|
| [role] | Approves new AI uses and this policy |
| [role] | Maintains the AI component register |
| [role] | Owns the evaluation suite |
| Component owners | Operate their component within this policy |

## 7. Review

This policy is reviewed every [interval] and after any incident.