Removing names is the start of a data review.

Separate three decisions that are often bundled together.#

A technically clean export can still conflict with a client commitment. A permitted use can still require stronger controls than an ordinary file attachment. Give each decision an owner and a place in the approval record. Avoid a single checkbox that says “data is anonymized” without describing what was checked.

NIST’s SP 800-188 describes de-identification as risk reduction that depends on techniques, governance and the sharing model. It also cautions that tools which mask personal information may not provide sufficient de-identification. It is guidance for government datasets, not a blanket legal clearance for a private company’s proposed transaction.

Review the full record, including what is linked to it.#

Ask the system owner to map a complete case: primary fields, comments, quoted email chains, attachments, filenames, export metadata and related tables. A review of the main spreadsheet alone can miss the information in screenshots, document properties or a linked conversation.

These are prompts for a review, not an exhaustive detector specification. The appropriate method depends on the data, recipient, applicable commitments and intended use. Ask a qualified reviewer to define that method and record the basis for the decision.

Preserve the useful relationship while testing the disclosure risk.#

A replacement label can preserve useful relationships: the same invented label can connect messages within an approved case. It can also preserve a route back to the original identity if a mapping table exists or the remaining context is revealing. Record the purpose of each retained link and who controls any mapping.

After transformations, rerun the usefulness check. If every meaningful decision has been removed, the original archive description no longer describes the proposed delivery. Update the brief, counts and buyer discussion. Do not represent an extensively redacted collection as equivalent to the unmodified source.

• Can a reviewer still identify the task and its outcome?
• Are the steps in the correct order?
• Do references point to material that is no longer present?
• Did the transformation create contradictory or misleading text?
• Are missing details labeled rather than silently invented?

Test the preparation process before increasing the scope.#

Use an internally approved test set with known issues. Include the formats that the real collection contains, such as tables, long notes and images. Ask the reviewer to explain what the process can detect and what remains outside it. A test that only covers clean text does not establish coverage for screenshots or audio.

Keep findings by issue type rather than recording only a single pass rate. One missed credential can require a different response from an incorrectly removed generic place name. Log the fix, rerun the relevant checks and keep the decision with the release version. Do not publish real examples from the test set in a public report.

Treat buyer-side access restrictions as part of the proposal. A private, limited evaluation may have a different disclosure profile from a public download or unrestricted onward distribution. Put the intended recipient, purpose and retention period into the buyer questionnaire, and have the agreement reflect the approved arrangement.

Use a release decision that can be reconstructed later.#

1. Identify the exact version and the person responsible for it.
2. Confirm scope and permissions for the recipient and proposed use.
3. Record completed checks, unresolved issues and exclusions.
4. Record the usefulness review after transformations.
5. Confirm delivery controls and the recipient’s authorized access.
6. Document the approval or the reason the release is paused.

Stop when the right to share is unresolved, the recipient cannot describe the use, required checks are incomplete or the collection contains material outside the approved scope. A paused collection is a valid outcome. It is preferable to describing an unreviewed export as ready.

The FTC’s AI privacy and confidentiality guidance stresses that companies must honor their data-use commitments. This guide organizes review questions; it does not determine your legal obligations or certify a release. Use the dataset brief to keep evidence and open questions together.

Questions owners ask