AI integration
Designing an AI redaction workflow for public records
Plan AI-assisted redaction for New Zealand public records. Follow a marked passage through context review, export checks and accountable release approval.

A draft release pack may contain a name in a letter, initials in an email chain and a description that identifies the same person to neighbours. AI can mark passages for an information officer to consider. A workable redaction system preserves the original, records each decision and lets an authorised person check the exact copy that will leave the organisation.
What should a redaction suggestion contain?
Give reviewers the proposed passage with its location, the reason it was flagged and enough surrounding text to judge context. Let them accept, reject or redraw the proposed boundary. Keep their reason and identity with the decision record. An entire page covered with highlights creates another reading task unless staff can resolve marks quickly.
In a fictional community hall complaint, a paragraph mentions the caretaker and a rare personal circumstance. Removing the name alone could still identify the person. Elsewhere, the same name appears as the public contact for booking the hall. Let the officer make different decisions and explain each. The Ombudsman’s official information guides are a starting point for the relevant decision-maker.
How do you test the copy that leaves the organisation?
Produce a separate release copy and open it in another viewer. Check visible pages, selectable text, notes, document properties and included attachments. Use an invented distinctive name in a test pack so an independent reviewer can identify unexpected appearances. Keep a record of the reviewed export version and the destination selected for release.
Then amend one page after sign-off. The release action should require approval of the new export. A previously approved file must stay distinguishable from its replacement. Ask the supplier to show how a reviewer compares the changed pages and who can release the final version when the original approver is away.
Add a second person to the export exercise. Give them the approved release copy and a fictional list of details expected to be absent, without showing the AI highlights. Ask them to read the remaining text for context and check whether the document is still intelligible. Record accidental removals alongside missed ones.
How will reviewers find what the model missed?
Prepare a reference copy by hand before testing the software. Track missed sensitive passages separately from unnecessary suggested redactions. Missing the only indirect identifier in a document matters even when the model marks every obvious email address. Ask staff to record review time and any information unnecessarily removed from the release.
Use invented material while agreeing how originals, review copies and provider logs will be handled. The Office of the Privacy Commissioner advises organisations using generative AI to consider privacy impacts and how providers use information. Rehearse a case that requires entirely manual review, including producing the final copy and decision record. That rehearsal gives the information team a way to keep working if the model changes, becomes unavailable or proves unsuitable for a particular collection.